From e3a0de6284c9a00459eb9a0b2627d10278033685 Mon Sep 17 00:00:00 2001 From: "IM.codes" Date: Wed, 8 Apr 2026 22:46:04 +0800 Subject: [PATCH 01/50] fix(session): honor cc presets in claude sdk sessions --- src/agent/providers/claude-code-sdk.ts | 7 +- src/agent/session-manager.ts | 35 +++-- src/agent/transport-provider.ts | 2 + src/agent/transport-session-runtime.ts | 3 + src/daemon/cc-presets.ts | 29 ++++ src/daemon/command-handler.ts | 34 +++++ test/agent/claude-code-sdk-provider.test.ts | 23 ++++ test/daemon/cc-presets.test.ts | 1 + test/e2e/sdk-transport-flow.test.ts | 139 ++++++++++++++++++++ 9 files changed, 262 insertions(+), 11 deletions(-) diff --git a/src/agent/providers/claude-code-sdk.ts b/src/agent/providers/claude-code-sdk.ts index 8fb1a2f7..c2e077e3 100644 --- a/src/agent/providers/claude-code-sdk.ts +++ b/src/agent/providers/claude-code-sdk.ts @@ -31,6 +31,7 @@ interface ClaudeSdkSessionState { env?: Record; model?: string; description?: string; + systemPrompt?: string; permissionMode: PermissionMode; effort?: TransportEffortLevel; started: boolean; @@ -123,6 +124,7 @@ export class ClaudeCodeSdkProvider implements TransportProvider { env: config.env ?? existing?.env, model: typeof config.agentId === 'string' ? config.agentId : existing?.model, description: config.description ?? existing?.description, + systemPrompt: config.systemPrompt ?? existing?.systemPrompt, permissionMode: this.resolvePermissionMode(), effort: config.effort ?? existing?.effort, started: !!(config.resumeId && config.skipCreate), @@ -243,6 +245,7 @@ export class ClaudeCodeSdkProvider implements TransportProvider { state.emittedToolStates.clear(); const resolvedBinary = this.resolveBinaryPath(this.config); + const baseSystemPrompt = extraSystemPrompt ?? state.description; const options: Record = { cwd: state.cwd, ...(state.env ? { env: { ...process.env, ...state.env } } : {}), @@ -252,7 +255,9 @@ export class ClaudeCodeSdkProvider implements TransportProvider { ...(state.started ? { resume: state.resumeId } : { sessionId: state.resumeId }), ...(state.model ? { model: state.model } : {}), ...(state.effort ? { effort: state.effort } : {}), - ...(extraSystemPrompt ? { appendSystemPrompt: extraSystemPrompt } : {}), + ...(baseSystemPrompt ?? state.systemPrompt ? { + appendSystemPrompt: [baseSystemPrompt, state.systemPrompt].filter(Boolean).join('\n\n'), + } : {}), }; // On Windows where claude resolved to a .cmd shim, override the SDK's // internal spawn so we can prepend `node script.js` and avoid spawn diff --git a/src/agent/session-manager.ts b/src/agent/session-manager.ts index b4335443..2dbde482 100644 --- a/src/agent/session-manager.ts +++ b/src/agent/session-manager.ts @@ -862,7 +862,9 @@ export async function relaunchSessionWithSettings( requestedModel: targetRequestedModel ?? undefined, effort: targetEffort ?? undefined, transportConfig: targetTransportConfig ?? undefined, - ccPreset: targetAgentType === 'claude-code' ? (targetCcPreset ?? undefined) : undefined, + ccPreset: (targetAgentType === 'claude-code' || targetAgentType === 'claude-code-sdk') + ? (targetCcPreset ?? undefined) + : undefined, ...(preserveTransportBinding ? { bindExistingKey: record.providerSessionId, skipCreate: true, @@ -1033,9 +1035,14 @@ export async function restoreTransportSessions(providerId: string): Promise | undefined; + let systemPrompt: string | undefined; + let effectiveRequestedModel = effectiveQwenModel; if (s.providerId === 'claude-code-sdk' && s.ccPreset) { - const { resolvePresetEnv } = await import('../daemon/cc-presets.js'); + const { resolvePresetEnv, getPresetTransportOverrides } = await import('../daemon/cc-presets.js'); extraEnv = await resolvePresetEnv(s.ccPreset, s.ccSessionId ?? undefined); + const presetOverrides = await getPresetTransportOverrides(s.ccPreset); + if (!effectiveRequestedModel && presetOverrides.model) effectiveRequestedModel = presetOverrides.model; + systemPrompt = presetOverrides.systemPrompt; } await runtime.initialize({ sessionKey: effectiveSessionKey, @@ -1045,12 +1052,14 @@ export async function restoreTransportSessions(providerId: string): Promise 0 ? { qwenAvailableModels: availableQwenModels } : {}), ...getQwenDisplayMetadata({ - model: effectiveQwenModel, + model: effectiveRequestedModel, authType: qwenRuntime?.authType ?? s.qwenAuthType, authLimit: qwenRuntime?.authLimit ?? s.qwenAuthLimit, quotaUsageLabel: (qwenRuntime?.authType ?? s.qwenAuthType) === 'qwen-oauth' ? getQwenOAuthQuotaUsageLabel() : undefined, @@ -1118,6 +1127,7 @@ export async function launchTransportSession(opts: LaunchOpts): Promise { let qwenAuthLimit: SessionRecord['qwenAuthLimit'] | undefined; let availableQwenModels: string[] | undefined; let sdkDisplay: Pick | undefined; + let transportSystemPrompt: string | undefined; const storedRequestedModel = !opts.fresh ? existing?.requestedModel : undefined; let requestedTransportModel = opts.requestedModel ?? storedRequestedModel ?? (agentType === 'qwen' ? (opts.qwenModel ?? existing?.qwenModel) : undefined); const effectiveTransportConfig = opts.transportConfig ?? existing?.transportConfig ?? {}; @@ -1150,8 +1160,11 @@ export async function launchTransportSession(opts: LaunchOpts): Promise { effectiveSkipCreate = true; } if (opts.ccPreset) { - const { resolvePresetEnv } = await import('../daemon/cc-presets.js'); + const { resolvePresetEnv, getPresetTransportOverrides } = await import('../daemon/cc-presets.js'); transportEnv = { ...(transportEnv ?? {}), ...(await resolvePresetEnv(opts.ccPreset, transportResumeId)) }; + const presetOverrides = await getPresetTransportOverrides(opts.ccPreset); + if (!requestedTransportModel && presetOverrides.model) requestedTransportModel = presetOverrides.model; + transportSystemPrompt = presetOverrides.systemPrompt; } sdkDisplay = await getClaudeSdkRuntimeConfig().catch(() => ({})); } else if (agentType === 'codex-sdk') { @@ -1167,6 +1180,7 @@ export async function launchTransportSession(opts: LaunchOpts): Promise { cwd: projectDir, label: label || name, description, + ...(transportSystemPrompt ? { systemPrompt: transportSystemPrompt } : {}), agentId: requestedTransportModel, bindExistingKey: effectiveBindExistingKey, skipCreate: effectiveSkipCreate, @@ -1214,6 +1228,7 @@ export async function launchTransportSession(opts: LaunchOpts): Promise { ...(sdkDisplay ?? {}), ...(opts.effort ? { effort: opts.effort } : {}), description, + ...(opts.ccPreset ? { ccPreset: opts.ccPreset } : {}), label, parentSession, userCreated: opts.userCreated, diff --git a/src/agent/transport-provider.ts b/src/agent/transport-provider.ts index 641aeb5e..dc479c0d 100644 --- a/src/agent/transport-provider.ts +++ b/src/agent/transport-provider.ts @@ -114,6 +114,8 @@ export interface SessionConfig { label?: string; /** Persona/system prompt injection — used for session description/role. */ description?: string; + /** Runtime/system prompt injection that should not be surfaced as user-facing description. */ + systemPrompt?: string; /** Parent session key for sub-sessions. */ parentSessionKey?: string; /** If binding to an already-existing remote session, use this key directly. */ diff --git a/src/agent/transport-session-runtime.ts b/src/agent/transport-session-runtime.ts index 404a6e8f..2f6c0180 100644 --- a/src/agent/transport-session-runtime.ts +++ b/src/agent/transport-session-runtime.ts @@ -31,6 +31,7 @@ export class TransportSessionRuntime implements SessionRuntime { private _providerSessionId: string | null = null; private _sending = false; private _description: string | undefined; + private _systemPrompt: string | undefined; private _agentId: string | undefined; private _effort: TransportEffortLevel | undefined; private _unsubscribes: Array<() => void> = []; @@ -103,6 +104,7 @@ export class TransportSessionRuntime implements SessionRuntime { /** Set providerSessionId directly (restore from store without initialize). */ setProviderSessionId(id: string): void { this._providerSessionId = id; } setDescription(desc: string): void { this._description = desc; } + setSystemPrompt(prompt: string): void { this._systemPrompt = prompt; } setAgentId(agentId: string): void { this._agentId = agentId; if (this._providerSessionId) { @@ -126,6 +128,7 @@ export class TransportSessionRuntime implements SessionRuntime { async initialize(config: SessionConfig): Promise { this._providerSessionId = await this.provider.createSession(config); this._description = config.description; + this._systemPrompt = config.systemPrompt; this._agentId = config.agentId; this._effort = config.effort; } diff --git a/src/daemon/cc-presets.ts b/src/daemon/cc-presets.ts index 3ccf67cb..7b15a05b 100644 --- a/src/daemon/cc-presets.ts +++ b/src/daemon/cc-presets.ts @@ -71,6 +71,11 @@ export async function resolvePresetEnv(presetName: string, ccSessionId?: string) const preset = await getPreset(presetName); if (!preset) return {}; const env = { ...preset.env }; + // Backward compatibility: older saved presets used ANTHROPIC_AUTH_TOKEN, + // while current Claude CLI/SDK auth reads ANTHROPIC_API_KEY in bare env mode. + if (env['ANTHROPIC_AUTH_TOKEN'] && !env['ANTHROPIC_API_KEY']) { + env['ANTHROPIC_API_KEY'] = env['ANTHROPIC_AUTH_TOKEN']; + } // Auto-fill model aliases from ANTHROPIC_MODEL if (env['ANTHROPIC_MODEL']) { for (const alias of MODEL_ALIASES) { @@ -87,6 +92,30 @@ export async function resolvePresetEnv(presetName: string, ccSessionId?: string) return env; } +export async function getPresetTransportOverrides(presetName: string): Promise<{ + model?: string; + systemPrompt?: string; +}> { + const preset = await getPreset(presetName); + if (!preset) return {}; + const env = await resolvePresetEnv(presetName); + const configuredModel = env['ANTHROPIC_MODEL']?.trim() || undefined; + const configuredBaseUrl = env['ANTHROPIC_BASE_URL']?.trim() || undefined; + const runtimeFacts = [ + `Authoritative runtime fact: this session is using the Claude Code preset "${preset.name}".`, + configuredBaseUrl ? `Authoritative provider endpoint: ${configuredBaseUrl}.` : undefined, + configuredModel ? `Authoritative runtime model: ${configuredModel}.` : undefined, + configuredModel ? `If the user asks which model you are using, answer exactly with "${configuredModel}".` : 'If the user asks which model or provider you are using, answer with the authoritative runtime facts above.', + configuredBaseUrl ? `If the user asks which provider or endpoint you are using, mention "${configuredBaseUrl}".` : undefined, + 'These runtime facts override any generic Claude Code tool schema, enum, or default.', + 'Do not answer with Sonnet, Opus, Haiku, or any inferred Claude default unless that exact value matches the authoritative runtime model above.', + ].filter(Boolean).join(' '); + return { + ...(configuredModel ? { model: configuredModel } : {}), + ...(runtimeFacts ? { systemPrompt: runtimeFacts } : {}), + }; +} + /** Default init message for non-Anthropic providers (no native web search). */ const DEFAULT_INIT_MESSAGE = 'For web searches, use: curl -s "https://html.duckduckgo.com/html/?q=QUERY" | head -200. Replace QUERY with URL-encoded search terms.'; diff --git a/src/daemon/command-handler.ts b/src/daemon/command-handler.ts index f5aa1391..d24c50ad 100644 --- a/src/daemon/command-handler.ts +++ b/src/daemon/command-handler.ts @@ -136,6 +136,25 @@ function getDefaultThinkingLevel(agentType: string | undefined): TransportEffort return supportsEffort(agentType) ? DEFAULT_TRANSPORT_EFFORT : undefined; } +function isTransportIdentityQuestion(text: string): boolean { + const trimmed = text.trim(); + if (!trimmed || trimmed.length > 200) return false; + return /(?:what|which|exact|current).*(?:model|provider|endpoint)|(?:model|provider|endpoint).*(?:using|use|current)|你.*(?:什么|哪个).*(?:模型|provider|服务商|端点)|当前.*(?:模型|provider|服务商|端点)|你在用什么模型/i.test(trimmed); +} + +async function resolveTransportIdentityAnswer(record: SessionRecord): Promise { + let model = record.activeModel ?? record.requestedModel ?? record.modelDisplay ?? undefined; + let provider = record.providerId ?? undefined; + if (record.ccPreset && record.agentType === 'claude-code-sdk') { + const { getPresetTransportOverrides, resolvePresetEnv } = await import('./cc-presets.js'); + const presetOverrides = await getPresetTransportOverrides(record.ccPreset); + const presetEnv = await resolvePresetEnv(record.ccPreset, record.ccSessionId); + model = model ?? presetOverrides.model; + provider = presetEnv['ANTHROPIC_BASE_URL']?.trim() || provider; + } + return model ?? provider ?? null; +} + async function syncSubSessionIfNeeded(sessionName: string, serverLink: ServerLink): Promise { if (!sessionName.startsWith('deck_sub_')) return; const subId = sessionName.slice('deck_sub_'.length); @@ -1235,6 +1254,19 @@ async function handleSend(cmd: Record, serverLink: ServerLink): } catch { /* */ } return; } + if (record && isTransportIdentityQuestion(text)) { + const identityAnswer = await resolveTransportIdentityAnswer(record); + if (identityAnswer) { + emitTransportUserMessage(text); + timelineEmitter.emit(sessionName, 'assistant.text', { text: identityAnswer, streaming: false }, { source: 'daemon', confidence: 'high' }); + const infoStatus = isLegacy ? 'accepted_legacy' : 'accepted'; + timelineEmitter.emit(sessionName, 'command.ack', { commandId: effectiveId, status: infoStatus }); + try { + serverLink.send({ type: 'command.ack', commandId: effectiveId, status: infoStatus, session: sessionName }); + } catch { /* */ } + return; + } + } const modelMatch = text.trim().match(/^\/model\s+(\S+)(?:\s+.*)?$/); const effortMatch = text.trim().match(/^\/(?:thinking|effort)\s+(\S+)\s*$/); if (record?.agentType === 'qwen' && modelMatch) { @@ -1838,6 +1870,7 @@ async function handleSubSessionStart(cmd: Record, serverLink: S const shellBin = cmd.shellBin as string | null | undefined; const ccSessionId = cmd.ccSessionId as string | null | undefined; const parentSession = cmd.parentSession as string | null | undefined; + const ccPreset = cmd.ccPreset as string | null | undefined; const requestedEffort: unknown = cmd.thinking ?? cmd.effort; const effort = isTransportEffortLevel(requestedEffort) ? requestedEffort @@ -1866,6 +1899,7 @@ async function handleSubSessionStart(cmd: Record, serverLink: S requestedModel: (cmd.requestedModel as string | undefined) ?? (cmd.model as string | undefined), transportConfig: (cmd.transportConfig as Record | undefined) ?? undefined, bindExistingKey, + ...(ccPreset ? { ccPreset } : {}), ...(type === 'claude-code-sdk' ? { ccSessionId: randomUUID(), fresh: true } : {}), ...(type === 'codex-sdk' ? { fresh: true } : {}), ...(effort ? { effort } : {}), diff --git a/test/agent/claude-code-sdk-provider.test.ts b/test/agent/claude-code-sdk-provider.test.ts index cfd13f4b..b9844a73 100644 --- a/test/agent/claude-code-sdk-provider.test.ts +++ b/test/agent/claude-code-sdk-provider.test.ts @@ -261,6 +261,29 @@ describe('ClaudeCodeSdkProvider', () => { }); }); + it('passes runtime-only system prompts without polluting description', async () => { + sdkMock.setNextMessages([ + { type: 'system', subtype: 'init', session_id: 'session-system', model: 'claude-sonnet-4-6' }, + { type: 'result', session_id: 'session-system', subtype: 'success', is_error: false, result: 'OK', usage: { input_tokens: 1, output_tokens: 1, cache_read_input_tokens: 0 } }, + ]); + + const provider = new ClaudeCodeSdkProvider(); + await provider.connect({ binaryPath: 'claude' }); + await provider.createSession({ + sessionKey: 'route-system', + cwd: '/tmp/project', + resumeId: 'session-system', + description: 'Visible description', + systemPrompt: 'Runtime note only', + }); + + await provider.send('route-system', 'hello'); + await flush(); + + const run = sdkMock.runs.at(-1)!; + expect(run.options.appendSystemPrompt).toBe('Visible description\n\nRuntime note only'); + }); + it('emits a fallback streaming delta from assistant text when the SDK does not send text_delta events', async () => { sdkMock.setNextMessages([ { type: 'system', subtype: 'init', session_id: 'session-fallback', model: 'claude-sonnet-4-6' }, diff --git a/test/daemon/cc-presets.test.ts b/test/daemon/cc-presets.test.ts index 3ad82cea..efa066a2 100644 --- a/test/daemon/cc-presets.test.ts +++ b/test/daemon/cc-presets.test.ts @@ -55,6 +55,7 @@ describe('cc presets', () => { await expect(resolvePresetEnv('MiniMax')).resolves.toMatchObject({ ANTHROPIC_BASE_URL: 'https://api.minimax.io/anthropic', ANTHROPIC_AUTH_TOKEN: 'test-token', + ANTHROPIC_API_KEY: 'test-token', ANTHROPIC_MODEL: 'MiniMax-M2.7', ANTHROPIC_SMALL_FAST_MODEL: 'MiniMax-M2.7', ANTHROPIC_DEFAULT_SONNET_MODEL: 'MiniMax-M2.7', diff --git a/test/e2e/sdk-transport-flow.test.ts b/test/e2e/sdk-transport-flow.test.ts index 41915b7d..ad5d4db8 100644 --- a/test/e2e/sdk-transport-flow.test.ts +++ b/test/e2e/sdk-transport-flow.test.ts @@ -389,6 +389,48 @@ describe('sdk transport flow e2e', () => { }); }); + it('preserves cc presets when settings restart switches a main session to claude-code-sdk', async () => { + mocks.store.set('deck_settings_preset_brain', { + name: 'deck_settings_preset_brain', + projectName: 'settings_preset', + role: 'brain', + agentType: 'claude-code', + projectDir: '/tmp/settings-preset', + state: 'idle', + restarts: 0, + restartTimestamps: [], + createdAt: 1, + updatedAt: 1, + ccSessionId: 'cc-settings-preset', + ccPreset: 'MiniMax', + }); + + const serverLink = { send: vi.fn() } as any; + handleWebCommand({ + type: 'session.restart', + sessionName: 'deck_settings_preset_brain', + agentType: 'claude-code-sdk', + }, serverLink); + await flushAsync(); + await waitForCondition(() => serverLink.send.mock.calls.some((call) => call[0]?.type === 'session_list')); + + const switched = mocks.store.get('deck_settings_preset_brain'); + expect(switched?.agentType).toBe('claude-code-sdk'); + expect(switched?.ccPreset).toBe('MiniMax'); + + handleWebCommand({ type: 'session.send', session: 'deck_settings_preset_brain', text: 'hello', commandId: 'cmd-settings-preset' }, serverLink); + await flushAsync(); + + const claudeCall = mocks.claudeCalls.at(-1); + expect(claudeCall?.options.env).toMatchObject({ + ANTHROPIC_BASE_URL: 'https://api.minimax.io/anthropic', + ANTHROPIC_API_KEY: expect.any(String), + ANTHROPIC_MODEL: 'MiniMax-M2.7', + }); + expect(claudeCall?.options.model).toBe('MiniMax-M2.7'); + expect(String(claudeCall?.options.appendSystemPrompt ?? '')).toContain('Authoritative runtime model: MiniMax-M2.7.'); + }); + it('pushes a corrective session_list when settings restart fails', async () => { const tmuxNewSession = newSession as ReturnType; tmuxNewSession.mockRejectedValueOnce(new Error('tmux create failed')); @@ -669,6 +711,103 @@ describe('sdk transport flow e2e', () => { expect(serverLink.send).not.toHaveBeenCalledWith(expect.objectContaining({ type: 'session.error' })); }); + it('persists and applies cc presets for claude-code-sdk main sessions', async () => { + const serverLink = { send: vi.fn() } as any; + + handleWebCommand({ + type: 'session.start', + project: 'ccsdk minimax', + dir: '/tmp/ccsdk-minimax-e2e', + agentType: 'claude-code-sdk', + ccPreset: 'MiniMax', + }, serverLink); + await flushAsync(); + await new Promise((resolve) => setTimeout(resolve, 50)); + + const sessionName = 'deck_ccsdk_minimax_brain'; + const record = mocks.store.get(sessionName); + expect(record?.ccPreset).toBe('MiniMax'); + + handleWebCommand({ type: 'session.send', session: sessionName, text: 'hello', commandId: 'cmd-ccsdk-minimax' }, serverLink); + await flushAsync(); + + const claudeCall = mocks.claudeCalls.at(-1); + expect(claudeCall?.options.env).toMatchObject({ + ANTHROPIC_BASE_URL: 'https://api.minimax.io/anthropic', + ANTHROPIC_API_KEY: expect.any(String), + ANTHROPIC_MODEL: 'MiniMax-M2.7', + }); + expect(claudeCall?.options.model).toBe('MiniMax-M2.7'); + expect(String(claudeCall?.options.appendSystemPrompt ?? '')).toContain('Authoritative runtime model: MiniMax-M2.7.'); + }); + + it('persists and applies cc presets for claude-code-sdk sub-sessions', async () => { + const serverLink = { send: vi.fn() } as any; + + handleWebCommand({ + type: 'subsession.start', + id: 'ccsdk_minimax_sub', + sessionType: 'claude-code-sdk', + cwd: '/tmp/ccsdk-minimax-sub-e2e', + parentSession: 'deck_parent_brain', + ccPreset: 'MiniMax', + }, serverLink); + await flushAsync(); + await waitForCondition(() => serverLink.send.mock.calls.length > 0); + + const sessionName = 'deck_sub_ccsdk_minimax_sub'; + const record = mocks.store.get(sessionName); + expect(record?.ccPreset).toBe('MiniMax'); + expect(serverLink.send).toHaveBeenCalledWith(expect.objectContaining({ + type: 'subsession.sync', + id: 'ccsdk_minimax_sub', + ccPresetId: 'MiniMax', + })); + + handleWebCommand({ type: 'session.send', session: sessionName, text: 'hello', commandId: 'cmd-ccsdk-minimax-sub' }, serverLink); + await flushAsync(); + + const claudeCall = mocks.claudeCalls.at(-1); + expect(claudeCall?.options.env).toMatchObject({ + ANTHROPIC_BASE_URL: 'https://api.minimax.io/anthropic', + ANTHROPIC_API_KEY: expect.any(String), + ANTHROPIC_MODEL: 'MiniMax-M2.7', + }); + expect(claudeCall?.options.model).toBe('MiniMax-M2.7'); + expect(String(claudeCall?.options.appendSystemPrompt ?? '')).toContain('Authoritative runtime model: MiniMax-M2.7.'); + }); + + it('answers claude-code-sdk runtime identity questions from authoritative preset metadata', async () => { + const serverLink = { send: vi.fn() } as any; + + handleWebCommand({ + type: 'session.start', + project: 'ccsdk identity', + dir: '/tmp/ccsdk-identity-e2e', + agentType: 'claude-code-sdk', + ccPreset: 'MiniMax', + }, serverLink); + await flushAsync(); + await new Promise((resolve) => setTimeout(resolve, 50)); + + mocks.claudeCalls.length = 0; + const sessionName = 'deck_ccsdk_identity_brain'; + handleWebCommand({ + type: 'session.send', + session: sessionName, + text: 'Reply with only the exact model or provider you are currently using.', + commandId: 'cmd-ccsdk-identity', + }, serverLink); + await flushAsync(); + + expect(mocks.claudeCalls).toEqual([]); + expect(mocks.emitted).toContainEqual(expect.objectContaining({ + session: sessionName, + type: 'assistant.text', + payload: expect.objectContaining({ text: 'MiniMax-M2.7', streaming: false }), + })); + }); + beforeEach(() => { mocks.store.clear(); mocks.emitted.length = 0; From 40ea144f1977d827845e62c532b7d43f1cd8a6d7 Mon Sep 17 00:00:00 2001 From: "IM.codes" Date: Wed, 8 Apr 2026 22:55:10 +0800 Subject: [PATCH 02/50] fix(ci): remove duplicate ccPreset declaration --- src/daemon/command-handler.ts | 2 -- 1 file changed, 2 deletions(-) diff --git a/src/daemon/command-handler.ts b/src/daemon/command-handler.ts index d24c50ad..f8ec7ff7 100644 --- a/src/daemon/command-handler.ts +++ b/src/daemon/command-handler.ts @@ -1915,8 +1915,6 @@ async function handleSubSessionStart(cmd: Record, serverLink: S } return; } - - const ccPreset = cmd.ccPreset as string | null | undefined; const subCcInitPrompt = cmd.ccInitPrompt as string | null | undefined; const description = cmd.description as string | null | undefined; From 2f622ed5cdaece2f94db4a55f025699fbb07e4a9 Mon Sep 17 00:00:00 2001 From: "IM.codes" Date: Wed, 8 Apr 2026 23:00:20 +0800 Subject: [PATCH 03/50] Throttle transport typewriter updates --- src/daemon/transport-relay.ts | 77 +++++++++++++- test/daemon/transport-relay.test.ts | 143 +++++++++++++++++++++++--- web/src/components/ChatView.tsx | 82 +++++++++++---- web/test/components/ChatView.test.tsx | 50 ++++++++- 4 files changed, 309 insertions(+), 43 deletions(-) diff --git a/src/daemon/transport-relay.ts b/src/daemon/transport-relay.ts index 73cb5889..57f94495 100644 --- a/src/daemon/transport-relay.ts +++ b/src/daemon/transport-relay.ts @@ -16,6 +16,72 @@ import { resolveContextWindow } from '../util/model-context.js'; let sendToServer: ((msg: Record) => void) | null = null; const inFlightMessages = new Map(); +const pendingStreamUpdates = new Map | null; +}>(); +const STREAM_UPDATE_INTERVAL_MS = 200; + +function emitStreamingAssistantText(sessionName: string, eventId: string, text: string): void { + timelineEmitter.emit(sessionName, 'assistant.text', { + text, + streaming: true, + }, { source: 'daemon', confidence: 'high', eventId }); +} + +function clearPendingStreamUpdate(eventId: string): void { + const pending = pendingStreamUpdates.get(eventId); + if (!pending) return; + if (pending.timer) clearTimeout(pending.timer); + pendingStreamUpdates.delete(eventId); +} + +function flushPendingStreamUpdate(eventId: string): void { + const pending = pendingStreamUpdates.get(eventId); + if (!pending || pending.pendingText == null) return; + pending.timer = null; + pending.lastEmitAt = Date.now(); + const nextText = pending.pendingText; + pending.pendingText = null; + emitStreamingAssistantText(pending.sessionName, pending.eventId, nextText); +} + +function emitThrottledStreamingAssistantText(sessionName: string, eventId: string, text: string): void { + const now = Date.now(); + let pending = pendingStreamUpdates.get(eventId); + if (!pending) { + pending = { + sessionName, + eventId, + lastEmitAt: 0, + pendingText: null, + timer: null, + }; + pendingStreamUpdates.set(eventId, pending); + } else { + pending.sessionName = sessionName; + } + + if (pending.lastEmitAt === 0 || now - pending.lastEmitAt >= STREAM_UPDATE_INTERVAL_MS) { + if (pending.timer) { + clearTimeout(pending.timer); + pending.timer = null; + } + pending.pendingText = null; + pending.lastEmitAt = now; + emitStreamingAssistantText(sessionName, eventId, text); + return; + } + + pending.pendingText = text; + if (pending.timer) return; + pending.timer = setTimeout(() => { + flushPendingStreamUpdate(eventId); + }, STREAM_UPDATE_INTERVAL_MS - (now - pending.lastEmitAt)); +} /** Set the send function (called once during server-link setup) */ export function setTransportRelaySend(fn: (msg: Record) => void): void { @@ -34,16 +100,17 @@ export function wireProviderToRelay(provider: TransportProvider): void { // Use delta.delta as the display text directly — the provider's internal // accumulator handles cumulative vs incremental differences. const stableEventId = `transport:${sessionName}:${delta.messageId}`; + const previous = inFlightMessages.get(sessionName); + if (previous && previous.messageId !== delta.messageId) { + clearPendingStreamUpdate(previous.eventId); + } inFlightMessages.set(sessionName, { messageId: delta.messageId, eventId: stableEventId, text: delta.delta, }); - timelineEmitter.emit(sessionName, 'assistant.text', { - text: delta.delta, - streaming: true, - }, { source: 'daemon', confidence: 'high', eventId: stableEventId }); + emitThrottledStreamingAssistantText(sessionName, stableEventId, delta.delta); }); provider.onComplete((providerSid: string, message: AgentMessage) => { @@ -60,6 +127,7 @@ export function wireProviderToRelay(provider: TransportProvider): void { ? tracked.eventId : `transport:${sessionName}:${message.id}`; inFlightMessages.delete(sessionName); + clearPendingStreamUpdate(stableEventId); timelineEmitter.emit(sessionName, 'assistant.text', { text: finalText, streaming: false, @@ -102,6 +170,7 @@ export function wireProviderToRelay(provider: TransportProvider): void { const tracked = inFlightMessages.get(sessionName); inFlightMessages.delete(sessionName); + if (tracked) clearPendingStreamUpdate(tracked.eventId); const errorText = tracked?.text ? `${tracked.text}\n\n⚠️ Error: ${error.message}` : `⚠️ Error: ${error.message}`; diff --git a/test/daemon/transport-relay.test.ts b/test/daemon/transport-relay.test.ts index 0ec27cdd..adf2ae24 100644 --- a/test/daemon/transport-relay.test.ts +++ b/test/daemon/transport-relay.test.ts @@ -140,33 +140,74 @@ describe('transport-relay (timeline-emitter based)', () => { expect(opts.eventId).toBe('transport:sess-a:msg-abc'); }); - it('multiple deltas pass through directly (provider handles accumulation)', () => { + it('throttles streaming updates to at most one emit every 200ms and keeps the latest text', () => { + vi.useFakeTimers(); + vi.setSystemTime(new Date('2026-04-08T00:00:00.000Z')); + + const { provider, fireDelta } = makeMockProvider(); + wireProviderToRelay(provider); + + fireDelta('sess-a', makeDelta({ messageId: 'msg-throttle', delta: 'a' })); + fireDelta('sess-a', makeDelta({ messageId: 'msg-throttle', delta: 'ab' })); + fireDelta('sess-a', makeDelta({ messageId: 'msg-throttle', delta: 'abc' })); + + expect(emitMock).toHaveBeenCalledTimes(1); + expect(emitMock.mock.calls[0][2].text).toBe('a'); + + vi.advanceTimersByTime(199); + expect(emitMock).toHaveBeenCalledTimes(1); + + vi.advanceTimersByTime(1); + expect(emitMock).toHaveBeenCalledTimes(2); + expect(emitMock.mock.calls[1][2].text).toBe('abc'); + + vi.useRealTimers(); + }); + + it('does not let a new message flush an old throttled delta later', () => { + vi.useFakeTimers(); + vi.setSystemTime(new Date('2026-04-08T00:00:00.000Z')); + const { provider, fireDelta } = makeMockProvider(); wireProviderToRelay(provider); - fireDelta('sess-a', makeDelta({ messageId: 'msg-1', delta: 'foo ' })); - fireDelta('sess-a', makeDelta({ messageId: 'msg-1', delta: 'bar ' })); - fireDelta('sess-a', makeDelta({ messageId: 'msg-1', delta: 'baz' })); + fireDelta('sess-a', makeDelta({ messageId: 'msg-old', delta: 'old-a' })); + fireDelta('sess-a', makeDelta({ messageId: 'msg-old', delta: 'old-ab' })); + fireDelta('sess-a', makeDelta({ messageId: 'msg-new', delta: 'new-a' })); + + expect(emitMock.mock.calls.filter(c => c[1] === 'assistant.text')).toHaveLength(2); + expect(emitMock.mock.calls[0][2].text).toBe('old-a'); + expect(emitMock.mock.calls[1][2].text).toBe('new-a'); - expect(emitMock).toHaveBeenCalledTimes(3); + vi.advanceTimersByTime(500); + const textCalls = emitMock.mock.calls.filter(c => c[1] === 'assistant.text'); + expect(textCalls).toHaveLength(2); + expect(textCalls.map(c => c[2].text)).not.toContain('old-ab'); - // Each delta is passed through directly — provider already accumulated - expect(emitMock.mock.calls[0][2].text).toBe('foo '); - expect(emitMock.mock.calls[1][2].text).toBe('bar '); - expect(emitMock.mock.calls[2][2].text).toBe('baz'); + vi.useRealTimers(); }); - it('uses the same stable eventId across multiple deltas for the same message', () => { + it('throttles independently per session', () => { + vi.useFakeTimers(); + vi.setSystemTime(new Date('2026-04-08T00:00:00.000Z')); + const { provider, fireDelta } = makeMockProvider(); wireProviderToRelay(provider); - fireDelta('sess-a', makeDelta({ messageId: 'msg-stable', delta: 'a' })); - fireDelta('sess-a', makeDelta({ messageId: 'msg-stable', delta: 'b' })); + fireDelta('sess-a', makeDelta({ messageId: 'msg-a', delta: 'A1' })); + fireDelta('sess-a', makeDelta({ messageId: 'msg-a', delta: 'A2' })); + fireDelta('sess-b', makeDelta({ messageId: 'msg-b', delta: 'B1' })); + fireDelta('sess-b', makeDelta({ messageId: 'msg-b', delta: 'B2' })); + + expect(emitMock.mock.calls.filter(c => c[1] === 'assistant.text')).toHaveLength(2); + vi.advanceTimersByTime(200); + + const textCalls = emitMock.mock.calls.filter(c => c[1] === 'assistant.text'); + expect(textCalls).toHaveLength(4); + expect(textCalls.filter(c => c[0] === 'sess-a').map(c => c[2].text)).toEqual(['A1', 'A2']); + expect(textCalls.filter(c => c[0] === 'sess-b').map(c => c[2].text)).toEqual(['B1', 'B2']); - const id1 = emitMock.mock.calls[0][3].eventId; - const id2 = emitMock.mock.calls[1][3].eventId; - expect(id1).toBe('transport:sess-a:msg-stable'); - expect(id2).toBe('transport:sess-a:msg-stable'); + vi.useRealTimers(); }); it('does NOT cache to JSONL via appendTransportEvent', () => { @@ -286,6 +327,52 @@ describe('transport-relay (timeline-emitter based)', () => { expect(textCall![3].eventId).toBe('transport:sess-1:msg-4'); }); + it('emits final completion immediately even when a throttled delta is pending', () => { + vi.useFakeTimers(); + vi.setSystemTime(new Date('2026-04-08T00:00:00.000Z')); + + const { provider, fireDelta, fireComplete } = makeMockProvider(); + wireProviderToRelay(provider); + + fireDelta('sess-1', makeDelta({ messageId: 'msg-final', delta: 'a' })); + fireDelta('sess-1', makeDelta({ messageId: 'msg-final', delta: 'ab' })); + emitMock.mockClear(); + + fireComplete('sess-1', makeMessage({ id: 'msg-final', content: 'final answer' })); + + const textCalls = emitMock.mock.calls.filter(c => c[1] === 'assistant.text'); + expect(textCalls).toHaveLength(1); + expect(textCalls[0][2].text).toBe('final answer'); + expect(textCalls[0][2].streaming).toBe(false); + + vi.advanceTimersByTime(500); + expect(emitMock.mock.calls.filter(c => c[1] === 'assistant.text')).toHaveLength(1); + + vi.useRealTimers(); + }); + + it('completion still emits immediately when another session has a pending throttled delta', () => { + vi.useFakeTimers(); + vi.setSystemTime(new Date('2026-04-08T00:00:00.000Z')); + + const { provider, fireDelta, fireComplete } = makeMockProvider(); + wireProviderToRelay(provider); + + fireDelta('sess-a', makeDelta({ messageId: 'msg-a', delta: 'A1' })); + fireDelta('sess-a', makeDelta({ messageId: 'msg-a', delta: 'A2' })); + emitMock.mockClear(); + + fireComplete('sess-b', makeMessage({ id: 'msg-b', sessionId: 'sess-b', content: 'done-b' })); + + const textCalls = emitMock.mock.calls.filter(c => c[1] === 'assistant.text'); + expect(textCalls).toHaveLength(1); + expect(textCalls[0][0]).toBe('sess-b'); + expect(textCalls[0][2].text).toBe('done-b'); + expect(textCalls[0][2].streaming).toBe(false); + + vi.useRealTimers(); + }); + it('caches to JSONL via appendTransportEvent with type assistant.text', async () => { const { provider, fireComplete } = makeMockProvider(); wireProviderToRelay(provider); @@ -416,6 +503,30 @@ describe('transport-relay (timeline-emitter based)', () => { expect(textCall?.[2]?.streaming).toBe(false); expect(textCall?.[2]?.text).toBe('partial answer\n\n⚠️ Error: boom'); }); + + it('emits error immediately even when a throttled delta is pending and suppresses the delayed flush', () => { + vi.useFakeTimers(); + vi.setSystemTime(new Date('2026-04-08T00:00:00.000Z')); + + const { provider, fireDelta, fireError } = makeMockProvider(); + wireProviderToRelay(provider); + + fireDelta('sess-err', makeDelta({ messageId: 'msg-err-2', delta: 'partial' })); + fireDelta('sess-err', makeDelta({ messageId: 'msg-err-2', delta: 'partial+' })); + emitMock.mockClear(); + + fireError('sess-err', { code: 'PROVIDER_ERROR', message: 'boom-now', recoverable: true }); + + const textCalls = emitMock.mock.calls.filter(c => c[1] === 'assistant.text'); + expect(textCalls).toHaveLength(1); + expect(textCalls[0][2].text).toBe('partial+\n\n⚠️ Error: boom-now'); + expect(textCalls[0][2].streaming).toBe(false); + + vi.advanceTimersByTime(500); + expect(emitMock.mock.calls.filter(c => c[1] === 'assistant.text')).toHaveLength(1); + + vi.useRealTimers(); + }); }); // ── emitTransportUserMessage ───────────────────────────────────────────── diff --git a/web/src/components/ChatView.tsx b/web/src/components/ChatView.tsx index 9d9e7c8e..a68046d6 100644 --- a/web/src/components/ChatView.tsx +++ b/web/src/components/ChatView.tsx @@ -55,6 +55,14 @@ interface ViewItem { lastTs?: number; } +interface AssistantBlockProps { + text: string; + ts: number; + onPathClick?: (p: string) => void; + onUrlClick?: (url: string) => void; + onDownload?: (path: string) => void; +} + const TOOL_INPUT_SUMMARY_KEYS = [ 'query', 'command', @@ -398,6 +406,33 @@ export function ChatView({ events, loading, refreshing: _refreshing, loadingOlde document.addEventListener('mouseup', onUp); }, [sessionId]); + const handlePathClick = useCallback((path: string) => { + setFileBrowserPath(path.replace(/^`+|`+$/g, '')); + }, []); + + const handleUrlClick = useCallback((url: string) => { + setPendingUrl(url); + }, []); + + const handleDownload = useCallback((path: string) => { + if (!serverId || !ws) return; + const reqId = ws.fsReadFile(path); + const unsub = ws.onMessage((msg) => { + if (msg.type !== 'fs.read_response' || msg.requestId !== reqId) return; + unsub(); + if (msg.downloadId) { + import('../api.js').then(({ downloadAttachment }) => { + downloadAttachment(serverId, msg.downloadId as string).catch(() => {}); + }); + } + }); + setTimeout(unsub, 30_000); + }, [serverId, ws]); + + const pathClickHandler = ws && !preview ? handlePathClick : undefined; + const urlClickHandler = !preview ? handleUrlClick : undefined; + const downloadHandler = serverId && ws ? handleDownload : undefined; + const viewItems = useMemo(() => buildViewItems(events), [events]); const scrollToBottom = () => { @@ -763,31 +798,19 @@ export function ChatView({ events, loading, refreshing: _refreshing, loadingOlde {!loading && viewItems.map((item, idx) => { const nextItem = viewItems[idx + 1]; const nextTs = nextItem?.ts ?? nextItem?.event?.ts; - const onPathClick = ws && !preview ? (p: string) => setFileBrowserPath(p.replace(/^`+|`+$/g, '')) : undefined; - const onUrlClick = !preview ? (url: string) => setPendingUrl(url) : undefined; - const onDownload = serverId && ws ? (p: string) => { - const reqId = ws.fsReadFile(p); - const unsub = ws.onMessage((msg) => { - if (msg.type !== 'fs.read_response' || msg.requestId !== reqId) return; - unsub(); - if (msg.downloadId) { - import('../api.js').then(({ downloadAttachment }) => { - downloadAttachment(serverId!, msg.downloadId as string).catch(() => {}); - }); - } - }); - // Auto-cleanup after 30s - setTimeout(unsub, 30_000); - } : undefined; return item.type === 'assistant-block' ? ( -
- - -
+ ) : item.type === 'tool-group' ? ( - + ) : ( - + ); })} {!loading &&
} @@ -1022,6 +1045,21 @@ function ToolCallGroup({ events, onPathClick }: { events: TimelineEvent[]; onPat // ToolInputFold removed — replaced by unified ToolBlockFold (CSS max-height based) +const AssistantBlock = memo(function AssistantBlock({ + text, + ts, + onPathClick, + onUrlClick, + onDownload, +}: AssistantBlockProps) { + return ( +
+ + +
+ ); +}); + function AttachmentDownloadButton({ att, serverId, onPathClick }: { att: { id: string; originalName?: string; size?: number; daemonPath?: string }; serverId: string; onPathClick?: (p: string) => void }) { const { t } = useTranslation(); const [error, setError] = useState(null); diff --git a/web/test/components/ChatView.test.tsx b/web/test/components/ChatView.test.tsx index 7a111b8c..c8980683 100644 --- a/web/test/components/ChatView.test.tsx +++ b/web/test/components/ChatView.test.tsx @@ -6,6 +6,8 @@ import { render, waitFor, cleanup, fireEvent } from '@testing-library/preact'; import { afterAll, afterEach, describe, expect, it, vi } from 'vitest'; import { ChatView } from '../../src/components/ChatView.js'; +const chatMarkdownRenderSpy = vi.hoisted(() => vi.fn()); + type ViewportListener = () => void; const visualViewportListeners = new Map>(); const visualViewportMock = { @@ -30,7 +32,10 @@ vi.mock('react-i18next', () => ({ })); vi.mock('../../src/components/ChatMarkdown.js', () => ({ - ChatMarkdown: ({ text }: { text: string }) =>
{text}
, + ChatMarkdown: ({ text }: { text: string }) => { + chatMarkdownRenderSpy(text); + return
{text}
; + }, })); vi.mock('../../src/components/FileBrowser.js', () => ({ @@ -48,6 +53,7 @@ describe('ChatView', () => { afterEach(() => { cleanup(); + chatMarkdownRenderSpy.mockClear(); visualViewportMock.height = 800; visualViewportListeners.clear(); }); @@ -158,6 +164,48 @@ describe('ChatView', () => { }); }); + it('does not rerender an unchanged assistant block when the parent chat rerenders', async () => { + const { rerender } = render( + , + ); + + expect(chatMarkdownRenderSpy.mock.calls.filter(([text]) => text === 'stable block')).toHaveLength(1); + + rerender( + , + ); + + expect(chatMarkdownRenderSpy.mock.calls.filter(([text]) => text === 'stable block')).toHaveLength(1); + }); + it('restores mobile keyboard scroll position from bottom offset instead of snapping to top', async () => { const initialEvents = [ { From cf304df6cad651b56385528455e5774f96b79fd1 Mon Sep 17 00:00:00 2001 From: "IM.codes" Date: Wed, 8 Apr 2026 23:13:54 +0800 Subject: [PATCH 04/50] fix(test): align transport streaming assertions with throttling --- test/daemon/oc-streaming-integration.test.ts | 15 +++++++++++++++ test/e2e/sdk-transport-flow.test.ts | 6 ++---- 2 files changed, 17 insertions(+), 4 deletions(-) diff --git a/test/daemon/oc-streaming-integration.test.ts b/test/daemon/oc-streaming-integration.test.ts index e4f13cbb..d7f7b0cd 100644 --- a/test/daemon/oc-streaming-integration.test.ts +++ b/test/daemon/oc-streaming-integration.test.ts @@ -74,6 +74,10 @@ function emitAgentEvent(payload: Record): void { lastWs().emit('message', JSON.stringify({ type: 'event', event: 'agent', payload })); } +function advanceStreamWindow(): void { + vi.advanceTimersByTime(250); +} + // ── Tests ──────────────────────────────────────────────────────────────────── describe('OC streaming integration: provider → relay → emitter', () => { @@ -105,9 +109,13 @@ describe('OC streaming integration: provider → relay → emitter', () => { emitAgentEvent({ runId, stream: 'lifecycle', data: { phase: 'start' }, key: 'test:sess' }); emitAgentEvent({ runId, stream: 'assistant', data: { delta: '你' }, key: 'test:sess' }); + advanceStreamWindow(); emitAgentEvent({ runId, stream: 'assistant', data: { delta: '好' }, key: 'test:sess' }); + advanceStreamWindow(); emitAgentEvent({ runId, stream: 'assistant', data: { delta: '世' }, key: 'test:sess' }); + advanceStreamWindow(); emitAgentEvent({ runId, stream: 'assistant', data: { delta: '界' }, key: 'test:sess' }); + advanceStreamWindow(); // Filter to assistant.text events for the sanitized session const textEvents = emittedEvents.filter( @@ -135,7 +143,9 @@ describe('OC streaming integration: provider → relay → emitter', () => { emitAgentEvent({ runId, stream: 'lifecycle', data: { phase: 'start' }, key: 'test:s2' }); emitAgentEvent({ runId, stream: 'assistant', data: { delta: 'Hello ' }, key: 'test:s2' }); + advanceStreamWindow(); emitAgentEvent({ runId, stream: 'assistant', data: { delta: 'World' }, key: 'test:s2' }); + advanceStreamWindow(); emitAgentEvent({ runId, stream: 'lifecycle', data: { phase: 'end' }, key: 'test:s2' }); const textEvents = emittedEvents.filter( @@ -161,7 +171,9 @@ describe('OC streaming integration: provider → relay → emitter', () => { // OC sends text=delta (both incremental, text field NOT cumulative) emitAgentEvent({ runId, stream: 'assistant', data: { text: '收到', delta: '收到' }, key: 'test:nc' }); + advanceStreamWindow(); emitAgentEvent({ runId, stream: 'assistant', data: { text: '主人', delta: '主人' }, key: 'test:nc' }); + advanceStreamWindow(); const textEvents = emittedEvents.filter( (e) => e.type === 'assistant.text' && e.sessionId === 'test___nc', @@ -203,8 +215,11 @@ describe('OC streaming integration: provider → relay → emitter', () => { emitAgentEvent({ runId, stream: 'lifecycle', data: { phase: 'start' }, key: 'test:dd' }); emitAgentEvent({ runId, stream: 'assistant', data: { delta: 'A' }, key: 'test:dd' }); + advanceStreamWindow(); emitAgentEvent({ runId, stream: 'assistant', data: { delta: 'B' }, key: 'test:dd' }); + advanceStreamWindow(); emitAgentEvent({ runId, stream: 'assistant', data: { delta: 'C' }, key: 'test:dd' }); + advanceStreamWindow(); emitAgentEvent({ runId, stream: 'lifecycle', data: { phase: 'end' }, key: 'test:dd' }); const textEvents = emittedEvents.filter( diff --git a/test/e2e/sdk-transport-flow.test.ts b/test/e2e/sdk-transport-flow.test.ts index ad5d4db8..5010e30e 100644 --- a/test/e2e/sdk-transport-flow.test.ts +++ b/test/e2e/sdk-transport-flow.test.ts @@ -848,9 +848,8 @@ describe('sdk transport flow e2e', () => { const toolResult = mocks.emitted.find((e) => e.session === SESSION_CC && e.type === 'tool.result'); const claudeCall = mocks.claudeCalls.at(-1); - expect(streaming.map((e) => e.payload.text)).toEqual(['Claude', 'Claude: hello']); + expect(streaming.map((e) => e.payload.text)).toEqual(['Claude']); expect(streaming[0]?.opts?.eventId).toBe(stableEventId); - expect(streaming[1]?.opts?.eventId).toBe(stableEventId); expect(final?.payload.text).toBe('Claude: hello'); expect(final?.opts?.eventId).toBe(stableEventId); expect(usage?.payload.model).toBe('claude-sonnet-4-6'); @@ -887,9 +886,8 @@ describe('sdk transport flow e2e', () => { const toolResult = mocks.emitted.find((e) => e.session === SESSION_CX && e.type === 'tool.result'); const ack = mocks.emitted.find((e) => e.session === SESSION_CX && e.type === 'command.ack'); - expect(streaming.map((e) => e.payload.text)).toEqual(['Codex', 'Codex: hello']); + expect(streaming.map((e) => e.payload.text)).toEqual(['Codex']); expect(streaming[0]?.opts?.eventId).toBe(`transport:${SESSION_CX}:msg-codex-e2e`); - expect(streaming[1]?.opts?.eventId).toBe(`transport:${SESSION_CX}:msg-codex-e2e`); expect(final?.payload.text).toBe('Codex: hello'); expect(final?.opts?.eventId).toBe(`transport:${SESSION_CX}:msg-codex-e2e`); expect(usage?.payload.inputTokens).toBe(7); From ca8de2e47b5247136c2021da985e6fc0defce3a5 Mon Sep 17 00:00:00 2001 From: "IM.codes" Date: Wed, 8 Apr 2026 23:22:13 +0800 Subject: [PATCH 05/50] fix(test): align qwen transport streaming assertions --- test/e2e/qwen-transport-flow.test.ts | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-) diff --git a/test/e2e/qwen-transport-flow.test.ts b/test/e2e/qwen-transport-flow.test.ts index 3bf5a89b..cfad5e06 100644 --- a/test/e2e/qwen-transport-flow.test.ts +++ b/test/e2e/qwen-transport-flow.test.ts @@ -215,9 +215,8 @@ describe('qwen transport flow e2e', () => { expect(user?.payload.text).toBe('hello'); expect(running).toBeDefined(); expect(thinking?.payload.text).toBe(''); - expect(streaming.map((e) => e.payload.text)).toEqual(['Qwen', 'Qwen: hello']); + expect(streaming.map((e) => e.payload.text)).toEqual(['Qwen']); expect(streaming[0]?.opts?.eventId).toBe(stableEventId); - expect(streaming[1]?.opts?.eventId).toBe(stableEventId); expect(final?.payload.text).toBe('Qwen: hello'); expect(final?.opts?.eventId).toBe(stableEventId); const usage = mocks.emitted.find((e) => e.session === SESSION && e.type === 'usage.update'); From 7a228542c4a72f2c6efe66b2bc240498ccbcd817 Mon Sep 17 00:00:00 2001 From: "IM.codes" Date: Wed, 8 Apr 2026 23:34:25 +0800 Subject: [PATCH 06/50] fix(test): stabilize transport preset e2e coverage --- test/e2e/sdk-transport-flow.test.ts | 33 +++++++++++++++++++++++++++++ 1 file changed, 33 insertions(+) diff --git a/test/e2e/sdk-transport-flow.test.ts b/test/e2e/sdk-transport-flow.test.ts index 5010e30e..f6d16c35 100644 --- a/test/e2e/sdk-transport-flow.test.ts +++ b/test/e2e/sdk-transport-flow.test.ts @@ -24,6 +24,39 @@ const mocks = vi.hoisted(() => { return { store, emitted, claudeCalls, codexCalls }; }); +const PRESET_ENV = { + ANTHROPIC_BASE_URL: 'https://api.minimax.io/anthropic', + ANTHROPIC_AUTH_TOKEN: 'test-token', + ANTHROPIC_API_KEY: 'test-token', + ANTHROPIC_MODEL: 'MiniMax-M2.7', + ANTHROPIC_SMALL_FAST_MODEL: 'MiniMax-M2.7', + ANTHROPIC_DEFAULT_SONNET_MODEL: 'MiniMax-M2.7', + ANTHROPIC_DEFAULT_OPUS_MODEL: 'MiniMax-M2.7', + ANTHROPIC_DEFAULT_HAIKU_MODEL: 'MiniMax-M2.7', + IMCODES_CONTEXT_WINDOW: '200000', +}; + +vi.mock('../../src/daemon/cc-presets.js', () => ({ + getPreset: vi.fn(async (name: string) => ( + name.trim().toLowerCase() === 'minimax' + ? { name: 'minimax', env: PRESET_ENV, contextWindow: 200000 } + : undefined + )), + resolvePresetEnv: vi.fn(async (name: string) => ( + name.trim().toLowerCase() === 'minimax' ? { ...PRESET_ENV } : {} + )), + getPresetTransportOverrides: vi.fn(async (name: string) => ( + name.trim().toLowerCase() === 'minimax' + ? { + model: 'MiniMax-M2.7', + systemPrompt: 'Authoritative runtime model: MiniMax-M2.7.', + } + : {} + )), + getPresetInitMessage: vi.fn(() => 'preset-init'), + invalidateCache: vi.fn(), +})); + vi.mock('node:child_process', async (importOriginal) => { const actual = await importOriginal(); const { EventEmitter } = await import('node:events'); From 3ee077f72849f9e7afc996caf19df3114cc715ad Mon Sep 17 00:00:00 2001 From: "IM.codes" Date: Thu, 9 Apr 2026 00:12:10 +0800 Subject: [PATCH 07/50] Keep mobile openspec menus onscreen --- web/src/components/SessionControls.tsx | 48 ++++- web/src/styles.css | 4 +- web/test/components/SessionControls.test.tsx | 185 +++++++++++++++++++ 3 files changed, 234 insertions(+), 3 deletions(-) diff --git a/web/src/components/SessionControls.tsx b/web/src/components/SessionControls.tsx index 02759661..4381496c 100644 --- a/web/src/components/SessionControls.tsx +++ b/web/src/components/SessionControls.tsx @@ -664,6 +664,38 @@ export function SessionControls({ ws, activeSession, inputRef, onAfterAction, on [openSpecLayoutTick, openSpecOpen], ); + const openSpecAuditDropdownStyle = isOpenSpecMobile + ? { + position: 'fixed', + left: 8, + right: 8, + bottom: 72, + minWidth: 0, + width: 'auto', + maxWidth: 'none', + } as const + : { + right: 0, + bottom: 'calc(100% + 6px)', + minWidth: 180, + } as const; + + const openSpecProposeDropdownStyle = isOpenSpecMobile + ? { + position: 'fixed', + left: 8, + right: 8, + bottom: 72, + minWidth: 0, + width: 'auto', + maxWidth: 'none', + } as const + : { + right: 0, + bottom: 'calc(100% + 6px)', + minWidth: 220, + } as const; + useEffect(() => { if (!openSpecOpen || typeof window === 'undefined') return; const refreshLayout = () => setOpenSpecLayoutTick((tick) => tick + 1); @@ -795,6 +827,12 @@ export function SessionControls({ ws, activeSession, inputRef, onAfterAction, on extra.p2pRounds = selection.rounds; if (selection.config.extraPrompt) extra.p2pExtraPrompt = selection.config.extraPrompt; if (selection.config.hopTimeoutMinutes != null) extra.p2pHopTimeoutMs = Math.min(selection.config.hopTimeoutMinutes * 60_000, 600_000); + if (mode === P2P_CONFIG_MODE) { + if (selection.config.advancedPresetKey) extra.p2pAdvancedPresetKey = selection.config.advancedPresetKey; + if (selection.config.advancedRounds) extra.p2pAdvancedRounds = selection.config.advancedRounds; + if (selection.config.advancedRunTimeoutMinutes != null) extra.p2pAdvancedRunTimeoutMinutes = selection.config.advancedRunTimeoutMinutes; + if (selection.config.contextReducer) extra.p2pContextReducer = selection.config.contextReducer; + } }, [p2pSavedConfig]); const buildSendPayload = useCallback((options?: string | BuildSendPayloadOptions): PendingSendPayload | null => { @@ -836,6 +874,12 @@ export function SessionControls({ ws, activeSession, inputRef, onAfterAction, on extra.p2pRounds = override?.rounds ?? cfg.rounds ?? 1; if (cfg.extraPrompt) extra.p2pExtraPrompt = cfg.extraPrompt; if (cfg.hopTimeoutMinutes != null) extra.p2pHopTimeoutMs = Math.min(cfg.hopTimeoutMinutes * 60_000, 600_000); + if (!override?.modeOverride || override.modeOverride === P2P_CONFIG_MODE) { + if (cfg.advancedPresetKey) extra.p2pAdvancedPresetKey = cfg.advancedPresetKey; + if (cfg.advancedRounds) extra.p2pAdvancedRounds = cfg.advancedRounds; + if (cfg.advancedRunTimeoutMinutes != null) extra.p2pAdvancedRunTimeoutMinutes = cfg.advancedRunTimeoutMinutes; + if (cfg.contextReducer) extra.p2pContextReducer = cfg.contextReducer; + } } // For non-config mode overrides (single or combo), send as p2pMode so the daemon uses it if (override?.modeOverride && override.modeOverride !== 'config') { @@ -1407,7 +1451,7 @@ export function SessionControls({ ws, activeSession, inputRef, onAfterAction, on {t('openspec.audit_action')} {openSpecAuditMenu === changeName && ( -
+
{openSpecProposeMenuOpen && ( -
+
+ + +
+ ); + } + if (stage === 'agents') { + return ( +
+ {visibleAgents.map((entry) => ( + + ))} +
+ ); + } + return ( +
+ {['audit', 'discuss', 'review', 'plan', 'brainstorm'].map((mode) => ( + + ))} +
+ ); + }, +})); + const uploadFileMock = vi.fn(); const getUserPrefMock = vi.fn().mockResolvedValue(null); const saveUserPrefMock = vi.fn().mockResolvedValue(undefined); @@ -85,6 +131,7 @@ vi.mock('../../src/api.js', () => ({ import { SessionControls } from '../../src/components/SessionControls.js'; import type { SessionInfo } from '../../src/types.js'; +import { BUILT_IN_ADVANCED_PRESETS } from '@shared/p2p-advanced.js'; const flushAsync = () => new Promise((resolve) => setTimeout(resolve, 0)); @@ -235,6 +282,69 @@ describe('SessionControls', () => { }); }); + it('sends advanced p2p config fields when config mode is used', async () => { + const ws = makeWs(); + getUserPrefMock.mockImplementation(async (key: unknown) => { + if (typeof key === 'string' && key.startsWith('p2p_session_config:')) { + const sessionKey = key.slice('p2p_session_config:'.length); + return JSON.stringify({ + sessions: { + [sessionKey]: { enabled: true, mode: 'audit' }, + 'deck_sub_abc': { enabled: true, mode: 'review' }, + }, + rounds: 3, + advancedPresetKey: 'openspec', + advancedRounds: BUILT_IN_ADVANCED_PRESETS.openspec, + advancedRunTimeoutMinutes: 45, + contextReducer: { + mode: 'clone_sdk_session', + templateSession: sessionKey, + }, + }); + } + return null; + }); + + render( + , + ); + await flushAsync(); + + const input = screen.getByRole('textbox') as HTMLDivElement; + input.textContent = '@'; + fireEvent.input(input); + fireEvent.click(screen.getByText('mock-select-all-config')); + + input.textContent = `${input.textContent} ship it`; + fireEvent.input(input); + fireEvent.click(screen.getByRole('button', { name: /^send$/i })); + + expect(ws.sendSessionCommand).toHaveBeenCalledWith('send', { + sessionName: 'my-session', + text: 'ship it', + p2pAtTargets: [ + { session: '__all__', mode: 'config' }, + ], + p2pSessionConfig: { + 'my-session': { enabled: true, mode: 'audit' }, + 'deck_sub_abc': { enabled: true, mode: 'review' }, + }, + p2pRounds: 3, + p2pAdvancedPresetKey: 'openspec', + p2pAdvancedRounds: BUILT_IN_ADVANCED_PRESETS.openspec, + p2pAdvancedRunTimeoutMinutes: 45, + p2pContextReducer: { + mode: 'clone_sdk_session', + templateSession: 'my-session', + }, + p2pLocale: 'en', + }); + }); + it('keeps the p2p button in solo mode after triggering a combo from the dropdown', async () => { render(); await flushAsync(); @@ -729,6 +839,81 @@ describe('SessionControls', () => { } }); + it('keeps the openspec audit submenu inside the mobile viewport', async () => { + const innerWidth = window.innerWidth; + Object.defineProperty(window, 'innerWidth', { configurable: true, value: 390 }); + + try { + const ws = makeWs(); + render( + , + ); + + fireEvent.click(screen.getByRole('button', { name: /openspec/i })); + ws.emit({ + type: 'fs.ls_response', + requestId: 'openspec-request', + status: 'ok', + resolvedPath: '/repo/openspec/changes', + entries: [ + { name: 'change-a', path: '/repo/openspec/changes/change-a', isDir: true, hidden: false }, + ], + }); + await flushAsync(); + + fireEvent.click(screen.getByRole('button', { name: 'expand change-a' })); + fireEvent.click(screen.getByRole('button', { name: 'audit_action' })); + + const submenu = document.querySelector('.openspec-submenu') as HTMLElement; + expect(submenu).toBeTruthy(); + expect(submenu.style.position).toBe('fixed'); + expect(submenu.style.left).toBe('8px'); + expect(submenu.style.right).toBe('8px'); + } finally { + Object.defineProperty(window, 'innerWidth', { configurable: true, value: innerWidth }); + } + }); + + it('keeps the openspec propose submenu inside the mobile viewport', async () => { + const innerWidth = window.innerWidth; + Object.defineProperty(window, 'innerWidth', { configurable: true, value: 390 }); + + try { + const ws = makeWs(); + render( + , + ); + + fireEvent.click(screen.getByRole('button', { name: /openspec/i })); + ws.emit({ + type: 'fs.ls_response', + requestId: 'openspec-request', + status: 'ok', + resolvedPath: '/repo/openspec/changes', + entries: [], + }); + await flushAsync(); + + fireEvent.click(screen.getByRole('button', { name: 'propose_action' })); + + const submenu = document.querySelector('.openspec-submenu') as HTMLElement; + expect(submenu).toBeTruthy(); + expect(submenu.style.position).toBe('fixed'); + expect(submenu.style.left).toBe('8px'); + expect(submenu.style.right).toBe('8px'); + } finally { + Object.defineProperty(window, 'innerWidth', { configurable: true, value: innerWidth }); + } + }); + it('clears input after send', () => { const ws = makeWs(); render(); From 7683cb310497908033062c1f76b8e5721797c7d0 Mon Sep 17 00:00:00 2001 From: "IM.codes" Date: Thu, 9 Apr 2026 00:20:21 +0800 Subject: [PATCH 08/50] Fix missing openspec test fixture --- web/test/components/SessionControls.test.tsx | 20 +++++++++++++++++--- 1 file changed, 17 insertions(+), 3 deletions(-) diff --git a/web/test/components/SessionControls.test.tsx b/web/test/components/SessionControls.test.tsx index d43dfb1a..01a28625 100644 --- a/web/test/components/SessionControls.test.tsx +++ b/web/test/components/SessionControls.test.tsx @@ -131,10 +131,24 @@ vi.mock('../../src/api.js', () => ({ import { SessionControls } from '../../src/components/SessionControls.js'; import type { SessionInfo } from '../../src/types.js'; -import { BUILT_IN_ADVANCED_PRESETS } from '@shared/p2p-advanced.js'; const flushAsync = () => new Promise((resolve) => setTimeout(resolve, 0)); +const TEST_OPENSPEC_ADVANCED_ROUNDS = [ + { + id: 'initial_audit', + mode: 'audit', + rounds: 1, + summaryMode: 'append', + }, + { + id: 'implementation_plan', + mode: 'plan', + rounds: 1, + summaryMode: 'append', + }, +] as const; + const makeWs = () => { const handlers = new Set<(msg: unknown) => void>(); return { @@ -294,7 +308,7 @@ describe('SessionControls', () => { }, rounds: 3, advancedPresetKey: 'openspec', - advancedRounds: BUILT_IN_ADVANCED_PRESETS.openspec, + advancedRounds: TEST_OPENSPEC_ADVANCED_ROUNDS, advancedRunTimeoutMinutes: 45, contextReducer: { mode: 'clone_sdk_session', @@ -335,7 +349,7 @@ describe('SessionControls', () => { }, p2pRounds: 3, p2pAdvancedPresetKey: 'openspec', - p2pAdvancedRounds: BUILT_IN_ADVANCED_PRESETS.openspec, + p2pAdvancedRounds: TEST_OPENSPEC_ADVANCED_ROUNDS, p2pAdvancedRunTimeoutMinutes: 45, p2pContextReducer: { mode: 'clone_sdk_session', From 31df38a1d098a385dc2f4efa8b212fbb5ceecfc7 Mon Sep 17 00:00:00 2001 From: "IM.codes" Date: Thu, 9 Apr 2026 00:24:03 +0800 Subject: [PATCH 09/50] feat(session): surface transport sdk status phases --- src/agent/providers/claude-code-sdk.ts | 46 +++++++++++++++++ src/agent/providers/codex-sdk.ts | 41 +++++++++++++++ src/agent/providers/qwen.ts | 55 +++++++++++++++++++++ src/agent/transport-provider.ts | 14 ++++++ src/daemon/transport-relay.ts | 15 +++++- test/agent/claude-code-sdk-provider.test.ts | 46 +++++++++++++++++ test/agent/codex-sdk-provider.test.ts | 24 +++++++++ test/agent/qwen-provider.test.ts | 24 +++++++++ test/daemon/transport-relay.test.ts | 34 +++++++++++++ web/src/thinking-utils.ts | 5 +- web/test/thinking-utils.test.ts | 19 +++++++ 11 files changed, 321 insertions(+), 2 deletions(-) create mode 100644 web/test/thinking-utils.test.ts diff --git a/src/agent/providers/claude-code-sdk.ts b/src/agent/providers/claude-code-sdk.ts index c2e077e3..434e3166 100644 --- a/src/agent/providers/claude-code-sdk.ts +++ b/src/agent/providers/claude-code-sdk.ts @@ -10,6 +10,7 @@ import type { ProviderError, SessionConfig, SessionInfoUpdate, + ProviderStatusUpdate, ToolCallEvent, } from '../transport-provider.js'; import { @@ -46,6 +47,7 @@ interface ClaudeSdkSessionState { pendingError?: ProviderError; toolCalls: Map; emittedToolStates: Map; + lastStatusSignature: string | null; } type ClaudeToolBlock = { @@ -92,6 +94,7 @@ export class ClaudeCodeSdkProvider implements TransportProvider { private errorCallbacks: Array<(sessionId: string, error: ProviderError) => void> = []; private toolCallCallbacks: Array<(sessionId: string, tool: ToolCallEvent) => void> = []; private sessionInfoCallbacks: Array<(sessionId: string, info: SessionInfoUpdate) => void> = []; + private statusCallbacks: Array<(sessionId: string, status: ProviderStatusUpdate) => void> = []; async connect(config: ProviderConfig): Promise { const binaryPath = this.resolveBinaryPath(config); @@ -138,6 +141,7 @@ export class ClaudeCodeSdkProvider implements TransportProvider { pendingComplete: undefined, toolCalls: new Map(), emittedToolStates: new Map(), + lastStatusSignature: null, }); this.emitSessionInfo(routeId, { resumeId, ...(config.effort ? { effort: config.effort } : {}) }); return routeId; @@ -187,6 +191,14 @@ export class ClaudeCodeSdkProvider implements TransportProvider { }; } + onStatus(cb: (sessionId: string, status: ProviderStatusUpdate) => void): () => void { + this.statusCallbacks.push(cb); + return () => { + const idx = this.statusCallbacks.indexOf(cb); + if (idx >= 0) this.statusCallbacks.splice(idx, 1); + }; + } + setSessionAgentId(sessionId: string, agentId: string): void { const state = this.sessions.get(sessionId); if (!state) return; @@ -243,6 +255,7 @@ export class ClaudeCodeSdkProvider implements TransportProvider { state.pendingError = undefined; state.toolCalls.clear(); state.emittedToolStates.clear(); + state.lastStatusSignature = null; const resolvedBinary = this.resolveBinaryPath(this.config); const baseSystemPrompt = extraSystemPrompt ?? state.description; @@ -349,6 +362,29 @@ export class ClaudeCodeSdkProvider implements TransportProvider { return; } + if (msg.type === 'system' && msg.subtype === 'compact_boundary') { + this.emitStatus(sessionId, state, { + status: 'compacting', + label: 'Compacting conversation...', + }); + return; + } + + if (msg.type === 'system' && msg.subtype === 'status') { + if (msg.status === 'compacting') { + this.emitStatus(sessionId, state, { + status: 'compacting', + label: 'Compacting conversation...', + }); + } else { + this.emitStatus(sessionId, state, { + status: msg.status, + label: null, + }); + } + return; + } + if (msg.type === 'stream_event') { const event = msg.event; if (event.type === 'message_start' && event.message?.id) { @@ -525,6 +561,16 @@ export class ClaudeCodeSdkProvider implements TransportProvider { for (const cb of this.sessionInfoCallbacks) cb(sessionId, info); } + private emitStatus(sessionId: string, state: ClaudeSdkSessionState, status: ProviderStatusUpdate): void { + const signature = JSON.stringify({ + status: status.status, + label: status.label ?? null, + }); + if (state.lastStatusSignature === signature) return; + state.lastStatusSignature = signature; + for (const cb of this.statusCallbacks) cb(sessionId, status); + } + private emitError(sessionId: string, error: ProviderError): void { if (this.sessions.get(sessionId)?.completed && error.code === PROVIDER_ERROR_CODES.CANCELLED) return; for (const cb of this.errorCallbacks) cb(sessionId, error); diff --git a/src/agent/providers/codex-sdk.ts b/src/agent/providers/codex-sdk.ts index 7e2f9614..4a936474 100644 --- a/src/agent/providers/codex-sdk.ts +++ b/src/agent/providers/codex-sdk.ts @@ -9,6 +9,7 @@ import type { ProviderError, SessionConfig, SessionInfoUpdate, + ProviderStatusUpdate, ToolCallEvent, } from '../transport-provider.js'; import { @@ -53,6 +54,7 @@ interface CodexSdkSessionState { cached_input_tokens: number; output_tokens: number; }; + lastStatusSignature: string | null; } function toolFromItem(item: Record, lifecycle: 'started' | 'completed'): ToolCallEvent | null { @@ -173,6 +175,7 @@ export class CodexSdkProvider implements TransportProvider { private errorCallbacks: Array<(sessionId: string, error: ProviderError) => void> = []; private toolCallCallbacks: Array<(sessionId: string, tool: ToolCallEvent) => void> = []; private sessionInfoCallbacks: Array<(sessionId: string, info: SessionInfoUpdate) => void> = []; + private statusCallbacks: Array<(sessionId: string, status: ProviderStatusUpdate) => void> = []; private child: ChildProcessWithoutNullStreams | null = null; private rl: ReadlineInterface | null = null; private nextRequestId = 1; @@ -221,6 +224,7 @@ export class CodexSdkProvider implements TransportProvider { pendingComplete: undefined, cancelled: false, lastUsage: undefined, + lastStatusSignature: null, }); if (config.resumeId || config.effort) this.emitSessionInfo(routeId, { ...(config.resumeId ? { resumeId: config.resumeId } : {}), ...(config.effort ? { effort: config.effort } : {}) }); return routeId; @@ -272,6 +276,14 @@ export class CodexSdkProvider implements TransportProvider { }; } + onStatus(cb: (sessionId: string, status: ProviderStatusUpdate) => void): () => void { + this.statusCallbacks.push(cb); + return () => { + const idx = this.statusCallbacks.indexOf(cb); + if (idx >= 0) this.statusCallbacks.splice(idx, 1); + }; + } + setSessionAgentId(sessionId: string, agentId: string): void { const state = this.sessions.get(sessionId); if (!state) return; @@ -302,6 +314,7 @@ export class CodexSdkProvider implements TransportProvider { state.pendingComplete = undefined; state.cancelled = false; state.lastUsage = undefined; + state.lastStatusSignature = null; await this.startTurn(sessionId, state, message); } @@ -470,6 +483,7 @@ export class CodexSdkProvider implements TransportProvider { const sessionId = this.threadToSession.get(params.threadId); const state = sessionId ? this.sessions.get(sessionId) : null; if (!sessionId || !state) return; + this.clearStatus(sessionId, state); state.currentMessageId = params.itemId; state.currentText += String(params.delta ?? ''); const delta: MessageDelta = { @@ -490,6 +504,16 @@ export class CodexSdkProvider implements TransportProvider { const item = params.item as Record | undefined; if (!item) return; + if (item.type === 'reasoning') { + this.emitStatus(sessionId, state, { + status: 'thinking', + label: 'Thinking...', + }); + return; + } + + this.clearStatus(sessionId, state); + const tool = toolFromItem(item, method === 'item/started' ? 'started' : 'completed'); if (tool) { for (const cb of this.toolCallCallbacks) cb(sessionId, tool); @@ -522,16 +546,19 @@ export class CodexSdkProvider implements TransportProvider { const status = turn.status; if (status === 'failed') { + this.clearStatus(sessionId, state); state.runningTurnId = undefined; this.emitError(sessionId, this.makeError(PROVIDER_ERROR_CODES.PROVIDER_ERROR, turn.error?.message ?? 'Codex turn failed', false, turn.error)); return; } if (status === 'interrupted') { + this.clearStatus(sessionId, state); state.runningTurnId = undefined; this.emitError(sessionId, this.makeError(PROVIDER_ERROR_CODES.CANCELLED, 'Codex turn cancelled', true)); return; } + this.clearStatus(sessionId, state); state.pendingComplete = { id: state.currentMessageId ?? `${sessionId}:agent-message`, sessionId, @@ -582,6 +609,20 @@ export class CodexSdkProvider implements TransportProvider { for (const cb of this.sessionInfoCallbacks) cb(sessionId, info); } + private emitStatus(sessionId: string, state: CodexSdkSessionState, status: ProviderStatusUpdate): void { + const signature = JSON.stringify({ + status: status.status, + label: status.label ?? null, + }); + if (state.lastStatusSignature === signature) return; + state.lastStatusSignature = signature; + for (const cb of this.statusCallbacks) cb(sessionId, status); + } + + private clearStatus(sessionId: string, state: CodexSdkSessionState): void { + this.emitStatus(sessionId, state, { status: null, label: null }); + } + private emitError(sessionId: string, error: ProviderError): void { for (const cb of this.errorCallbacks) cb(sessionId, error); } diff --git a/src/agent/providers/qwen.ts b/src/agent/providers/qwen.ts index 01282761..d2665d2a 100644 --- a/src/agent/providers/qwen.ts +++ b/src/agent/providers/qwen.ts @@ -10,6 +10,7 @@ import type { ProviderCapabilities, ProviderConfig, ProviderError, + ProviderStatusUpdate, SessionConfig, ToolCallEvent, } from '../transport-provider.js'; @@ -46,6 +47,7 @@ interface QwenSessionState { toolUseByIndex: Map; toolUseById: Map; emittedToolSignatures: Map; + lastStatusSignature: string | null; } function toQwenReasoning(effort: TransportEffortLevel): false | { effort: 'low' | 'medium' | 'high' } { @@ -185,6 +187,7 @@ export class QwenProvider implements TransportProvider { private completeCallbacks: Array<(sessionId: string, message: AgentMessage) => void> = []; private errorCallbacks: Array<(sessionId: string, error: ProviderError) => void> = []; private toolCallCallbacks: Array<(sessionId: string, tool: ToolCallEvent) => void> = []; + private statusCallbacks: Array<(sessionId: string, status: ProviderStatusUpdate) => void> = []; async connect(config: ProviderConfig): Promise { const resolved = resolveExecutableForSpawn(QWEN_BIN); @@ -226,6 +229,7 @@ export class QwenProvider implements TransportProvider { toolUseByIndex: existing?.toolUseByIndex ?? new Map(), toolUseById: existing?.toolUseById ?? new Map(), emittedToolSignatures: existing?.emittedToolSignatures ?? new Map(), + lastStatusSignature: existing?.lastStatusSignature ?? null, }); return sessionId; } @@ -267,6 +271,14 @@ export class QwenProvider implements TransportProvider { this.toolCallCallbacks.push(cb); } + onStatus(cb: (sessionId: string, status: ProviderStatusUpdate) => void): () => void { + this.statusCallbacks.push(cb); + return () => { + const idx = this.statusCallbacks.indexOf(cb); + if (idx >= 0) this.statusCallbacks.splice(idx, 1); + }; + } + setSessionAgentId(sessionId: string, agentId: string): void { const state = this.sessions.get(sessionId); if (!state) return; @@ -304,6 +316,7 @@ export class QwenProvider implements TransportProvider { toolUseByIndex: new Map(), toolUseById: new Map(), emittedToolSignatures: new Map(), + lastStatusSignature: null, }; if (state.child && !state.child.killed) { throw this.makeError(PROVIDER_ERROR_CODES.PROVIDER_ERROR, 'Qwen session is already busy', true); @@ -317,6 +330,7 @@ export class QwenProvider implements TransportProvider { state.toolUseByIndex.clear(); state.toolUseById.clear(); state.emittedToolSignatures.clear(); + state.lastStatusSignature = null; const args = [ '-p', message, @@ -424,11 +438,20 @@ export class QwenProvider implements TransportProvider { const event = payload.event; if (!event) return; if (event.type === 'message_start') { + this.clearStatus(sessionId, state); state.currentMessageId = event.message?.id ?? randomUUID(); state.currentText = ''; return; } + if (event.type === 'content_block_start' && event.content_block?.type === 'thinking') { + this.emitStatus(sessionId, state, { + status: 'thinking', + label: 'Thinking...', + }); + return; + } if (event.type === 'content_block_start' && event.content_block?.type === 'tool_use') { + this.clearStatus(sessionId, state); const toolId = event.content_block.id ?? randomUUID(); const toolName = event.content_block.name ?? 'tool'; const toolInput = event.content_block.input; @@ -463,6 +486,7 @@ export class QwenProvider implements TransportProvider { return; } if (event.type === 'content_block_delta' && event.delta?.type === 'text_delta' && typeof event.delta.text === 'string') { + this.clearStatus(sessionId, state); state.currentMessageId ??= randomUUID(); state.currentText += event.delta.text; this.deltaCallbacks.forEach((cb) => cb(sessionId, { @@ -473,6 +497,13 @@ export class QwenProvider implements TransportProvider { })); return; } + if (event.type === 'content_block_delta' && event.delta?.type === 'thinking_delta' && typeof event.delta.thinking === 'string') { + this.emitStatus(sessionId, state, { + status: 'thinking', + label: 'Thinking...', + }); + return; + } if (event.type === 'content_block_delta' && event.delta?.type === 'input_json_delta' && typeof event.index === 'number') { const tool = state.toolUseByIndex.get(event.index); if (!tool) return; @@ -500,6 +531,14 @@ export class QwenProvider implements TransportProvider { } if (payload.type === 'assistant') { + if ((payload.message?.content ?? []).some((block) => block?.type === 'thinking')) { + this.emitStatus(sessionId, state, { + status: 'thinking', + label: 'Thinking...', + }); + } else { + this.clearStatus(sessionId, state); + } for (const block of payload.message?.content ?? []) { if (block?.type === 'tool_use' && block.id) { if (hasMeaningfulToolValue(block.input)) { @@ -530,6 +569,7 @@ export class QwenProvider implements TransportProvider { } if (payload.type === 'user') { + this.clearStatus(sessionId, state); for (const block of payload.message?.content ?? []) { if (block?.type !== 'tool_result' || !block.tool_use_id) continue; const output = stringifyToolResultContent(block.content); @@ -551,6 +591,7 @@ export class QwenProvider implements TransportProvider { } if (payload.type === 'result') { + this.clearStatus(sessionId, state); if (payload.is_error) { emitError(payload.error?.message || stderrBuf || 'Qwen execution failed', payload); return; @@ -635,6 +676,20 @@ export class QwenProvider implements TransportProvider { return { code, message, recoverable, details }; } + private emitStatus(sessionId: string, state: QwenSessionState, status: ProviderStatusUpdate): void { + const signature = JSON.stringify({ + status: status.status, + label: status.label ?? null, + }); + if (state.lastStatusSignature === signature) return; + state.lastStatusSignature = signature; + for (const cb of this.statusCallbacks) cb(sessionId, status); + } + + private clearStatus(sessionId: string, state: QwenSessionState): void { + this.emitStatus(sessionId, state, { status: null, label: null }); + } + private async ensureSettingsPath(state: QwenSessionState): Promise { if (!state.settingsDir) { state.settingsDir = await mkdtemp(path.join(os.tmpdir(), 'imcodes-qwen-thinking-')); diff --git a/src/agent/transport-provider.ts b/src/agent/transport-provider.ts index dc479c0d..bd5833aa 100644 --- a/src/agent/transport-provider.ts +++ b/src/agent/transport-provider.ts @@ -180,6 +180,14 @@ export interface SessionInfoUpdate { effort?: TransportEffortLevel; } +/** Provider-reported transient execution status (e.g. compacting). */ +export interface ProviderStatusUpdate { + /** Machine-readable transient status. Null clears any previously active status. */ + status: string | null; + /** Human-readable label shown in the footer/status line. Null clears the label. */ + label?: string | null; +} + // ── TransportProvider interface ───────────────────────────────────────────── /** @@ -286,6 +294,12 @@ export interface TransportProvider { */ onSessionInfo?(cb: (sessionId: string, info: SessionInfoUpdate) => void): () => void; + /** + * Register a callback for transient provider status changes (e.g. compacting). + * Used by SDK-backed providers that can surface phases before a final response arrives. + */ + onStatus?(cb: (sessionId: string, status: ProviderStatusUpdate) => void): () => void; + /** * Register a callback for approval requests from the agent. * Only call when capabilities.approval is true. diff --git a/src/daemon/transport-relay.ts b/src/daemon/transport-relay.ts index 57f94495..6f8e711b 100644 --- a/src/daemon/transport-relay.ts +++ b/src/daemon/transport-relay.ts @@ -5,7 +5,7 @@ * JSONL watchers), so ChatView renders them without any special handling. * Also cached to local JSONL for replay on reconnect/restart. */ -import type { TransportProvider, ProviderError } from '../agent/transport-provider.js'; +import type { TransportProvider, ProviderError, ProviderStatusUpdate } from '../agent/transport-provider.js'; import type { MessageDelta, AgentMessage, ToolCallEvent } from '../../shared/agent-message.js'; import { TRANSPORT_MSG } from '../../shared/transport-events.js'; import { resolveSessionName } from '../agent/session-manager.js'; @@ -244,6 +244,19 @@ export function wireProviderToRelay(provider: TransportProvider): void { ...(tool.detail !== undefined ? { detail: tool.detail } : {}), }); }); + + provider.onStatus?.((providerSid: string, status: ProviderStatusUpdate) => { + const sessionName = resolveSessionName(providerSid); + if (!sessionName) { + logger.debug({ providerSid }, 'transport-relay: unresolved route for status — dropped'); + return; + } + + timelineEmitter.emit(sessionName, 'agent.status', { + status: status.status, + ...(status.label !== undefined ? { label: status.label } : {}), + }, { source: 'daemon', confidence: 'high' }); + }); } /** Emit user.message through timeline when user sends to a transport session. */ diff --git a/test/agent/claude-code-sdk-provider.test.ts b/test/agent/claude-code-sdk-provider.test.ts index b9844a73..a9daf58a 100644 --- a/test/agent/claude-code-sdk-provider.test.ts +++ b/test/agent/claude-code-sdk-provider.test.ts @@ -411,6 +411,52 @@ describe('ClaudeCodeSdkProvider', () => { const run = sdkMock.runs.at(-1)!; expect(run.options.effort).toBe('high'); }); + + it('emits compacting status updates from Claude SDK system status messages', async () => { + sdkMock.setNextMessages([ + { type: 'system', subtype: 'init', session_id: 'session-status', model: 'claude-sonnet-4-6' }, + { type: 'system', subtype: 'status', session_id: 'session-status', status: 'compacting' }, + { type: 'system', subtype: 'status', session_id: 'session-status', status: null }, + { type: 'result', session_id: 'session-status', subtype: 'success', is_error: false, result: 'OK', usage: { input_tokens: 1, output_tokens: 1, cache_read_input_tokens: 0 } }, + ]); + + const provider = new ClaudeCodeSdkProvider(); + await provider.connect({ binaryPath: 'claude' }); + await provider.createSession({ sessionKey: 'route-status', cwd: '/tmp/project', resumeId: 'session-status' }); + + const statuses: Array<{ status: string | null; label?: string | null }> = []; + provider.onStatus?.((_sid, status) => statuses.push(status)); + + await provider.send('route-status', 'hello'); + await flush(); + + expect(statuses).toEqual([ + { status: 'compacting', label: 'Compacting conversation...' }, + { status: null, label: null }, + ]); + }); + + it('emits compacting status from compact boundary system messages', async () => { + sdkMock.setNextMessages([ + { type: 'system', subtype: 'init', session_id: 'session-boundary', model: 'claude-sonnet-4-6' }, + { type: 'system', subtype: 'compact_boundary', session_id: 'session-boundary' }, + { type: 'result', session_id: 'session-boundary', subtype: 'success', is_error: false, result: 'OK', usage: { input_tokens: 1, output_tokens: 1, cache_read_input_tokens: 0 } }, + ]); + + const provider = new ClaudeCodeSdkProvider(); + await provider.connect({ binaryPath: 'claude' }); + await provider.createSession({ sessionKey: 'route-boundary', cwd: '/tmp/project', resumeId: 'session-boundary' }); + + const statuses: Array<{ status: string | null; label?: string | null }> = []; + provider.onStatus?.((_sid, status) => statuses.push(status)); + + await provider.send('route-boundary', 'hello'); + await flush(); + + expect(statuses).toEqual([ + { status: 'compacting', label: 'Compacting conversation...' }, + ]); + }); }); type ToolEventSnapshot = { name: string; status: string; input: unknown; output?: unknown; detail?: unknown }; diff --git a/test/agent/codex-sdk-provider.test.ts b/test/agent/codex-sdk-provider.test.ts index b4a8ff0c..1bbc2fd8 100644 --- a/test/agent/codex-sdk-provider.test.ts +++ b/test/agent/codex-sdk-provider.test.ts @@ -309,4 +309,28 @@ describe('CodexSdkProvider', () => { const turnStartReq = child.requests.find((req) => req.method === 'turn/start'); expect(turnStartReq?.params?.effort).toBe('high'); }); + + it('emits thinking status from reasoning items and clears it on streamed assistant text', async () => { + const provider = new CodexSdkProvider(); + await provider.connect({ binaryPath: 'codex' }); + await provider.createSession({ sessionKey: 'route-status', cwd: '/tmp/project' }); + + const statuses: Array<{ status: string | null; label?: string | null }> = []; + provider.onStatus?.((_sid, status) => statuses.push(status)); + + await provider.send('route-status', 'hello'); + const child = childProcessMock.children[0]; + child.emits({ + method: 'item/started', + params: { threadId: 'thread-1', turnId: 'turn-1', item: { id: 'reason-1', type: 'reasoning', text: 'Planning next step' } }, + }); + child.emits({ method: 'item/agentMessage/delta', params: { threadId: 'thread-1', turnId: 'turn-1', itemId: 'msg-1', delta: 'O' } }); + child.emits({ method: 'turn/completed', params: { threadId: 'thread-1', turn: { id: 'turn-1', status: 'completed', error: null } } }); + await flush(); + + expect(statuses).toEqual([ + { status: 'thinking', label: 'Thinking...' }, + { status: null, label: null }, + ]); + }); }); diff --git a/test/agent/qwen-provider.test.ts b/test/agent/qwen-provider.test.ts index e7af72c4..94ebee18 100644 --- a/test/agent/qwen-provider.test.ts +++ b/test/agent/qwen-provider.test.ts @@ -404,4 +404,28 @@ describe('QwenProvider', () => { }, ]); }); + + it('emits thinking status from qwen thinking blocks and clears it on text output', async () => { + const provider = new QwenProvider(); + await provider.connect({}); + await provider.createSession({ sessionKey: 'sess-thinking', cwd: '/tmp/project' }); + + const statuses: Array<{ status: string | null; label?: string | null }> = []; + provider.onStatus?.((_sid, status) => statuses.push(status)); + + await provider.send('sess-thinking', 'think'); + const run = lastSpawn(); + run.child.stdout.write(`${JSON.stringify({ type: 'stream_event', event: { type: 'message_start', message: { id: 'msg-thinking' } } })}\n`); + run.child.stdout.write(`${JSON.stringify({ type: 'stream_event', event: { type: 'content_block_start', index: 0, content_block: { type: 'thinking' } } })}\n`); + run.child.stdout.write(`${JSON.stringify({ type: 'stream_event', event: { type: 'content_block_delta', index: 0, delta: { type: 'thinking_delta', thinking: 'Analyzing...' } } })}\n`); + run.child.stdout.write(`${JSON.stringify({ type: 'stream_event', event: { type: 'content_block_delta', index: 1, delta: { type: 'text_delta', text: 'Done' } } })}\n`); + run.child.emit('close', 0, null); + await flushIO(); + + expect(statuses).toEqual([ + { status: null, label: null }, + { status: 'thinking', label: 'Thinking...' }, + { status: null, label: null }, + ]); + }); }); diff --git a/test/daemon/transport-relay.test.ts b/test/daemon/transport-relay.test.ts index adf2ae24..75ec8944 100644 --- a/test/daemon/transport-relay.test.ts +++ b/test/daemon/transport-relay.test.ts @@ -48,12 +48,14 @@ type DeltaCb = (sessionId: string, delta: MessageDelta) => void; type CompleteCb = (sessionId: string, message: AgentMessage) => void; type ErrorCb = (sessionId: string, error: { code: string; message: string; recoverable: boolean }) => void; type ToolCb = (sessionId: string, tool: ToolCallEvent) => void; +type StatusCb = (sessionId: string, status: { status: string | null; label?: string | null }) => void; function makeMockProvider() { let deltaCb: DeltaCb | undefined; let completeCb: CompleteCb | undefined; let errorCb: ErrorCb | undefined; let toolCb: ToolCb | undefined; + let statusCb: StatusCb | undefined; return { provider: { @@ -61,11 +63,13 @@ function makeMockProvider() { onComplete: (cb: CompleteCb) => { completeCb = cb; return () => { completeCb = undefined; }; }, onError: (cb: ErrorCb) => { errorCb = cb; return () => { errorCb = undefined; }; }, onToolCall: (cb: ToolCb) => { toolCb = cb; }, + onStatus: (cb: StatusCb) => { statusCb = cb; return () => { statusCb = undefined; }; }, } as unknown as TransportProvider, fireDelta: (sid: string, delta: MessageDelta) => deltaCb?.(sid, delta), fireComplete: (sid: string, msg: AgentMessage) => completeCb?.(sid, msg), fireError: (sid: string, err: { code: string; message: string; recoverable: boolean }) => errorCb?.(sid, err), fireTool: (sid: string, tool: ToolCallEvent) => toolCb?.(sid, tool), + fireStatus: (sid: string, status: { status: string | null; label?: string | null }) => statusCb?.(sid, status), }; } @@ -641,6 +645,36 @@ describe('transport-relay (timeline-emitter based)', () => { expect(call![3].eventId).toBe('transport-tool:sess-tool:tool-1:result'); }); }); + + describe('onStatus', () => { + it('emits agent.status into the timeline for provider status updates', () => { + const { provider, fireStatus } = makeMockProvider(); + wireProviderToRelay(provider); + + fireStatus('sess-status', { status: 'compacting', label: 'Compacting conversation...' }); + + expect(emitMock).toHaveBeenCalledWith( + 'sess-status', + 'agent.status', + { status: 'compacting', label: 'Compacting conversation...' }, + expect.objectContaining({ source: 'daemon', confidence: 'high' }), + ); + }); + + it('emits unlabeled status updates so the frontend can clear stale status text', () => { + const { provider, fireStatus } = makeMockProvider(); + wireProviderToRelay(provider); + + fireStatus('sess-status', { status: null, label: null }); + + expect(emitMock).toHaveBeenCalledWith( + 'sess-status', + 'agent.status', + { status: null, label: null }, + expect.objectContaining({ source: 'daemon', confidence: 'high' }), + ); + }); + }); }); // ── useTimeline same-ID replacement (logic extracted for unit testing) ─────── diff --git a/web/src/thinking-utils.ts b/web/src/thinking-utils.ts index 140c0275..a6fb5896 100644 --- a/web/src/thinking-utils.ts +++ b/web/src/thinking-utils.ts @@ -57,7 +57,10 @@ export function getActiveThinkingTs(events: Array<{ type: string; ts: number; pa export function getActiveStatusText(events: Array<{ type: string; payload?: Record }>): string | null { for (let i = events.length - 1; i >= 0; i--) { const e = events[i]; - if (e.type === 'agent.status' && e.payload?.label) return String(e.payload.label); + if (e.type === 'agent.status') { + if (e.payload?.label) return String(e.payload.label); + return null; + } if (e.type !== 'agent.status') break; } return null; diff --git a/web/test/thinking-utils.test.ts b/web/test/thinking-utils.test.ts new file mode 100644 index 00000000..fd61cb4c --- /dev/null +++ b/web/test/thinking-utils.test.ts @@ -0,0 +1,19 @@ +import { describe, expect, it } from 'vitest'; + +import { getActiveStatusText } from '../src/thinking-utils.js'; + +describe('getActiveStatusText', () => { + it('returns the latest trailing status label', () => { + expect(getActiveStatusText([ + { type: 'assistant.text', payload: { text: 'done' } }, + { type: 'agent.status', payload: { status: 'compacting', label: 'Compacting conversation...' } }, + ])).toBe('Compacting conversation...'); + }); + + it('treats an unlabeled trailing status as an explicit clear', () => { + expect(getActiveStatusText([ + { type: 'agent.status', payload: { status: 'compacting', label: 'Compacting conversation...' } }, + { type: 'agent.status', payload: { status: null, label: null } }, + ])).toBeNull(); + }); +}); From 5a48617308f97397ceac0807b3e44f7812f7eef8 Mon Sep 17 00:00:00 2001 From: "IM.codes" Date: Thu, 9 Apr 2026 00:33:16 +0800 Subject: [PATCH 10/50] Fix SessionControls advanced config typing --- web/src/components/SessionControls.tsx | 31 ++++++++++++++++---------- 1 file changed, 19 insertions(+), 12 deletions(-) diff --git a/web/src/components/SessionControls.tsx b/web/src/components/SessionControls.tsx index 4381496c..5d3a6ab3 100644 --- a/web/src/components/SessionControls.tsx +++ b/web/src/components/SessionControls.tsx @@ -127,6 +127,13 @@ function getP2pMenuItemColor(mode: string, active: boolean): string { return getP2pModeColor(mode); } +type OptionalP2pAdvancedConfig = { + advancedPresetKey?: unknown; + advancedRounds?: unknown; + advancedRunTimeoutMinutes?: unknown; + contextReducer?: unknown; +}; + interface PendingAtTarget { session: string; mode: string; @@ -166,6 +173,16 @@ type ManualP2pResolveResult = { type P2pConfigTab = 'participants' | 'combos'; +function appendOptionalAdvancedP2pConfig(extra: Record, config: P2pSavedConfig): void { + const advanced = config as P2pSavedConfig & OptionalP2pAdvancedConfig; + if (advanced.advancedPresetKey) extra.p2pAdvancedPresetKey = advanced.advancedPresetKey; + if (advanced.advancedRounds) extra.p2pAdvancedRounds = advanced.advancedRounds; + if (advanced.advancedRunTimeoutMinutes != null) { + extra.p2pAdvancedRunTimeoutMinutes = advanced.advancedRunTimeoutMinutes; + } + if (advanced.contextReducer) extra.p2pContextReducer = advanced.contextReducer; +} + // Enter moved after ↓ arrow const SHORTCUTS: Array<{ label: string; title: string; data: string; wide?: boolean }> = [ { label: 'Esc', title: 'Escape', data: '\x1b' }, @@ -827,12 +844,7 @@ export function SessionControls({ ws, activeSession, inputRef, onAfterAction, on extra.p2pRounds = selection.rounds; if (selection.config.extraPrompt) extra.p2pExtraPrompt = selection.config.extraPrompt; if (selection.config.hopTimeoutMinutes != null) extra.p2pHopTimeoutMs = Math.min(selection.config.hopTimeoutMinutes * 60_000, 600_000); - if (mode === P2P_CONFIG_MODE) { - if (selection.config.advancedPresetKey) extra.p2pAdvancedPresetKey = selection.config.advancedPresetKey; - if (selection.config.advancedRounds) extra.p2pAdvancedRounds = selection.config.advancedRounds; - if (selection.config.advancedRunTimeoutMinutes != null) extra.p2pAdvancedRunTimeoutMinutes = selection.config.advancedRunTimeoutMinutes; - if (selection.config.contextReducer) extra.p2pContextReducer = selection.config.contextReducer; - } + if (mode === P2P_CONFIG_MODE) appendOptionalAdvancedP2pConfig(extra, selection.config); }, [p2pSavedConfig]); const buildSendPayload = useCallback((options?: string | BuildSendPayloadOptions): PendingSendPayload | null => { @@ -874,12 +886,7 @@ export function SessionControls({ ws, activeSession, inputRef, onAfterAction, on extra.p2pRounds = override?.rounds ?? cfg.rounds ?? 1; if (cfg.extraPrompt) extra.p2pExtraPrompt = cfg.extraPrompt; if (cfg.hopTimeoutMinutes != null) extra.p2pHopTimeoutMs = Math.min(cfg.hopTimeoutMinutes * 60_000, 600_000); - if (!override?.modeOverride || override.modeOverride === P2P_CONFIG_MODE) { - if (cfg.advancedPresetKey) extra.p2pAdvancedPresetKey = cfg.advancedPresetKey; - if (cfg.advancedRounds) extra.p2pAdvancedRounds = cfg.advancedRounds; - if (cfg.advancedRunTimeoutMinutes != null) extra.p2pAdvancedRunTimeoutMinutes = cfg.advancedRunTimeoutMinutes; - if (cfg.contextReducer) extra.p2pContextReducer = cfg.contextReducer; - } + if (!override?.modeOverride || override.modeOverride === P2P_CONFIG_MODE) appendOptionalAdvancedP2pConfig(extra, cfg); } // For non-config mode overrides (single or combo), send as p2pMode so the daemon uses it if (override?.modeOverride && override.modeOverride !== 'config') { From adc175019075a6766282bb6a3b7d62501c1d9355 Mon Sep 17 00:00:00 2001 From: "IM.codes" Date: Thu, 9 Apr 2026 00:56:23 +0800 Subject: [PATCH 11/50] Implement smart P2P advanced workflow support --- shared/p2p-advanced.ts | 381 +++++ shared/p2p-modes.ts | 9 + shared/p2p-status.ts | 24 + src/daemon/command-handler.ts | 21 +- src/daemon/cron-executor.ts | 9 +- src/daemon/p2p-orchestrator.ts | 886 ++++++++++- test/daemon/cron-executor.test.ts | 51 +- test/daemon/p2p-orchestrator.test.ts | 1454 +++++++++++++++++- test/daemon/p2p-parser.test.ts | 70 +- test/shared/p2p-advanced.test.ts | 184 +++ web/src/app.tsx | 60 +- web/src/components/P2pConfigPanel.tsx | 416 ++++- web/src/components/P2pProgressCard.tsx | 6 +- web/src/i18n/locales/en.json | 22 + web/src/i18n/locales/es.json | 22 + web/src/i18n/locales/ja.json | 22 + web/src/i18n/locales/ko.json | 22 + web/src/i18n/locales/ru.json | 22 + web/src/i18n/locales/zh-CN.json | 22 + web/src/i18n/locales/zh-TW.json | 22 + web/src/p2p-run-mapping.ts | 106 ++ web/test/components/P2pConfigPanel.test.tsx | 50 + web/test/components/P2pProgressCard.test.tsx | 103 ++ web/test/p2p-run-mapping.test.ts | 219 +++ 24 files changed, 4016 insertions(+), 187 deletions(-) create mode 100644 shared/p2p-advanced.ts create mode 100644 test/shared/p2p-advanced.test.ts create mode 100644 web/src/p2p-run-mapping.ts create mode 100644 web/test/p2p-run-mapping.test.ts diff --git a/shared/p2p-advanced.ts b/shared/p2p-advanced.ts new file mode 100644 index 00000000..1f8543c5 --- /dev/null +++ b/shared/p2p-advanced.ts @@ -0,0 +1,381 @@ +import { isTransportSessionAgentType } from './agent-types.js'; + +const LEGACY_MODE_KEYS = new Set(['audit', 'review', 'plan', 'brainstorm', 'discuss']); +const COMBO_SEPARATOR = '>'; + +export type P2pAdvancedPresetKey = 'openspec'; +export type P2pRoundPreset = + | 'discussion' + | 'openspec_propose' + | 'proposal_audit' + | 'implementation' + | 'implementation_audit' + | 'custom'; +export type P2pRoundExecutionMode = 'single_main' | 'multi_dispatch'; +export type P2pRoundPermissionScope = 'analysis_only' | 'artifact_generation' | 'implementation'; +export type P2pRoundVerdictPolicy = 'none' | 'smart_gate' | 'forced_rework'; +export type P2pContextReducerMode = 'reuse_existing_session' | 'clone_sdk_session'; +export type P2pVerdictMarker = 'PASS' | 'REWORK'; +export type P2pDispatchStyle = 'initiator_only' | 'worker_hops'; +export type P2pSynthesisStyle = 'none' | 'initiator_summary'; + +export interface P2pContextReducerConfig { + mode: P2pContextReducerMode; + sessionName?: string; + templateSession?: string; +} + +export interface P2pAdvancedJumpRule { + targetRoundId: string; + marker?: P2pVerdictMarker; + minTriggers: number; + maxTriggers: number; +} + +export interface P2pAdvancedRound { + id: string; + title: string; + preset: P2pRoundPreset; + executionMode: P2pRoundExecutionMode; + permissionScope: P2pRoundPermissionScope; + timeoutMinutes?: number; + artifactOutputs?: string[]; + promptAppend?: string; + verdictPolicy?: P2pRoundVerdictPolicy; + jumpRule?: P2pAdvancedJumpRule; +} + +export interface P2pParticipantSnapshotEntry { + sessionName: string; + agentType: string; + parentSession?: string | null; +} + +export interface P2pHelperDiagnostic { + code: + | 'P2P_HELPER_PRIMARY_FAILED' + | 'P2P_HELPER_FALLBACK_FAILED' + | 'P2P_HELPER_CLEANUP_FAILED' + | 'P2P_COMPRESSION_SKIPPED_NO_FALLBACK' + | 'P2P_VERDICT_MISSING'; + attempt: number; + sourceSession?: string | null; + templateSession?: string | null; + fallbackSession?: string | null; + timestamp: number; + message?: string; +} + +export interface P2pResolvedRound { + id: string; + title: string; + modeKey: string; + preset: P2pRoundPreset; + executionMode: P2pRoundExecutionMode; + permissionScope: P2pRoundPermissionScope; + timeoutMinutes: number; + timeoutMs: number; + promptAppend: string; + verdictPolicy: P2pRoundVerdictPolicy; + jumpRule?: P2pAdvancedJumpRule; + dispatchStyle: P2pDispatchStyle; + synthesisStyle: P2pSynthesisStyle; + requiresVerdict: boolean; + presetPrompt: string; + summaryPrompt?: string; + authoritativeVerdictWriter: 'initiator_summary' | 'initiator_only' | null; + allowRouting: boolean; + artifactOutputs: string[]; + artifactConvention: 'none' | 'explicit' | 'openspec_convention'; +} + +export interface ResolveP2pRoundPlanOptions { + modeOverride?: string; + roundsOverride?: number; + hopTimeoutMinutes?: number; + advancedPresetKey?: string | null; + advancedRounds?: P2pAdvancedRound[] | null; + advancedRunTimeoutMinutes?: number | null; + contextReducer?: P2pContextReducerConfig | null; + participants?: P2pParticipantSnapshotEntry[] | null; +} + +export interface P2pResolvedPlan { + advanced: boolean; + rounds: P2pResolvedRound[]; + overallRunTimeoutMinutes?: number; + contextReducer?: P2pContextReducerConfig; + helperEligibleSnapshot?: P2pParticipantSnapshotEntry[]; +} + +const DEFAULT_HOP_TIMEOUT_MINUTES = 8; +const DEFAULT_ADVANCED_RUN_TIMEOUT_MINUTES = 30; + +function parseModePipeline(mode: string): string[] { + if (mode.includes(COMBO_SEPARATOR)) { + return mode.split(COMBO_SEPARATOR).map((entry) => entry.trim()).filter(Boolean); + } + return [mode]; +} + +function isValidLegacyMode(mode: string): boolean { + return LEGACY_MODE_KEYS.has(mode); +} + +function validateLegacyMode(mode: string): void { + const pipeline = parseModePipeline(mode); + if (pipeline.length === 0 || pipeline.some((entry) => !isValidLegacyMode(entry))) { + throw new Error(`Invalid P2P mode pipeline: ${mode}`); + } +} + +function cloneRound(value: T): T { + return JSON.parse(JSON.stringify(value)) as T; +} + +function createOpenSpecPreset(): P2pAdvancedRound[] { + return [ + { + id: 'discussion', + title: 'Discussion', + preset: 'discussion', + executionMode: 'multi_dispatch', + permissionScope: 'analysis_only', + timeoutMinutes: 5, + verdictPolicy: 'none', + }, + { + id: 'openspec_propose', + title: 'OpenSpec Propose', + preset: 'openspec_propose', + executionMode: 'single_main', + permissionScope: 'artifact_generation', + timeoutMinutes: 8, + verdictPolicy: 'none', + }, + { + id: 'proposal_audit', + title: 'Proposal Audit', + preset: 'proposal_audit', + executionMode: 'single_main', + permissionScope: 'analysis_only', + timeoutMinutes: 6, + verdictPolicy: 'none', + }, + { + id: 'implementation', + title: 'Implementation', + preset: 'implementation', + executionMode: 'multi_dispatch', + permissionScope: 'implementation', + timeoutMinutes: 8, + verdictPolicy: 'none', + }, + { + id: 'implementation_audit', + title: 'Implementation Audit', + preset: 'implementation_audit', + executionMode: 'single_main', + permissionScope: 'analysis_only', + timeoutMinutes: 6, + verdictPolicy: 'smart_gate', + jumpRule: { + targetRoundId: 'implementation', + marker: 'REWORK', + minTriggers: 0, + maxTriggers: 2, + }, + }, + ]; +} + +export const BUILT_IN_ADVANCED_PRESETS: Record = { + openspec: createOpenSpecPreset(), +}; + +const PRESET_PROMPTS: Record = { + discussion: 'Clarify the request, collect missing constraints, and synthesize the strongest next-step understanding from the evidence in the discussion file and referenced code.', + openspec_propose: 'Produce an OpenSpec-ready proposal/design/tasks result from the discussion and code context. Write concrete artifacts, acceptance criteria, and implementation scope rather than broad notes.', + proposal_audit: 'Audit the proposal artifacts for missing scope, missing acceptance criteria, contradictions, and weak assumptions. Strengthen the proposal without changing the requested objective.', + implementation: 'Execute the implementation work required by the current round. Prefer concrete code and tests over commentary, while staying within the stated scope and artifact targets.', + implementation_audit: 'Audit the implementation result against the requested scope, artifact outputs, and acceptance criteria. End with an authoritative verdict marker.', + custom: 'Follow the configured round contract exactly. Stay within the declared permission scope and use the configured outputs and prompt append as the operative instruction.', +}; + +const SUMMARY_PROMPTS: Partial> = { + discussion: 'Synthesize the key points, areas of agreement, and open questions from this round. Then assign concrete follow-up focus for the next round.', + implementation: 'Synthesize the implementation outputs from the worker evidence. Produce one authoritative implementation summary that references the latest completed attempt.', + implementation_audit: 'Write one authoritative audit synthesis and end with exactly one verdict marker line: `` or ``.', +}; + +function buildLegacyResolvedRound(mode: string, roundIndex: number, totalRounds: number, hopTimeoutMinutes?: number): P2pResolvedRound { + const pipeline = parseModePipeline(mode); + const modeKey = pipeline[Math.min(roundIndex - 1, pipeline.length - 1)] ?? mode; + return { + id: `legacy_${roundIndex}`, + title: `Round ${roundIndex}`, + modeKey, + preset: 'custom', + executionMode: 'multi_dispatch', + permissionScope: 'analysis_only', + timeoutMinutes: hopTimeoutMinutes ?? DEFAULT_HOP_TIMEOUT_MINUTES, + timeoutMs: (hopTimeoutMinutes ?? DEFAULT_HOP_TIMEOUT_MINUTES) * 60_000, + promptAppend: '', + verdictPolicy: 'none', + dispatchStyle: 'worker_hops', + synthesisStyle: 'initiator_summary', + requiresVerdict: false, + presetPrompt: '', + summaryPrompt: totalRounds === roundIndex ? undefined : 'Synthesize the key points, areas of agreement, and open questions from this round. Then assign concrete follow-up focus for the next round.', + authoritativeVerdictWriter: null, + allowRouting: false, + artifactOutputs: [], + artifactConvention: 'none', + }; +} + +function defaultArtifactConvention(round: P2pAdvancedRound): 'none' | 'explicit' | 'openspec_convention' { + if (round.preset === 'openspec_propose' && (!round.artifactOutputs || round.artifactOutputs.length === 0)) { + return 'openspec_convention'; + } + if (round.permissionScope === 'artifact_generation') return 'explicit'; + return 'none'; +} + +function normalizeAdvancedRound(round: P2pAdvancedRound): P2pResolvedRound { + const verdictPolicy = round.verdictPolicy ?? 'none'; + const artifactConvention = defaultArtifactConvention(round); + const artifactOutputs = artifactConvention === 'openspec_convention' + ? ['openspec/changes'] + : [...(round.artifactOutputs ?? [])]; + const synthesisStyle: P2pSynthesisStyle = round.executionMode === 'multi_dispatch' ? 'initiator_summary' : 'none'; + const requiresVerdict = verdictPolicy !== 'none'; + const authoritativeVerdictWriter = requiresVerdict + ? (round.executionMode === 'multi_dispatch' ? 'initiator_summary' : 'initiator_only') + : null; + const allowRouting = round.preset !== 'proposal_audit' && requiresVerdict && !!round.jumpRule; + return { + id: round.id, + title: round.title, + modeKey: round.preset === 'custom' ? 'custom' : round.preset, + preset: round.preset, + executionMode: round.executionMode, + permissionScope: round.permissionScope, + timeoutMinutes: round.timeoutMinutes ?? DEFAULT_HOP_TIMEOUT_MINUTES, + timeoutMs: (round.timeoutMinutes ?? DEFAULT_HOP_TIMEOUT_MINUTES) * 60_000, + promptAppend: round.promptAppend?.trim() ?? '', + verdictPolicy, + jumpRule: round.jumpRule ? cloneRound(round.jumpRule) : undefined, + dispatchStyle: round.executionMode === 'single_main' ? 'initiator_only' : 'worker_hops', + synthesisStyle, + requiresVerdict, + presetPrompt: PRESET_PROMPTS[round.preset], + summaryPrompt: synthesisStyle === 'initiator_summary' ? SUMMARY_PROMPTS[round.preset] : undefined, + authoritativeVerdictWriter, + allowRouting, + artifactOutputs, + artifactConvention, + }; +} + +function validateAdvancedRoundIds(rounds: P2pAdvancedRound[]): void { + const seen = new Set(); + for (const round of rounds) { + if (!round.id.trim()) throw new Error('Advanced P2P round ids must be non-empty'); + if (seen.has(round.id)) throw new Error(`Duplicate advanced P2P round id: ${round.id}`); + seen.add(round.id); + } +} + +function validateContextReducer( + reducer: P2pContextReducerConfig | null | undefined, + participants: P2pParticipantSnapshotEntry[] | null | undefined, +): P2pContextReducerConfig | undefined { + if (!reducer) return undefined; + const snapshot = participants ?? []; + const lookup = new Map(snapshot.map((entry) => [entry.sessionName, entry])); + if (reducer.mode === 'reuse_existing_session') { + if (!reducer.sessionName) throw new Error('contextReducer.sessionName is required for reuse_existing_session'); + const target = lookup.get(reducer.sessionName); + if (!target || !isTransportSessionAgentType(target.agentType)) { + throw new Error(`Reducer session is not an eligible SDK-backed participant: ${reducer.sessionName}`); + } + } else { + if (!reducer.templateSession) throw new Error('contextReducer.templateSession is required for clone_sdk_session'); + const template = lookup.get(reducer.templateSession); + if (!template || !isTransportSessionAgentType(template.agentType)) { + throw new Error(`Reducer template is not an eligible SDK-backed participant: ${reducer.templateSession}`); + } + } + return cloneRound(reducer); +} + +function validateAdvancedRounds(rounds: P2pAdvancedRound[]): void { + validateAdvancedRoundIds(rounds); + const ids = new Set(rounds.map((round) => round.id)); + for (const round of rounds) { + const verdictPolicy = round.verdictPolicy ?? 'none'; + const artifactConvention = defaultArtifactConvention(round); + if (round.permissionScope === 'artifact_generation' && artifactConvention === 'explicit' && (!round.artifactOutputs || round.artifactOutputs.length === 0)) { + throw new Error(`Artifact-generation round "${round.id}" must declare artifact outputs`); + } + if (verdictPolicy === 'forced_rework') { + if (!round.jumpRule) throw new Error(`forced_rework round "${round.id}" requires a jumpRule`); + if (round.jumpRule.minTriggers < 0) throw new Error(`forced_rework round "${round.id}" has invalid minTriggers`); + if (round.jumpRule.maxTriggers < round.jumpRule.minTriggers) throw new Error(`forced_rework round "${round.id}" has invalid maxTriggers`); + } + if (round.jumpRule) { + if (!ids.has(round.jumpRule.targetRoundId)) throw new Error(`Round "${round.id}" jumps to unknown target "${round.jumpRule.targetRoundId}"`); + const currentIndex = rounds.findIndex((entry) => entry.id === round.id); + const targetIndex = rounds.findIndex((entry) => entry.id === round.jumpRule?.targetRoundId); + if (targetIndex >= currentIndex) throw new Error(`Round "${round.id}" must jump backward to an earlier round`); + if (round.preset === 'proposal_audit') throw new Error('proposal_audit cannot drive routing in v1'); + } + } +} + +export function resolveP2pRoundPlan(options: ResolveP2pRoundPlanOptions): P2pResolvedPlan { + const { + modeOverride, + roundsOverride, + hopTimeoutMinutes, + advancedPresetKey, + advancedRounds, + advancedRunTimeoutMinutes, + contextReducer, + participants, + } = options; + + const advancedRequested = !!advancedPresetKey || !!advancedRounds?.length; + if (!advancedRequested) { + const mode = modeOverride ?? 'discuss'; + validateLegacyMode(mode); + const comboRounds = parseModePipeline(mode).length; + const totalRounds = Math.max(1, roundsOverride ?? comboRounds); + return { + advanced: false, + rounds: Array.from({ length: totalRounds }, (_, index) => buildLegacyResolvedRound(mode, index + 1, totalRounds, hopTimeoutMinutes)), + }; + } + + if (advancedPresetKey && advancedPresetKey !== 'openspec') { + throw new Error(`Unknown advanced P2P preset: ${advancedPresetKey}`); + } + + const presetRounds = advancedPresetKey === 'openspec' + ? cloneRound(BUILT_IN_ADVANCED_PRESETS.openspec) + : []; + const rawRounds = advancedRounds?.length ? cloneRound(advancedRounds) : presetRounds; + if (rawRounds.length === 0) throw new Error('Advanced P2P requires at least one round'); + validateAdvancedRounds(rawRounds); + const validatedReducer = validateContextReducer(contextReducer, participants); + const helperEligibleSnapshot = (participants ?? []).filter((entry) => isTransportSessionAgentType(entry.agentType)); + + return { + advanced: true, + rounds: rawRounds.map((round) => normalizeAdvancedRound(round)), + overallRunTimeoutMinutes: advancedRunTimeoutMinutes ?? DEFAULT_ADVANCED_RUN_TIMEOUT_MINUTES, + contextReducer: validatedReducer, + helperEligibleSnapshot, + }; +} diff --git a/shared/p2p-modes.ts b/shared/p2p-modes.ts index ec9e2448..5285bae5 100644 --- a/shared/p2p-modes.ts +++ b/shared/p2p-modes.ts @@ -1,4 +1,5 @@ /** P2P Quick Discussion mode configuration. */ +import type { P2pAdvancedPresetKey, P2pAdvancedRound, P2pContextReducerConfig } from './p2p-advanced.js'; /** The "config" meta-mode — each session uses its own saved default mode. */ export const P2P_CONFIG_MODE = 'config' as const; @@ -18,6 +19,14 @@ export interface P2pSavedConfig { extraPrompt?: string; /** Per-hop timeout in minutes. Default: 8. */ hopTimeoutMinutes?: number; + /** Built-in advanced workflow preset key. */ + advancedPresetKey?: P2pAdvancedPresetKey; + /** Advanced round overrides / full custom workflow definition. */ + advancedRounds?: P2pAdvancedRound[]; + /** Whole-run timeout for advanced workflows in minutes. */ + advancedRunTimeoutMinutes?: number; + /** Optional context compression/helper config for advanced workflows. */ + contextReducer?: P2pContextReducerConfig; } diff --git a/shared/p2p-status.ts b/shared/p2p-status.ts index 0e1c5740..534b2c49 100644 --- a/shared/p2p-status.ts +++ b/shared/p2p-status.ts @@ -5,6 +5,7 @@ * server/web consumers, and expose richer parallel-hop progress through * additive fields. */ +import type { P2pHelperDiagnostic } from './p2p-advanced.js'; export const P2P_RUN_STATUS_VALUES = [ 'queued', @@ -175,6 +176,29 @@ export interface P2pRunUpdatePayload { hop_counts?: P2pHopCounts; completed_round_hops_count?: number; terminal_reason?: 'completed' | 'timed_out' | 'failed' | 'cancelled' | null; + advanced_p2p_enabled?: boolean; + current_round_id?: string | null; + current_execution_step?: number | null; + current_round_attempt?: number | null; + round_attempt_counts?: Record; + round_jump_counts?: Record; + routing_history?: Array<{ + fromRoundId?: string | null; + toRoundId?: string | null; + trigger?: string | null; + atStep: number; + atAttempt?: number | null; + timestamp: number; + }>; + helper_diagnostics?: P2pHelperDiagnostic[]; + advanced_nodes?: Array<{ + id: string; + title: string; + preset?: string; + status: P2pProgressNodeStatus; + attempt?: number; + step?: number; + }>; [key: string]: unknown; } diff --git a/src/daemon/command-handler.ts b/src/daemon/command-handler.ts index f8ec7ff7..c06791bb 100644 --- a/src/daemon/command-handler.ts +++ b/src/daemon/command-handler.ts @@ -30,6 +30,7 @@ import { promisify } from 'node:util'; const execAsync = promisify(execCb); import { startP2pRun, cancelP2pRun, getP2pRun, listP2pRuns, serializeP2pRun, type P2pTarget } from './p2p-orchestrator.js'; import { getComboRoundCount, parseModePipeline, P2P_CONFIG_MODE, type P2pSessionConfig } from '../../shared/p2p-modes.js'; +import type { P2pAdvancedRound, P2pContextReducerConfig } from '../../shared/p2p-advanced.js'; import { CRON_MSG } from '../../shared/cron-types.js'; import { executeCronJob } from './cron-executor.js'; import { TRANSPORT_MSG } from '../../shared/transport-events.js'; @@ -1061,6 +1062,10 @@ async function handleSend(cmd: Record, serverLink: ServerLink): let p2pExtraPrompt = (cmd as any).p2pExtraPrompt as string | undefined; const p2pLocale = (cmd as any).p2pLocale as string | undefined; const p2pHopTimeoutMs = (cmd as any).p2pHopTimeoutMs as number | undefined; + const p2pAdvancedPresetKey = (cmd as any).p2pAdvancedPresetKey as string | undefined; + const p2pAdvancedRounds = (cmd as any).p2pAdvancedRounds as P2pAdvancedRound[] | undefined; + const p2pAdvancedRunTimeoutMinutes = (cmd as any).p2pAdvancedRunTimeoutMinutes as number | undefined; + const p2pContextReducer = (cmd as any).p2pContextReducer as P2pContextReducerConfig | undefined; const p2pModeField = (cmd as any).p2pMode as string | undefined; const p2pAtTargets = (cmd as any).p2pAtTargets as Array<{ session: string; mode: string }> | undefined; const explicitTargets = directTargetSession @@ -1200,7 +1205,21 @@ async function handleSend(cmd: Record, serverLink: ServerLink): const langInstr = `Use the user's selected i18n language (${langName}) for the discussion.`; p2pExtraPrompt = p2pExtraPrompt ? `${p2pExtraPrompt}\n${langInstr}` : langInstr; } - const run = await startP2pRun(sessionName, tokens.agents, tokens.cleanText, fileContents, serverLink, p2pRounds, p2pExtraPrompt, resolvedMode || undefined, p2pHopTimeoutMs); + const run = await startP2pRun({ + initiatorSession: sessionName, + targets: tokens.agents, + userText: tokens.cleanText, + fileContents, + serverLink, + rounds: p2pRounds, + extraPrompt: p2pExtraPrompt, + modeOverride: resolvedMode || undefined, + hopTimeoutMs: p2pHopTimeoutMs, + advancedPresetKey: p2pAdvancedPresetKey, + advancedRounds: p2pAdvancedRounds, + advancedRunTimeoutMs: p2pAdvancedRunTimeoutMinutes != null ? p2pAdvancedRunTimeoutMinutes * 60_000 : undefined, + contextReducer: p2pContextReducer, + }); const status = isLegacy ? 'accepted_legacy' : 'accepted'; timelineEmitter.emit(sessionName, 'command.ack', { commandId: effectiveId, status }); try { diff --git a/src/daemon/cron-executor.ts b/src/daemon/cron-executor.ts index 887771dd..ae305634 100644 --- a/src/daemon/cron-executor.ts +++ b/src/daemon/cron-executor.ts @@ -162,7 +162,14 @@ export async function executeCronJob(msg: CronDispatchMessage, serverLink: Serve } logger.info({ jobId, jobName, initiator: name, targets: targets.length, mode }, 'Cron: starting P2P discussion'); - const run = await startP2pRun(name, targets, topic, [], serverLink, rounds ?? 1); + const run = await startP2pRun({ + initiatorSession: name, + targets, + userText: topic, + fileContents: [], + serverLink, + rounds: rounds ?? 1, + }); // Link cron execution to P2P discussion so frontend can navigate try { serverLink.send({ type: 'cron.p2p_linked', jobId, discussionId: run.discussionId, runId: run.id }); diff --git a/src/daemon/p2p-orchestrator.ts b/src/daemon/p2p-orchestrator.ts index 4c335648..93df4b85 100644 --- a/src/daemon/p2p-orchestrator.ts +++ b/src/daemon/p2p-orchestrator.ts @@ -13,8 +13,17 @@ import { randomUUID } from 'node:crypto'; import { sendKeysDelayedEnter } from '../agent/tmux.js'; import { detectStatusAsync } from '../agent/detect.js'; import { getSession } from '../store/session-store.js'; -import { getTransportRuntime } from '../agent/session-manager.js'; +import { getTransportRuntime, launchTransportSession, stopTransportRuntimeSession } from '../agent/session-manager.js'; import { P2P_BASELINE_PROMPT, getP2pMode, getModeForRound, isComboMode, parseModePipeline, roundPrompt, type P2pMode } from '../../shared/p2p-modes.js'; +import { + resolveP2pRoundPlan, + type P2pAdvancedRound, + type P2pContextReducerConfig, + type P2pHelperDiagnostic, + type P2pParticipantSnapshotEntry, + type P2pResolvedPlan, + type P2pResolvedRound, +} from '../../shared/p2p-advanced.js'; import { formatP2pParticipantIdentity, shortP2pSessionName } from '../../shared/p2p-participant.js'; import { P2P_TERMINAL_HOP_STATUSES, @@ -41,6 +50,22 @@ export interface P2pTarget { mode: string; // mode key e.g. 'audit' } +export interface StartP2pRunOptions { + initiatorSession: string; + targets: P2pTarget[]; + userText: string; + fileContents: Array<{ path: string; content: string }>; + serverLink: ServerLink | null; + rounds?: number; + extraPrompt?: string; + modeOverride?: string; + hopTimeoutMs?: number; + advancedPresetKey?: string; + advancedRounds?: P2pAdvancedRound[]; + advancedRunTimeoutMs?: number; + contextReducer?: P2pContextReducerConfig; +} + interface P2pHopRuntime extends P2pHopProgress { section_header: string; artifact_path: string; @@ -56,6 +81,7 @@ export interface P2pRun { initiatorSession: string; currentTargetSession: string | null; finalReturnSession: string; + /** Compatibility-only projection for legacy consumers; advanced runs drain this per execution round. */ remainingTargets: P2pTarget[]; /** Total number of hop targets (excluding initiator phases). Fixed at creation time. */ totalTargets: number; @@ -69,9 +95,9 @@ export interface P2pRun { userText: string; timeoutMs: number; resultSummary: string | null; - /** Sessions whose hops completed successfully. */ + /** Compatibility-only projection for legacy consumers; advanced loop retries may repeat sessions here. */ completedHops: P2pTarget[]; - /** Sessions whose hops were skipped (timeout, sendKeys failure, idle without file change). */ + /** Compatibility-only projection for legacy consumers; advanced loop retries may repeat sessions here. */ skippedHops: string[]; error: string | null; createdAt: string; @@ -90,6 +116,26 @@ export interface P2pRun { /** Parallel hop runtime state across all rounds. */ hopStates: P2pHopRuntime[]; activeTargetSessions: string[]; + advancedP2pEnabled: boolean; + resolvedRounds?: P2pResolvedRound[]; + helperEligibleSnapshot: P2pParticipantSnapshotEntry[]; + contextReducer?: P2pContextReducerConfig; + advancedRunTimeoutMs?: number; + deadlineAt?: number | null; + currentRoundId?: string | null; + currentExecutionStep: number; + currentRoundAttempt: number; + roundAttemptCounts: Record; + roundJumpCounts: Record; + routingHistory: Array<{ + fromRoundId?: string | null; + toRoundId?: string | null; + trigger?: string | null; + atStep: number; + atAttempt?: number | null; + timestamp: number; + }>; + helperDiagnostics: P2pHelperDiagnostic[]; /** Internal: set to true when cancel requested */ _cancelled: boolean; } @@ -113,6 +159,13 @@ export function serializeP2pRun(run: P2pRun): P2pRunUpdatePayload { const currentHopState = activeHopStates[0] ?? null; const currentHop = currentHopState?.session ?? run.activeTargetSessions[0] ?? run.currentTargetSession; const hopCounts = countHopStates(run.hopStates); + const routingHistory = Array.isArray(run.routingHistory) ? run.routingHistory : []; + const latestStepByRoundId = routingHistory.reduce>((acc, entry) => { + if (typeof entry.toRoundId === 'string' && typeof entry.atStep === 'number') { + acc[entry.toRoundId] = entry.atStep; + } + return acc; + }, {}); return { id: run.id, @@ -187,6 +240,30 @@ export function serializeP2pRun(run: P2pRun): P2pRunUpdatePayload { terminal_reason: run.status === 'completed' || run.status === 'timed_out' || run.status === 'failed' || run.status === 'cancelled' ? run.status : null, + advanced_p2p_enabled: run.advancedP2pEnabled || undefined, + current_round_id: run.currentRoundId ?? null, + current_execution_step: run.currentExecutionStep || null, + current_round_attempt: run.currentRoundAttempt || null, + round_attempt_counts: run.advancedP2pEnabled ? { ...run.roundAttemptCounts } : undefined, + round_jump_counts: run.advancedP2pEnabled ? { ...run.roundJumpCounts } : undefined, + routing_history: run.advancedP2pEnabled ? [...routingHistory] : undefined, + helper_diagnostics: run.advancedP2pEnabled && run.helperDiagnostics.length > 0 ? [...run.helperDiagnostics] : undefined, + advanced_nodes: run.advancedP2pEnabled && run.resolvedRounds + ? run.resolvedRounds.map((round) => ({ + id: round.id, + title: round.title, + preset: round.preset, + status: (() => { + if (run.currentRoundId === round.id) { + return P2P_TERMINAL_RUN_STATUSES.has(run.status) ? (run.status === 'completed' ? 'done' : 'skipped') : 'active'; + } + if ((run.roundAttemptCounts[round.id] ?? 0) > 0) return 'done'; + return 'pending'; + })(), + attempt: run.roundAttemptCounts[round.id] ?? 0, + step: latestStepByRoundId[round.id], + })) + : undefined, // Full node list for segmented progress display — compatibility projection all_nodes: (() => { type NodeInfo = { @@ -275,6 +352,7 @@ let IDLE_POLL_MS = 3_000; let GRACE_PERIOD_DEFAULT_MS = 180_000; // 3 min — complex analysis (subagent research + write) takes time let MIN_PROCESSING_MS = 30_000; // Don't trust idle detection until 30s after dispatch let FILE_SETTLE_CYCLES = 3; // File must stop growing for 3 poll cycles (9s) to be "settled" +let ROUND_HOP_CLEANUP_DELAY_MS = 30_000; /** Override poll interval for tests. */ export function _setIdlePollMs(ms: number): void { IDLE_POLL_MS = ms; } @@ -284,6 +362,8 @@ export function _setGracePeriodMs(ms: number): void { GRACE_PERIOD_DEFAULT_MS = export function _setMinProcessingMs(ms: number): void { MIN_PROCESSING_MS = ms; } /** Override file settle cycles for tests. */ export function _setFileSettleCycles(n: number): void { FILE_SETTLE_CYCLES = n; } +/** Override round hop artifact cleanup delay for tests. */ +export function _setRoundHopCleanupDelayMs(ms: number): void { ROUND_HOP_CLEANUP_DELAY_MS = ms; } // ── Idle event registry (callback-driven, no polling) ───────────────────── @@ -344,19 +424,103 @@ function waitForIdleEvent(session: string, timeoutMs: number): IdleWaiterHandle // ── Start a P2P run ─────────────────────────────────────────────────────── -export async function startP2pRun( - initiatorSession: string, - targets: P2pTarget[], - userText: string, - fileContents: Array<{ path: string; content: string }>, - serverLink: ServerLink | null, - rounds?: number, - extraPrompt?: string, - /** Explicit mode override — used for combo pipelines (e.g. "brainstorm>discuss>plan"). */ - modeOverride?: string, - /** Custom per-hop timeout in ms. Overrides mode default (300s). */ - hopTimeoutMs?: number, +function buildHelperEligibleSnapshot(initiatorSession: string, targets: P2pTarget[]): P2pParticipantSnapshotEntry[] { + const seen = new Set(); + const names = [initiatorSession, ...targets.map((target) => target.session)]; + const snapshot: P2pParticipantSnapshotEntry[] = []; + for (const sessionName of names) { + if (seen.has(sessionName)) continue; + seen.add(sessionName); + const record = getSession(sessionName); + snapshot.push({ + sessionName, + agentType: record?.agentType ?? 'unknown', + parentSession: record?.parentSession ?? null, + }); + } + return snapshot; +} + +function normalizeStartP2pRunArgs( + args: [ + StartP2pRunOptions + ] | [ + string, + P2pTarget[], + string, + Array<{ path: string; content: string }>, + ServerLink | null, + number | undefined, + string | undefined, + string | undefined, + number | undefined, + ], +): StartP2pRunOptions { + if (args.length === 1 && typeof args[0] === 'object' && args[0] !== null && 'initiatorSession' in args[0]) { + return args[0] as StartP2pRunOptions; + } + const [ + initiatorSession, + targets, + userText, + fileContents, + serverLink, + rounds, + extraPrompt, + modeOverride, + hopTimeoutMs, + ] = args as [ + string, + P2pTarget[], + string, + Array<{ path: string; content: string }>, + ServerLink | null, + number | undefined, + string | undefined, + string | undefined, + number | undefined, + ]; + return { + initiatorSession, + targets, + userText, + fileContents, + serverLink, + rounds, + extraPrompt, + modeOverride, + hopTimeoutMs, + }; +} + +export async function startP2pRun(...args: + [StartP2pRunOptions] | [ + string, + P2pTarget[], + string, + Array<{ path: string; content: string }>, + ServerLink | null, + number | undefined, + string | undefined, + string | undefined, + number | undefined, + ] ): Promise { + const { + initiatorSession, + targets, + userText, + fileContents, + serverLink, + rounds, + extraPrompt, + modeOverride, + hopTimeoutMs, + advancedPresetKey, + advancedRounds, + advancedRunTimeoutMs, + contextReducer, + } = normalizeStartP2pRunArgs(args); // Validate same domain const mainSession = extractMainSession(initiatorSession); for (const t of targets) { @@ -365,6 +529,17 @@ export async function startP2pRun( } } + const helperEligibleSnapshot = buildHelperEligibleSnapshot(initiatorSession, targets); + const resolvedPlan: P2pResolvedPlan = resolveP2pRoundPlan({ + modeOverride: modeOverride ?? targets[0]?.mode ?? 'discuss', + roundsOverride: rounds, + hopTimeoutMinutes: hopTimeoutMs != null ? Math.ceil(hopTimeoutMs / 60_000) : undefined, + advancedPresetKey, + advancedRounds, + advancedRunTimeoutMinutes: advancedRunTimeoutMs != null ? Math.ceil(advancedRunTimeoutMs / 60_000) : undefined, + contextReducer, + participants: helperEligibleSnapshot, + }); const mode = modeOverride ?? targets[0]?.mode ?? 'discuss'; const modeConfig = getP2pMode(isComboMode(mode) ? parseModePipeline(mode)[0] : mode); const runId = randomUUID().slice(0, 12); @@ -394,7 +569,9 @@ export async function startP2pRun( await writeFile(contextFilePath, seed, 'utf8'); const P2P_MAX_ROUNDS = 6; - const totalRounds = Math.min(P2P_MAX_ROUNDS, Math.max(1, rounds ?? 1)); + const totalRounds = resolvedPlan.advanced + ? resolvedPlan.rounds.length + : Math.min(P2P_MAX_ROUNDS, Math.max(1, rounds ?? 1)); const run: P2pRun = { id: runId, discussionId, @@ -426,6 +603,23 @@ export async function startP2pRun( hopStartedAt: Date.now(), hopStates: [], activeTargetSessions: [], + advancedP2pEnabled: resolvedPlan.advanced, + resolvedRounds: resolvedPlan.advanced ? resolvedPlan.rounds : undefined, + helperEligibleSnapshot: resolvedPlan.helperEligibleSnapshot ?? helperEligibleSnapshot, + contextReducer: resolvedPlan.contextReducer, + advancedRunTimeoutMs: resolvedPlan.advanced && resolvedPlan.overallRunTimeoutMinutes != null + ? resolvedPlan.overallRunTimeoutMinutes * 60_000 + : undefined, + deadlineAt: resolvedPlan.advanced && resolvedPlan.overallRunTimeoutMinutes != null + ? Date.now() + (resolvedPlan.overallRunTimeoutMinutes * 60_000) + : null, + currentRoundId: resolvedPlan.advanced ? (resolvedPlan.rounds[0]?.id ?? null) : null, + currentExecutionStep: 0, + currentRoundAttempt: 1, + roundAttemptCounts: {}, + roundJumpCounts: {}, + routingHistory: [], + helperDiagnostics: [], _cancelled: false, }; @@ -625,6 +819,11 @@ async function cleanupRoundHopArtifacts(roundHops: P2pHopRuntime[]): Promise { void cleanupRoundHopArtifacts(roundHops); }, ROUND_HOP_CLEANUP_DELAY_MS); +} + function updateHopStatus(run: P2pRun, hop: P2pHopRuntime | null | undefined, status: P2pHopStatus, error: string | null = null): void { if (!hop) return; hop.status = status; @@ -639,6 +838,10 @@ function updateHopStatus(run: P2pRun, hop: P2pHopRuntime | null | undefined, sta } async function executeChain(run: P2pRun, modeConfig: P2pMode | undefined, serverLink: ServerLink | null): Promise { + if (run.advancedP2pEnabled && run.resolvedRounds?.length) { + await executeAdvancedChain(run, serverLink); + return; + } const totalHops = run.allTargets.length; // ── Multi-round loop ── @@ -681,76 +884,75 @@ async function executeChain(run: P2pRun, modeConfig: P2pMode | undefined, server // ── Phase 2: Sub-session hops ── run.activePhase = 'hop'; const roundHops = await createRoundHopStates(run, targets, roundModeKey); - run.activeTargetSessions = roundHops.map((hop) => hop.session); - const hopResults = await Promise.allSettled(targets.map(async (target, i) => { - if (run._cancelled) return false; - const hop = roundHops[i]; - const hopMode = combo ? roundModeKey : target.mode; - const hopLabel = `${discussionParticipantName(target.session)} — ${capitalize(hopMode)} (hop ${i + 1}/${totalHops}${roundLabel})`; - hop.section_header = hopLabel; - const hopModeConfig = combo ? roundModeConfig : (getP2pMode(target.mode) ?? modeConfig); - const hopPrompt = buildHopPrompt(run, hopModeConfig, { - session: target.session, - sectionHeader: hopLabel, - instruction: `Read the discussion file and provide your ${hopMode} analysis. Append your output to the file.\nIMPORTANT: This is ANALYSIS ONLY. Do NOT implement fixes, do NOT edit code files, do NOT run commands. Only write your analysis into this discussion file.`, + try { + run.activeTargetSessions = roundHops.map((hop) => hop.session); + const hopResults = await Promise.allSettled(targets.map(async (target, i) => { + if (run._cancelled) return false; + const hop = roundHops[i]; + const hopMode = combo ? roundModeKey : target.mode; + const hopLabel = `${discussionParticipantName(target.session)} — ${capitalize(hopMode)} (hop ${i + 1}/${totalHops}${roundLabel})`; + hop.section_header = hopLabel; + const hopModeConfig = combo ? roundModeConfig : (getP2pMode(target.mode) ?? modeConfig); + const hopPrompt = buildHopPrompt(run, hopModeConfig, { + session: target.session, + sectionHeader: hopLabel, + instruction: `Read the discussion file and provide your ${hopMode} analysis. Append your output to the file.\nIMPORTANT: This is ANALYSIS ONLY. Do NOT implement fixes, do NOT edit code files, do NOT run commands. Only write your analysis into this discussion file.`, + isInitial: false, + filePath: hop.artifact_path, + }, rp); + logger.info({ runId: run.id, target: target.session, mode: hopMode, hop: i + 1, totalHops, round: run.currentRound }, 'P2P: Phase 2 — dispatching hop'); + return dispatchHop(run, target.session, hopPrompt, serverLink, { + sectionHeader: hopLabel, + hop, + filePath: hop.artifact_path, + }); + })); + run.activeTargetSessions = []; + run.currentTargetSession = null; + if (run._cancelled || isTerminal(run.status)) return; + logger.info({ + runId: run.id, + round: run.currentRound, + settled: hopResults.length, + completed: roundHops.filter((hop) => hop.status === 'completed').length, + }, 'P2P: Phase 2 — round barrier settled'); + await appendRoundEvidence(run, roundHops); + if (run._cancelled || isTerminal(run.status)) return; + + run.remainingTargets = []; + + // ── Round summary: Initiator synthesizes this round ── + if (run._cancelled) return; + run.runPhase = 'summarizing'; + run.summaryPhase = 'running'; + run.activePhase = 'summary'; + const isLastRound = run.currentRound === run.rounds; + const summaryModeConfig = isLastRound && combo + ? getModeForRound(run.mode, run.rounds) // last pipeline mode for final summary + : roundModeConfig; + const roundSummaryHeader = isLastRound + ? `${discussionParticipantNameWithMode(run.initiatorSession, roundModeKey)} — Final Summary` + : `${discussionParticipantNameWithMode(run.initiatorSession, roundModeKey)} — Round ${run.currentRound}/${run.rounds} Summary`; + const roundSummaryInstruction = isLastRound + ? `${summaryModeConfig?.summaryPrompt ?? 'Synthesize a final summary that captures the consensus, key decisions, and any remaining disagreements across all rounds.'}\nBefore writing the summary, use the hop evidence already appended into the discussion file for this round. If the user context clearly specifies a destination file for the final plan, write the complete plan there. Otherwise, write the complete plan at the end of the discussion file.` + : `Synthesize the key points, areas of agreement, and open questions from this round. Then assign specific focus areas or questions for each participant in the next round (round ${run.currentRound + 1}). Append to the file.\nIMPORTANT: This is ANALYSIS ONLY. Do NOT implement fixes, do NOT edit code files, do NOT run commands. Only write your analysis into this discussion file.`; + const roundSummaryPrompt = buildHopPrompt(run, summaryModeConfig, { + session: run.initiatorSession, + sectionHeader: roundSummaryHeader, + instruction: `${roundSummaryInstruction}\nThe orchestrator has already appended each completed hop's evidence into the discussion file. If you write the final plan to another file, still append a short completion note under the new final-summary heading in the discussion file that records the chosen output file path.`, isInitial: false, - filePath: hop.artifact_path, }, rp); - logger.info({ runId: run.id, target: target.session, mode: hopMode, hop: i + 1, totalHops, round: run.currentRound }, 'P2P: Phase 2 — dispatching hop'); - return dispatchHop(run, target.session, hopPrompt, serverLink, { - sectionHeader: hopLabel, - hop, - filePath: hop.artifact_path, + logger.info({ runId: run.id, round: run.currentRound, isLastRound, roundMode: roundModeKey }, isLastRound ? 'P2P: Final summary — initiator' : 'P2P: Round summary — initiator'); + const summaryOk = await dispatchHop(run, run.initiatorSession, roundSummaryPrompt, serverLink, { + sectionHeader: roundSummaryHeader, + required: true, }); - })); - run.activeTargetSessions = []; - run.currentTargetSession = null; - if (run._cancelled || isTerminal(run.status)) return; - logger.info({ - runId: run.id, - round: run.currentRound, - settled: hopResults.length, - completed: roundHops.filter((hop) => hop.status === 'completed').length, - }, 'P2P: Phase 2 — round barrier settled'); - await appendRoundEvidence(run, roundHops); - if (run._cancelled || isTerminal(run.status)) return; - - if (run.currentRound === run.rounds) { - run.remainingTargets = []; - } else { - run.remainingTargets = []; + if (!summaryOk) return; + run.summaryPhase = 'completed'; + if (run._cancelled || isTerminal(run.status)) return; + } finally { + scheduleRoundHopArtifactCleanup(roundHops); } - - // ── Round summary: Initiator synthesizes this round ── - if (run._cancelled) return; - run.runPhase = 'summarizing'; - run.summaryPhase = 'running'; - run.activePhase = 'summary'; - const isLastRound = run.currentRound === run.rounds; - const summaryModeConfig = isLastRound && combo - ? getModeForRound(run.mode, run.rounds) // last pipeline mode for final summary - : roundModeConfig; - const roundSummaryHeader = isLastRound - ? `${discussionParticipantNameWithMode(run.initiatorSession, roundModeKey)} — Final Summary` - : `${discussionParticipantNameWithMode(run.initiatorSession, roundModeKey)} — Round ${run.currentRound}/${run.rounds} Summary`; - const roundSummaryInstruction = isLastRound - ? `${summaryModeConfig?.summaryPrompt ?? 'Synthesize a final summary that captures the consensus, key decisions, and any remaining disagreements across all rounds.'}\nBefore writing the summary, use the hop evidence already appended into the discussion file for this round. If the user context clearly specifies a destination file for the final plan, write the complete plan there. Otherwise, write the complete plan at the end of the discussion file.` - : `Synthesize the key points, areas of agreement, and open questions from this round. Then assign specific focus areas or questions for each participant in the next round (round ${run.currentRound + 1}). Append to the file.\nIMPORTANT: This is ANALYSIS ONLY. Do NOT implement fixes, do NOT edit code files, do NOT run commands. Only write your analysis into this discussion file.`; - const roundSummaryPrompt = buildHopPrompt(run, summaryModeConfig, { - session: run.initiatorSession, - sectionHeader: roundSummaryHeader, - instruction: `${roundSummaryInstruction}\nThe orchestrator has already appended each completed hop's evidence into the discussion file. If you write the final plan to another file, still append a short completion note under the new final-summary heading in the discussion file that records the chosen output file path.`, - isInitial: false, - }, rp); - logger.info({ runId: run.id, round: run.currentRound, isLastRound, roundMode: roundModeKey }, isLastRound ? 'P2P: Final summary — initiator' : 'P2P: Round summary — initiator'); - const summaryOk = await dispatchHop(run, run.initiatorSession, roundSummaryPrompt, serverLink, { - sectionHeader: roundSummaryHeader, - required: true, - }); - if (!summaryOk) return; - run.summaryPhase = 'completed'; - setTimeout(() => { void cleanupRoundHopArtifacts(roundHops); }, 30_000); - if (run._cancelled || isTerminal(run.status)) return; } if (run._cancelled || isTerminal(run.status)) return; @@ -782,6 +984,487 @@ async function executeChain(run: P2pRun, modeConfig: P2pMode | undefined, server }, 60_000); } +function addHelperDiagnostic(run: P2pRun, diagnostic: Omit): void { + run.helperDiagnostics.push({ ...diagnostic, timestamp: Date.now() }); +} + +function parseVerdictFromContent(content: string): 'PASS' | 'REWORK' | null { + const matches = [...content.matchAll(//g)]; + const verdict = matches.at(-1)?.[1]; + return verdict === 'PASS' || verdict === 'REWORK' ? verdict : null; +} + +function helperFallbackCandidates(run: P2pRun, exclude: string[]): string[] { + const excluded = new Set(exclude); + return run.helperEligibleSnapshot + .filter((entry) => entry.sessionName !== run.initiatorSession) + .filter((entry) => !!entry.parentSession || entry.sessionName.startsWith('deck_sub_')) + .filter((entry) => !excluded.has(entry.sessionName)) + .map((entry) => entry.sessionName); +} + +function ensureRunDeadline(run: P2pRun, serverLink: ServerLink | null): boolean { + if (!run.advancedRunTimeoutMs || !run.deadlineAt) return true; + if (Date.now() <= run.deadlineAt) return true; + failRun(run, 'timed_out', 'advanced_run_timeout', serverLink); + return false; +} + +async function launchClonedHelperSession(run: P2pRun, templateSession: string): Promise { + const template = getSession(templateSession); + if (!template || !template.runtimeType || template.runtimeType !== 'transport') { + throw new Error(`Helper template is not an eligible SDK transport session: ${templateSession}`); + } + const helperName = `deck_p2p_helper_${run.id}_${run.currentExecutionStep}_${randomUUID().slice(0, 6)}`; + await launchTransportSession({ + name: helperName, + projectName: template.projectName, + role: 'w1', + agentType: template.agentType as any, + projectDir: template.projectDir, + skipStore: true, + fresh: true, + requestedModel: template.requestedModel ?? template.activeModel ?? undefined, + transportConfig: template.transportConfig ?? undefined, + description: `P2P helper for ${run.id}`, + label: helperName, + effort: template.effort, + ...(template.ccPreset ? { ccPreset: template.ccPreset } : {}), + }); + return helperName; +} + +async function teardownHelperSession(run: P2pRun, sessionName: string): Promise { + try { + await stopTransportRuntimeSession(sessionName); + } catch (err) { + addHelperDiagnostic(run, { + code: 'P2P_HELPER_CLEANUP_FAILED', + attempt: run.currentRoundAttempt, + sourceSession: sessionName, + message: err instanceof Error ? err.message : String(err), + }); + } +} + +async function cleanupReducerSummaryFile( + run: P2pRun, + summaryPath: string, + sourceSession?: string | null, + templateSession?: string | null, + fallbackSession?: string | null, +): Promise { + try { + await unlink(summaryPath); + } catch (err) { + if ((err as NodeJS.ErrnoException | undefined)?.code === 'ENOENT') return; + addHelperDiagnostic(run, { + code: 'P2P_HELPER_CLEANUP_FAILED', + attempt: run.currentRoundAttempt, + sourceSession: sourceSession ?? null, + templateSession: templateSession ?? null, + fallbackSession: fallbackSession ?? null, + message: err instanceof Error ? err.message : String(err), + }); + } +} + +async function reduceAdvancedContext( + run: P2pRun, + round: P2pResolvedRound, + serverLink: ServerLink | null, +): Promise { + if (!run.contextReducer) return null; + const info = await stat(run.contextFilePath).catch(() => null); + if (!info || info.size < 32_000) return null; + + const summaryPath = join(dirname(run.contextFilePath), `${run.id}.reducer.${run.currentExecutionStep}.md`); + const sectionHeader = `P2P Helper Summary — ${round.title} (step ${run.currentExecutionStep})`; + const reducerPrompt = [ + `[P2P Helper Task — ${run.id}]`, + `Read the discussion file at ${run.contextFilePath}.`, + `Produce a compact context reduction for the next round.`, + `Focus on: latest implementation attempt, latest audit findings, declared artifact targets, and only the most relevant unresolved issues.`, + `Write the result to ${summaryPath}.`, + `Add a new heading "## ${sectionHeader}" and put the reduced context under it.`, + `Do not change workflow verdicts, do not route, and do not edit any code files.`, + `Start immediately.`, + ].join('\n'); + + const readReducedSummary = async () => { + const content = await readFile(summaryPath, 'utf8').catch(() => ''); + const section = extractHeadingSection(content, sectionHeader) ?? content; + return section.trim() ? section.trim() : null; + }; + + const attemptWithSession = async (sessionName: string, codeOnFailure: P2pHelperDiagnostic['code'], templateSession?: string | null) => { + const ok = await dispatchHop(run, sessionName, reducerPrompt, serverLink, { + sectionHeader, + filePath: summaryPath, + required: false, + }); + if (!ok) { + addHelperDiagnostic(run, { + code: codeOnFailure, + attempt: run.currentRoundAttempt, + sourceSession: sessionName, + templateSession: templateSession ?? null, + }); + return null; + } + return readReducedSummary(); + }; + + await writeFile(summaryPath, '# Helper Summary\n\n', 'utf8'); + try { + if (run.contextReducer.mode === 'reuse_existing_session' && run.contextReducer.sessionName) { + const primaryResult = await attemptWithSession( + run.contextReducer.sessionName, + 'P2P_HELPER_PRIMARY_FAILED', + run.contextReducer.sessionName, + ); + if (primaryResult) return primaryResult; + } else if (run.contextReducer.mode === 'clone_sdk_session' && run.contextReducer.templateSession) { + let helperName: string | null = null; + try { + helperName = await launchClonedHelperSession(run, run.contextReducer.templateSession); + const primaryResult = await attemptWithSession( + helperName, + 'P2P_HELPER_PRIMARY_FAILED', + run.contextReducer.templateSession, + ); + if (primaryResult) return primaryResult; + } catch (err) { + addHelperDiagnostic(run, { + code: 'P2P_HELPER_PRIMARY_FAILED', + attempt: run.currentRoundAttempt, + templateSession: run.contextReducer.templateSession, + message: err instanceof Error ? err.message : String(err), + }); + } finally { + if (helperName) await teardownHelperSession(run, helperName); + } + } + + const fallbackSession = helperFallbackCandidates(run, [ + run.contextReducer.sessionName ?? '', + run.contextReducer.templateSession ?? '', + ])[0]; + if (!fallbackSession) { + addHelperDiagnostic(run, { + code: 'P2P_COMPRESSION_SKIPPED_NO_FALLBACK', + attempt: run.currentRoundAttempt, + templateSession: run.contextReducer.templateSession ?? null, + sourceSession: run.contextReducer.sessionName ?? null, + }); + return null; + } + const fallbackResult = await attemptWithSession(fallbackSession, 'P2P_HELPER_FALLBACK_FAILED', fallbackSession); + if (fallbackResult) return fallbackResult; + failRun(run, 'failed', `helper_fallback_failed:${fallbackSession}`, serverLink); + return null; + } finally { + await cleanupReducerSummaryFile( + run, + summaryPath, + run.contextReducer.sessionName ?? null, + run.contextReducer.templateSession ?? null, + ); + } +} + +async function captureArtifactBaseline(run: P2pRun, round: P2pResolvedRound): Promise> { + const baseline = new Map(); + const record = getSession(run.initiatorSession); + const projectDir = record?.projectDir ?? process.cwd(); + if (round.artifactConvention === 'openspec_convention') { + const target = join(projectDir, 'openspec', 'changes'); + try { + const entries = await readdir(target); + baseline.set(target, entries.join('\n')); + } catch { + baseline.set(target, null); + } + return baseline; + } + for (const output of round.artifactOutputs) { + const absPath = join(projectDir, output); + try { + baseline.set(absPath, await readFile(absPath, 'utf8')); + } catch { + baseline.set(absPath, null); + } + } + return baseline; +} + +async function validateArtifactOutputsForRound(run: P2pRun, round: P2pResolvedRound, baseline: Map): Promise { + if (round.artifactConvention === 'none') return; + if (round.artifactConvention === 'openspec_convention') { + const target = [...baseline.keys()][0]; + const before = baseline.get(target) ?? null; + try { + const afterEntries = (await readdir(target)).join('\n'); + if (afterEntries === before) throw new Error('openspec_convention artifacts were not observably updated'); + return; + } catch (err) { + throw new Error(err instanceof Error ? err.message : String(err)); + } + } + for (const [absPath, before] of baseline.entries()) { + let after: string | null = null; + try { after = await readFile(absPath, 'utf8'); } catch { after = null; } + if (after == null) throw new Error(`Expected artifact missing after round: ${absPath}`); + if (after === before) throw new Error(`Expected artifact not observably updated: ${absPath}`); + } +} + +async function readAppendedContent(filePath: string, baselineSize: number): Promise { + const buffer = await readFile(filePath); + return buffer.subarray(Math.min(buffer.length, baselineSize)).toString('utf8'); +} + +function buildAdvancedRoundPrefix(run: P2pRun, round: P2pResolvedRound): string { + return `[Advanced Round ${run.currentExecutionStep} — ${round.title} — Attempt ${run.currentRoundAttempt}]`; +} + +function buildAdvancedPromptCommon( + run: P2pRun, + round: P2pResolvedRound, + targetSession: string, + filePath: string, + sectionHeader: string, + reducerSummary: string | null, + instruction: string, +): string { + const parts: string[] = []; + parts.push(buildAdvancedRoundPrefix(run, round)); + parts.push(''); + parts.push(P2P_BASELINE_PROMPT); + if (round.presetPrompt) parts.push(round.presetPrompt); + parts.push(''); + parts.push(`[P2P Advanced Task — run ${run.id}]`); + parts.push(`Discussion file: ${filePath}`); + parts.push(`Your identity for this round is "${discussionParticipantName(targetSession)}".`); + parts.push(`Round id: ${round.id}`); + parts.push(`Permission scope: ${round.permissionScope}`); + if (round.artifactConvention === 'openspec_convention') { + parts.push('Required artifact contract: write OpenSpec artifacts under repository OpenSpec conventions inside openspec/changes/.'); + } else if (round.artifactOutputs.length > 0) { + parts.push(`Required artifact outputs: ${round.artifactOutputs.join(', ')}`); + } + if (reducerSummary) { + parts.push(''); + parts.push('Reduced context for this attempt:'); + parts.push(reducerSummary); + } + if (round.promptAppend) { + parts.push(''); + parts.push(`Additional round instructions: ${round.promptAppend}`); + } + parts.push(''); + parts.push(instruction); + parts.push(`Add a new heading "## ${sectionHeader}" at the end of the discussion file and write your result below it.`); + parts.push('Do not ask for confirmation. Start immediately.'); + if (run.extraPrompt) { + parts.push(''); + parts.push(`Additional instructions: ${run.extraPrompt}`); + } + return parts.join('\n'); +} + +function buildAdvancedHopPrompt( + run: P2pRun, + round: P2pResolvedRound, + target: P2pTarget, + filePath: string, + sectionHeader: string, + reducerSummary: string | null, +): string { + const instruction = round.permissionScope === 'analysis_only' + ? 'Read the discussion file and provide analysis only. Do not edit code or other files.' + : round.permissionScope === 'artifact_generation' + ? 'Read the discussion file and produce the required artifacts. You may write only the round outputs and the discussion note for this round.' + : 'Read the discussion file and perform the implementation work required by this round. You may edit code and tests as needed, then append a concise execution note to the discussion file.'; + return buildAdvancedPromptCommon(run, round, target.session, filePath, sectionHeader, reducerSummary, instruction); +} + +function buildAdvancedSynthesisPrompt( + run: P2pRun, + round: P2pResolvedRound, + sectionHeader: string, + reducerSummary: string | null, +): string { + const instruction = round.summaryPrompt + ?? 'Synthesize the evidence appended in this round into one authoritative summary.'; + return buildAdvancedPromptCommon( + run, + round, + run.initiatorSession, + run.contextFilePath, + sectionHeader, + reducerSummary, + instruction, + ); +} + +async function executeAdvancedChain(run: P2pRun, serverLink: ServerLink | null): Promise { + const rounds = run.resolvedRounds ?? []; + let roundIndex = 0; + while (roundIndex < rounds.length) { + if (run._cancelled || isTerminal(run.status)) return; + if (!ensureRunDeadline(run, serverLink)) return; + const round = rounds[roundIndex]; + run.timeoutMs = round.timeoutMs; + run.currentRound = roundIndex + 1; + run.currentRoundId = round.id; + run.currentExecutionStep += 1; + run.roundAttemptCounts[round.id] = (run.roundAttemptCounts[round.id] ?? 0) + 1; + run.currentRoundAttempt = run.roundAttemptCounts[round.id]; + run.runPhase = 'round_execution'; + run.summaryPhase = null; + run.activePhase = round.dispatchStyle === 'initiator_only' ? 'initial' : 'hop'; + pushState(run, serverLink); + + const artifactBaseline = await captureArtifactBaseline(run, round); + const reducerSummary = await reduceAdvancedContext(run, round, serverLink); + if (run._cancelled || isTerminal(run.status)) return; + + let authoritativeSegment = ''; + if (round.dispatchStyle === 'initiator_only') { + const sectionHeader = `${discussionParticipantName(run.initiatorSession)} — ${round.title} (attempt ${run.currentRoundAttempt})`; + const baselineBuffer = await readFile(run.contextFilePath).catch(() => Buffer.from('')); + const prompt = buildAdvancedHopPrompt( + run, + round, + { session: run.initiatorSession, mode: round.modeKey }, + run.contextFilePath, + sectionHeader, + reducerSummary, + ); + const ok = await dispatchHop(run, run.initiatorSession, prompt, serverLink, { + sectionHeader, + required: true, + }); + if (!ok) return; + authoritativeSegment = await readAppendedContent(run.contextFilePath, baselineBuffer.length); + } else { + const targets = [...run.allTargets]; + const roundHops = await createRoundHopStates(run, targets, round.modeKey); + try { + run.activeTargetSessions = roundHops.map((hop) => hop.session); + await Promise.allSettled(targets.map(async (target, index) => { + const hop = roundHops[index]; + const sectionHeader = `${discussionParticipantName(target.session)} — ${round.title} (hop ${index + 1}/${targets.length}, attempt ${run.currentRoundAttempt})`; + hop.section_header = sectionHeader; + const prompt = buildAdvancedHopPrompt(run, round, target, hop.artifact_path, sectionHeader, reducerSummary); + return dispatchHop(run, target.session, prompt, serverLink, { + sectionHeader, + hop, + filePath: hop.artifact_path, + }); + })); + run.activeTargetSessions = []; + run.currentTargetSession = null; + if (run._cancelled || isTerminal(run.status)) return; + await appendRoundEvidence(run, roundHops); + if (round.synthesisStyle === 'initiator_summary') { + run.runPhase = 'summarizing'; + run.summaryPhase = 'running'; + run.activePhase = 'summary'; + const sectionHeader = `${discussionParticipantName(run.initiatorSession)} — ${round.title} Synthesis (attempt ${run.currentRoundAttempt})`; + const baselineBuffer = await readFile(run.contextFilePath).catch(() => Buffer.from('')); + const prompt = buildAdvancedSynthesisPrompt(run, round, sectionHeader, reducerSummary); + const ok = await dispatchHop(run, run.initiatorSession, prompt, serverLink, { + sectionHeader, + required: true, + }); + if (!ok) return; + authoritativeSegment = await readAppendedContent(run.contextFilePath, baselineBuffer.length); + run.summaryPhase = 'completed'; + } + } finally { + scheduleRoundHopArtifactCleanup(roundHops); + } + } + + await validateArtifactOutputsForRound(run, round, artifactBaseline).catch((err) => { + failRun(run, 'failed', err instanceof Error ? err.message : String(err), serverLink); + }); + if (run._cancelled || isTerminal(run.status)) return; + + const verdict = round.requiresVerdict ? parseVerdictFromContent(authoritativeSegment) : null; + const effectiveVerdict = round.requiresVerdict + ? (verdict ?? (() => { + addHelperDiagnostic(run, { + code: 'P2P_VERDICT_MISSING', + attempt: run.currentRoundAttempt, + sourceSession: run.initiatorSession, + message: `Missing verdict marker for round ${round.id}`, + }); + return 'REWORK' as const; + })()) + : null; + + const jump = round.allowRouting && round.jumpRule + ? (() => { + const jumpCount = run.roundJumpCounts[round.id] ?? 0; + const belowMax = jumpCount < round.jumpRule!.maxTriggers; + if (!belowMax) return null; + if (round.verdictPolicy === 'forced_rework') { + if (jumpCount < round.jumpRule.minTriggers) return round.jumpRule.targetRoundId; + return effectiveVerdict === (round.jumpRule.marker ?? 'REWORK') ? round.jumpRule.targetRoundId : null; + } + return effectiveVerdict === (round.jumpRule.marker ?? 'REWORK') ? round.jumpRule.targetRoundId : null; + })() + : null; + + if (jump) { + run.roundJumpCounts[round.id] = (run.roundJumpCounts[round.id] ?? 0) + 1; + run.routingHistory.push({ + fromRoundId: round.id, + toRoundId: jump, + trigger: effectiveVerdict, + atStep: run.currentExecutionStep, + atAttempt: run.currentRoundAttempt, + timestamp: Date.now(), + }); + roundIndex = rounds.findIndex((entry) => entry.id === jump); + continue; + } + + roundIndex += 1; + } + + if (!ensureRunDeadline(run, serverLink) || run._cancelled || isTerminal(run.status)) return; + run.runPhase = 'summarizing'; + run.summaryPhase = 'running'; + run.activePhase = 'summary'; + const finalRound = rounds[Math.max(rounds.length - 1, 0)]; + run.timeoutMs = finalRound?.timeoutMs ?? run.timeoutMs; + const finalPrompt = buildHopPrompt(run, getP2pMode(finalRound?.modeKey ?? run.mode), { + session: run.initiatorSession, + sectionHeader: `${discussionParticipantNameWithMode(run.initiatorSession, finalRound?.modeKey ?? run.mode)} — Final Summary`, + instruction: `${getP2pMode(finalRound?.modeKey ?? run.mode)?.summaryPrompt ?? 'Synthesize a final summary that captures the consensus, key decisions, and any remaining disagreements across all rounds.'}\nBefore writing the summary, use the hop evidence already appended into the discussion file for this round. If the user context clearly specifies a destination file for the final plan, write the complete plan there. Otherwise, write the complete plan at the end of the discussion file.`, + isInitial: false, + }); + const summaryOk = await dispatchHop(run, run.initiatorSession, finalPrompt, serverLink, { + sectionHeader: `${discussionParticipantNameWithMode(run.initiatorSession, finalRound?.modeKey ?? run.mode)} — Final Summary`, + required: true, + }); + if (!summaryOk) return; + run.summaryPhase = 'completed'; + + let fullContent = ''; + try { + fullContent = await readFile(run.contextFilePath, 'utf8'); + run.resultSummary = fullContent.slice(-2000); + } catch { /* ignore */ } + run.completedAt = new Date().toISOString(); + transition(run, 'completed', serverLink); + setTimeout(() => { activeRuns.delete(run.id); }, 60_000); +} + // ── Single hop dispatch + wait ──────────────────────────────────────────── interface DispatchHopOptions { @@ -852,6 +1535,15 @@ async function dispatchHop( } }; + const abortForOverallTimeout = async () => { + if (hop) { + await finishHop('timed_out', 'advanced_run_timeout'); + } else { + run.currentTargetSession = null; + run.activeTargetSessions = run.activeTargetSessions.filter((item) => item !== session); + } + }; + for (let attempt = 0; attempt <= MAX_RETRIES; attempt++) { if (run._cancelled) { await finishHop('cancelled'); @@ -898,6 +1590,11 @@ async function dispatchHop( let headingFoundAt = 0; while (Date.now() < deadline) { + if (!ensureRunDeadline(run, serverLink)) { + idleWaiter.cancel(); + await abortForOverallTimeout(); + return false; + } if (Date.now() >= hardDeadline) { logger.warn({ runId: run.id, session }, 'P2P: hard deadline reached, force-skipping hop'); break; @@ -908,6 +1605,11 @@ async function dispatchHop( return false; } await sleep(IDLE_POLL_MS); + if (!ensureRunDeadline(run, serverLink)) { + idleWaiter.cancel(); + await abortForOverallTimeout(); + return false; + } if (Date.now() >= hardDeadline) { logger.warn({ runId: run.id, session }, 'P2P: hard deadline reached, force-skipping hop'); break; @@ -1126,11 +1828,16 @@ function transition(run: P2pRun, status: P2pRunStatus, serverLink: ServerLink | run.summaryPhase = 'completed'; } else if (status === 'cancelled') { run.runPhase = 'cancelled'; - } else if (status === 'failed' || status === 'timed_out') { + } else if (status === 'failed') { run.runPhase = 'failed'; } if (P2P_TERMINAL_RUN_STATUSES.has(status)) { run.completedAt = run.completedAt ?? new Date().toISOString(); + if (run.advancedP2pEnabled) { + void cleanupRoundHopArtifacts(run.hopStates); + } else { + scheduleRoundHopArtifactCleanup(run.hopStates); + } } run.updatedAt = new Date().toISOString(); logger.info({ runId: run.id, status }, 'P2P run state transition'); @@ -1143,8 +1850,17 @@ function failRun(run: P2pRun, errorType: string, message: string, serverLink: Se run.updatedAt = new Date().toISOString(); const status: P2pRunStatus = errorType === 'timed_out' ? 'timed_out' : 'failed'; run.status = status; - run.runPhase = 'failed'; - if (run.activePhase === 'summary') run.summaryPhase = 'failed'; + if (status === 'failed') { + run.runPhase = 'failed'; + } + if (run.activePhase === 'summary') { + run.summaryPhase = 'failed'; + } + if (run.advancedP2pEnabled) { + void cleanupRoundHopArtifacts(run.hopStates); + } else { + scheduleRoundHopArtifactCleanup(run.hopStates); + } logger.warn({ runId: run.id, errorType, message }, 'P2P run failed'); pushState(run, serverLink); } diff --git a/test/daemon/cron-executor.test.ts b/test/daemon/cron-executor.test.ts index c0bfe1a0..c6102150 100644 --- a/test/daemon/cron-executor.test.ts +++ b/test/daemon/cron-executor.test.ts @@ -286,17 +286,18 @@ describe('executeCronJob', () => { await executeCronJob(msg, mockServerLink); - expect(startP2pRun).toHaveBeenCalledWith( - 'deck_myapp_brain', - [ + expect((startP2pRun as ReturnType).mock.calls[0]).toHaveLength(1); + expect(startP2pRun).toHaveBeenCalledWith(expect.objectContaining({ + initiatorSession: 'deck_myapp_brain', + targets: [ { session: 'deck_myapp_w1', mode: 'audit' }, { session: 'deck_myapp_w2', mode: 'audit' }, ], - 'code review', - [], - mockServerLink, - 3, - ); + userText: 'code review', + fileContents: [], + serverLink: mockServerLink, + rounds: 3, + })); }); // 14. P2P with no valid participants — skips @@ -362,14 +363,15 @@ describe('executeCronJob', () => { await executeCronJob(msg, mockServerLink); - expect(startP2pRun).toHaveBeenCalledWith( - 'deck_myapp_brain', - [{ session: 'deck_myapp_w1', mode: 'review' }], - 'quick sync', - [], - mockServerLink, - 1, - ); + expect((startP2pRun as ReturnType).mock.calls[0]).toHaveLength(1); + expect(startP2pRun).toHaveBeenCalledWith(expect.objectContaining({ + initiatorSession: 'deck_myapp_brain', + targets: [{ session: 'deck_myapp_w1', mode: 'review' }], + userText: 'quick sync', + fileContents: [], + serverLink: mockServerLink, + rounds: 1, + })); }); it('logs warning for unknown action type', async () => { @@ -453,14 +455,15 @@ describe('executeCronJob', () => { await executeCronJob(msg, mockServerLink); - expect(startP2pRun).toHaveBeenCalledWith( - 'deck_myapp_brain', - [{ session: 'deck_sub_worker1', mode: 'review' }], - 'architecture review', - [], - mockServerLink, - 1, - ); + expect((startP2pRun as ReturnType).mock.calls[0]).toHaveLength(1); + expect(startP2pRun).toHaveBeenCalledWith(expect.objectContaining({ + initiatorSession: 'deck_myapp_brain', + targets: [{ session: 'deck_sub_worker1', mode: 'review' }], + userText: 'architecture review', + fileContents: [], + serverLink: mockServerLink, + rounds: 1, + })); }); it('sends command result with executionId when assistant output completes', async () => { diff --git a/test/daemon/p2p-orchestrator.test.ts b/test/daemon/p2p-orchestrator.test.ts index 3576400e..c33e8988 100644 --- a/test/daemon/p2p-orchestrator.test.ts +++ b/test/daemon/p2p-orchestrator.test.ts @@ -1,7 +1,7 @@ import { describe, it, expect, vi, beforeEach, afterEach } from 'vitest'; const isDarwin = process.platform === 'darwin'; -import { mkdir, readFile, rm, appendFile, writeFile, utimes, access } from 'node:fs/promises'; +import { mkdir, readFile, readdir, rm, appendFile, writeFile, utimes, access } from 'node:fs/promises'; import { join } from 'node:path'; import { tmpdir } from 'node:os'; @@ -12,6 +12,8 @@ const { getSessionMock, detectStatusMock, detectStatusAsyncMock, + launchTransportSessionMock, + stopTransportRuntimeSessionMock, serverLinkMock, } = vi.hoisted(() => ({ sendKeysDelayedEnterMock: vi.fn().mockResolvedValue(undefined), @@ -20,6 +22,8 @@ const { getSessionMock: vi.fn(), detectStatusMock: vi.fn().mockReturnValue('idle'), detectStatusAsyncMock: vi.fn().mockResolvedValue('idle'), + launchTransportSessionMock: vi.fn().mockResolvedValue(undefined), + stopTransportRuntimeSessionMock: vi.fn().mockResolvedValue(undefined), serverLinkMock: { send: vi.fn() }, })); @@ -41,6 +45,8 @@ vi.mock('../../src/agent/detect.js', () => ({ vi.mock('../../src/agent/session-manager.js', () => ({ getTransportRuntime: vi.fn(), + launchTransportSession: launchTransportSessionMock, + stopTransportRuntimeSession: stopTransportRuntimeSessionMock, })); vi.mock('../../src/util/logger.js', () => ({ @@ -63,6 +69,7 @@ import { _setGracePeriodMs, _setIdlePollMs, _setMinProcessingMs, + _setRoundHopCleanupDelayMs, type P2pRun, type P2pRunStatus, } from '../../src/daemon/p2p-orchestrator.js'; @@ -93,12 +100,27 @@ async function waitForStatus(runId: string, expected: P2pRunStatus[], maxMs = 10 throw new Error(`Run ${runId} ended in ${run.status}, expected ${expected.join(', ')}`); } +async function waitForNoRoundHopArtifacts(projectDir: string, runId: string, maxMs = 1000): Promise { + const discussionsDir = join(projectDir, '.imc', 'discussions'); + const start = Date.now(); + while (Date.now() - start < maxMs) { + const remainingArtifacts = (await readdir(discussionsDir).catch(() => [] as string[])) + .filter((name) => name.startsWith(runId) && /\.round\d+\.hop\d+\.md$/.test(name)); + if (remainingArtifacts.length === 0) return; + await new Promise((resolve) => setTimeout(resolve, 25)); + } + const remainingArtifacts = (await readdir(discussionsDir).catch(() => [] as string[])) + .filter((name) => name.startsWith(runId) && /\.round\d+\.hop\d+\.md$/.test(name)); + expect(remainingArtifacts).toEqual([]); +} + beforeEach(async () => { vi.clearAllMocks(); _setIdlePollMs(20); _setGracePeriodMs(80); _setMinProcessingMs(0); _setFileSettleCycles(1); + _setRoundHopCleanupDelayMs(30_000); tempProjectDir = join(tmpdir(), `p2p-par-${Date.now()}-${Math.random().toString(36).slice(2, 8)}`); await mkdir(tempProjectDir, { recursive: true }); @@ -133,6 +155,7 @@ afterEach(async () => { _setGracePeriodMs(180000); _setMinProcessingMs(30000); _setFileSettleCycles(3); + _setRoundHopCleanupDelayMs(30_000); await rm(tempProjectDir, { recursive: true, force: true }).catch(() => {}); }); @@ -152,6 +175,52 @@ describe('P2P orchestrator — parallel rounds', () => { expect(done.hopStates[0].artifact_path).toContain(`${done.id}.round1.hop1.md`); }); + it('keeps legacy single-mode runs on the legacy payload path without advanced fields', async () => { + const run = await startP2pRun( + 'deck_proj_brain', + [{ session: 'deck_proj_w1', mode: 'audit' }], + 'legacy single mode', + [], + serverLinkMock as any, + ); + + const done = await waitForStatus(run.id, ['completed']); + const payload = serializeP2pRun(done); + + expect(done.advancedP2pEnabled).toBe(false); + expect(payload.advanced_p2p_enabled).toBeUndefined(); + expect(payload.helper_diagnostics).toBeUndefined(); + expect(payload.all_nodes?.length).toBeGreaterThan(0); + expect(payload.mode_key).toBe('audit'); + expect(done.completedHops.map((hop) => hop.session)).toEqual(['deck_proj_w1']); + expect(done.skippedHops).toEqual([]); + expect(done.remainingTargets).toEqual([]); + }); + + it('preserves legacy combo-mode sequencing without advanced fields', async () => { + const run = await startP2pRun( + 'deck_proj_brain', + [{ session: 'deck_proj_w1', mode: 'brainstorm>discuss' as any }], + 'legacy combo mode', + [], + serverLinkMock as any, + 2, + undefined, + 'brainstorm>discuss', + ); + + const done = await waitForStatus(run.id, ['completed']); + const comboHops = done.hopStates.filter((hop) => hop.session === 'deck_proj_w1'); + + expect(done.advancedP2pEnabled).toBe(false); + expect(comboHops.map((hop) => hop.round_index)).toEqual([1, 2]); + expect(comboHops.map((hop) => hop.mode)).toEqual(['brainstorm', 'discuss']); + expect(done.completedHops.map((hop) => hop.session)).toEqual(['deck_proj_w1', 'deck_proj_w1']); + expect(done.skippedHops).toEqual([]); + expect(done.remainingTargets).toEqual([]); + expect(done.resultSummary).toContain('Final Summary'); + }); + it.skipIf(isDarwin)('cleans stale orphan hop artifacts when a new run starts', async () => { const discussionsDir = join(tempProjectDir, '.imc', 'discussions'); await mkdir(discussionsDir, { recursive: true }); @@ -534,6 +603,244 @@ describe('P2P orchestrator — parallel rounds', () => { expect(payload.hop_counts?.completed).toBeGreaterThanOrEqual(1); }); + it('preserves the active run phase when an advanced whole-run timeout fires', async () => { + let runId = ''; + sendKeysDelayedEnterMock.mockImplementationOnce(async (session: string, prompt: string) => { + const filePath = pathFromPrompt(prompt); + const heading = headingFromPrompt(prompt); + await appendFile(filePath, `\n## ${heading}\n\nSlow output from ${session}.\n`, 'utf8'); + setTimeout(() => { + const current = getP2pRun(runId); + if (current) current.deadlineAt = Date.now() - 1; + notifySessionIdle(session); + }, 20); + }); + + const run = await startP2pRun({ + initiatorSession: 'deck_proj_brain', + targets: [], + userText: 'timeout before final summary', + fileContents: [], + serverLink: serverLinkMock as any, + advancedRunTimeoutMs: 10, + advancedRounds: [ + { + id: 'implementation', + title: 'Implementation', + preset: 'implementation', + executionMode: 'single_main', + permissionScope: 'implementation', + }, + ], + }); + runId = run.id; + + const done = await waitForStatus(run.id, ['timed_out'], 15_000); + const payload = serializeP2pRun(done); + + expect(payload.status).toBe('timed_out'); + expect(payload.run_phase).toBe('round_execution'); + expect(payload.summary_phase).toBeNull(); + expect(payload.error).toContain('advanced_run_timeout'); + }, 20_000); + + it('serializes helper diagnostics only for advanced runs with the documented shape', () => { + const timestamp = Date.now(); + const run: P2pRun = { + id: 'run_helper_diag', + discussionId: 'disc_helper_diag', + mainSession: 'deck_proj_brain', + initiatorSession: 'deck_proj_brain', + currentTargetSession: null, + finalReturnSession: 'deck_proj_brain', + remainingTargets: [], + totalTargets: 0, + mode: 'plan', + status: 'running', + runPhase: 'round_execution', + summaryPhase: null, + activePhase: 'hop', + contextFilePath: '/tmp/run_helper_diag.md', + userText: 'helper diagnostics payload', + timeoutMs: 120000, + resultSummary: null, + completedHops: [], + skippedHops: [], + error: null, + createdAt: new Date().toISOString(), + updatedAt: new Date().toISOString(), + completedAt: null, + rounds: 1, + currentRound: 1, + allTargets: [], + extraPrompt: '', + hopStartedAt: Date.now(), + hopStates: [], + activeTargetSessions: [], + advancedP2pEnabled: true, + resolvedRounds: [], + helperEligibleSnapshot: [], + contextReducer: undefined, + advancedRunTimeoutMs: undefined, + deadlineAt: null, + currentRoundId: 'implementation', + currentExecutionStep: 2, + currentRoundAttempt: 3, + roundAttemptCounts: { implementation: 3 }, + roundJumpCounts: {}, + routingHistory: [], + helperDiagnostics: [ + { + code: 'P2P_HELPER_FALLBACK_FAILED', + attempt: 3, + sourceSession: 'deck_sub_helper', + templateSession: 'deck_proj_brain', + fallbackSession: 'deck_sub_helper', + timestamp, + message: 'fallback session timed out', + }, + ], + _cancelled: false, + }; + + const payload = serializeP2pRun(run); + + expect(payload.helper_diagnostics).toEqual([ + { + code: 'P2P_HELPER_FALLBACK_FAILED', + attempt: 3, + sourceSession: 'deck_sub_helper', + templateSession: 'deck_proj_brain', + fallbackSession: 'deck_sub_helper', + timestamp, + message: 'fallback session timed out', + }, + ]); + + const legacyPayload = serializeP2pRun({ + ...run, + id: 'run_helper_diag_legacy', + advancedP2pEnabled: false, + }); + expect(legacyPayload.helper_diagnostics).toBeUndefined(); + }); + + it('serializes the latest routing step for looped advanced rounds', () => { + const run: P2pRun = { + id: 'run_advanced_latest_step', + discussionId: 'disc_latest_step', + mainSession: 'deck_proj_brain', + initiatorSession: 'deck_proj_brain', + currentTargetSession: null, + finalReturnSession: 'deck_proj_brain', + remainingTargets: [], + totalTargets: 0, + mode: 'discuss', + status: 'running', + runPhase: 'round_execution', + summaryPhase: null, + activePhase: 'hop', + contextFilePath: '/tmp/run_advanced_latest_step.md', + userText: 'latest step serialization', + timeoutMs: 120000, + resultSummary: null, + completedHops: [], + skippedHops: [], + error: null, + createdAt: new Date().toISOString(), + updatedAt: new Date().toISOString(), + completedAt: null, + rounds: 2, + currentRound: 1, + allTargets: [], + extraPrompt: '', + hopStartedAt: Date.now(), + hopStates: [], + activeTargetSessions: [], + advancedP2pEnabled: true, + resolvedRounds: [ + { + id: 'implementation', + title: 'Implementation', + preset: 'implementation', + executionMode: 'multi_dispatch', + permissionScope: 'implementation', + promptSuffix: '', + timeoutMs: 60_000, + requiresVerdict: false, + allowRouting: false, + jumpRule: null, + output: { kind: 'discussion_append' }, + helperTask: null, + modeKey: 'discuss', + verdictPolicy: 'none', + }, + { + id: 'implementation_audit', + title: 'Implementation Audit', + preset: 'implementation_audit', + executionMode: 'single_main', + permissionScope: 'analysis_only', + promptSuffix: '', + timeoutMs: 60_000, + requiresVerdict: true, + allowRouting: true, + jumpRule: { + targetRoundId: 'implementation', + marker: 'REWORK', + minTriggers: 0, + maxTriggers: 2, + }, + output: { kind: 'discussion_append' }, + helperTask: null, + modeKey: 'audit', + verdictPolicy: 'smart_gate', + }, + ], + helperEligibleSnapshot: ['deck_proj_brain'], + contextReducer: undefined, + advancedRunTimeoutMs: undefined, + deadlineAt: null, + currentRoundId: 'implementation', + currentExecutionStep: 5, + currentRoundAttempt: 3, + roundAttemptCounts: { + implementation: 3, + implementation_audit: 2, + }, + roundJumpCounts: { + implementation_audit: 2, + }, + routingHistory: [ + { + fromRoundId: 'implementation_audit', + toRoundId: 'implementation', + trigger: 'REWORK', + atStep: 3, + atAttempt: 1, + timestamp: 1, + }, + { + fromRoundId: 'implementation_audit', + toRoundId: 'implementation', + trigger: 'REWORK', + atStep: 5, + atAttempt: 2, + timestamp: 2, + }, + ], + helperDiagnostics: [], + _cancelled: false, + }; + + const payload = serializeP2pRun(run); + + expect(payload.advanced_nodes?.find((node) => node.id === 'implementation')).toMatchObject({ + attempt: 3, + step: 5, + }); + }); + it('projects all active hops into all_nodes for parallel round progress', () => { const run: P2pRun = { id: 'run_parallel', @@ -617,4 +924,1149 @@ describe('P2P orchestrator — parallel rounds', () => { expect(payload.current_target_session).toBe('deck_proj_w1'); expect(payload.active_hop_number).toBe(1); }); + + it('falls back to an sdk child helper when the primary reducer session fails', async () => { + getSessionMock.mockImplementation((name: string) => { + if (name === 'deck_proj_brain') return { agentType: 'claude-code-sdk', projectDir: tempProjectDir, parentSession: undefined, label: 'brain' }; + if (name === 'deck_sub_helper') return { agentType: 'qwen', projectDir: tempProjectDir, parentSession: 'deck_proj_brain', label: 'helper' }; + return null; + }); + + sendKeysDelayedEnterMock.mockImplementation(async (session: string, prompt: string) => { + if (prompt.includes('[P2P Helper Task') && session === 'deck_proj_brain') { + throw new Error('primary reducer unavailable'); + } + const filePath = [...prompt.matchAll(/\/\S+?\.md/g)].at(-1)?.[0] ?? pathFromPrompt(prompt); + const heading = headingFromPrompt(prompt); + await appendFile(filePath, `\n## ${heading}\n\nOutput from ${session}.\n`, 'utf8'); + setTimeout(() => notifySessionIdle(session), 20); + }); + + const run = await startP2pRun({ + initiatorSession: 'deck_proj_brain', + targets: [{ session: 'deck_sub_helper', mode: 'audit' }], + userText: 'x'.repeat(40_000), + fileContents: [], + serverLink: serverLinkMock as any, + advancedRounds: [ + { + id: 'implementation', + title: 'Implementation', + preset: 'implementation', + executionMode: 'single_main', + permissionScope: 'implementation', + }, + ], + contextReducer: { + mode: 'reuse_existing_session', + sessionName: 'deck_proj_brain', + }, + }); + + const done = await waitForStatus(run.id, ['completed'], 15_000); + expect(done.status).toBe('completed'); + expect(done.helperEligibleSnapshot).toEqual([ + { sessionName: 'deck_proj_brain', agentType: 'claude-code-sdk', parentSession: null }, + { sessionName: 'deck_sub_helper', agentType: 'qwen', parentSession: 'deck_proj_brain' }, + ]); + expect(done.helperDiagnostics).toEqual(expect.arrayContaining([ + expect.objectContaining({ + code: 'P2P_HELPER_PRIMARY_FAILED', + sourceSession: 'deck_proj_brain', + }), + ])); + const reducerPath = join(tempProjectDir, '.imc', 'discussions', `${done.id}.reducer.1.md`); + await expect(access(reducerPath)).rejects.toThrow(); + expect(done.error).toBeNull(); + }, 15_000); + + it('cleans up clone-mode temporary sdk helper sessions after reducer use', async () => { + let helperName: string | null = null; + launchTransportSessionMock.mockImplementation(async (opts: any) => { + helperName = opts.name; + }); + getSessionMock.mockImplementation((name: string) => { + if (name === 'deck_proj_brain') { + return { + agentType: 'claude-code-sdk', + runtimeType: 'transport', + projectDir: tempProjectDir, + projectName: 'proj', + parentSession: undefined, + label: 'brain', + requestedModel: 'claude-sonnet', + transportConfig: { source: 'test' }, + }; + } + if (helperName && name === helperName) { + return { + agentType: 'claude-code-sdk', + runtimeType: 'transport', + projectDir: tempProjectDir, + projectName: 'proj', + parentSession: 'deck_proj_brain', + label: helperName, + requestedModel: 'claude-sonnet', + transportConfig: { source: 'test' }, + }; + } + return null; + }); + + sendKeysDelayedEnterMock.mockImplementation(async (session: string, prompt: string) => { + const filePath = [...prompt.matchAll(/\/\S+?\.md/g)].at(-1)?.[0] ?? pathFromPrompt(prompt); + const heading = headingFromPrompt(prompt); + await appendFile(filePath, `\n## ${heading}\n\nOutput from ${session}.\n`, 'utf8'); + setTimeout(() => notifySessionIdle(session), 20); + }); + + const run = await startP2pRun({ + initiatorSession: 'deck_proj_brain', + targets: [], + userText: `clone sdk helper cleanup ${'x'.repeat(40_000)}`, + fileContents: [], + serverLink: serverLinkMock as any, + advancedRounds: [ + { + id: 'implementation', + title: 'Implementation', + preset: 'implementation', + executionMode: 'single_main', + permissionScope: 'implementation', + }, + ], + contextReducer: { + mode: 'clone_sdk_session', + templateSession: 'deck_proj_brain', + }, + }); + + const done = await waitForStatus(run.id, ['completed'], 15_000); + expect(done.status).toBe('completed'); + expect(launchTransportSessionMock).toHaveBeenCalledTimes(1); + expect(stopTransportRuntimeSessionMock).toHaveBeenCalledWith(helperName); + }, 20_000); + + it('fails the run when both the primary reducer path and fallback child fail', async () => { + getSessionMock.mockImplementation((name: string) => { + if (name === 'deck_proj_brain') return { agentType: 'claude-code-sdk', projectDir: tempProjectDir, parentSession: undefined, label: 'brain' }; + if (name === 'deck_sub_helper') return { agentType: 'qwen', projectDir: tempProjectDir, parentSession: 'deck_proj_brain', label: 'helper' }; + return null; + }); + + sendKeysDelayedEnterMock.mockImplementation(async (session: string, prompt: string) => { + if (prompt.includes('[P2P Helper Task')) { + throw new Error(`helper failed for ${session}`); + } + const filePath = pathFromPrompt(prompt); + const heading = headingFromPrompt(prompt); + await appendFile(filePath, `\n## ${heading}\n\nOutput from ${session}.\n`, 'utf8'); + setTimeout(() => notifySessionIdle(session), 20); + }); + + const run = await startP2pRun({ + initiatorSession: 'deck_proj_brain', + targets: [{ session: 'deck_sub_helper', mode: 'audit' }], + userText: 'x'.repeat(40_000), + fileContents: [], + serverLink: serverLinkMock as any, + advancedRounds: [ + { + id: 'implementation', + title: 'Implementation', + preset: 'implementation', + executionMode: 'single_main', + permissionScope: 'implementation', + }, + ], + contextReducer: { + mode: 'reuse_existing_session', + sessionName: 'deck_proj_brain', + }, + }); + + const done = await waitForStatus(run.id, ['failed'], 15_000); + expect(done.error).toContain('helper_fallback_failed:deck_sub_helper'); + expect(done.helperDiagnostics).toEqual(expect.arrayContaining([ + expect.objectContaining({ code: 'P2P_HELPER_PRIMARY_FAILED', sourceSession: 'deck_proj_brain' }), + expect.objectContaining({ code: 'P2P_HELPER_FALLBACK_FAILED', sourceSession: 'deck_sub_helper' }), + ])); + }, 20_000); + + it('filters cli participants out of the advanced helper snapshot stored on the run', async () => { + getSessionMock.mockImplementation((name: string) => { + if (name === 'deck_proj_brain') return { agentType: 'claude-code-sdk', projectDir: tempProjectDir, parentSession: undefined, label: 'brain' }; + if (name === 'deck_sub_helper') return { agentType: 'qwen', projectDir: tempProjectDir, parentSession: 'deck_proj_brain', label: 'helper' }; + if (name === 'deck_proj_cli') return { agentType: 'codex', projectDir: tempProjectDir, parentSession: undefined, label: 'cli' }; + return null; + }); + + const run = await startP2pRun({ + initiatorSession: 'deck_proj_brain', + targets: [ + { session: 'deck_sub_helper', mode: 'audit' }, + { session: 'deck_proj_cli', mode: 'review' }, + ], + userText: 'helper snapshot should stay sdk-only', + fileContents: [], + serverLink: serverLinkMock as any, + advancedRounds: [ + { + id: 'implementation', + title: 'Implementation', + preset: 'implementation', + executionMode: 'single_main', + permissionScope: 'implementation', + }, + ], + contextReducer: { + mode: 'clone_sdk_session', + templateSession: 'deck_proj_brain', + }, + }); + + expect(run.helperEligibleSnapshot).toEqual([ + { sessionName: 'deck_proj_brain', agentType: 'claude-code-sdk', parentSession: null }, + { sessionName: 'deck_sub_helper', agentType: 'qwen', parentSession: 'deck_proj_brain' }, + ]); + + await cancelP2pRun(run.id, serverLinkMock as any); + }); + + it('fails before creating a run when advanced config is rejected by the resolver', async () => { + const beforeIds = listP2pRuns().map((run) => run.id); + + await expect(startP2pRun({ + initiatorSession: 'deck_proj_brain', + targets: [], + userText: 'invalid advanced config should not start', + fileContents: [], + serverLink: serverLinkMock as any, + advancedRounds: [ + { + id: 'implementation', + title: 'Implementation', + preset: 'implementation', + executionMode: 'single_main', + permissionScope: 'implementation', + }, + { + id: 'implementation_audit', + title: 'Implementation Audit', + preset: 'implementation_audit', + executionMode: 'single_main', + permissionScope: 'analysis_only', + verdictPolicy: 'smart_gate', + jumpRule: { + targetRoundId: 'implementation_audit', + marker: 'REWORK', + minTriggers: 0, + maxTriggers: 2, + }, + }, + ], + })).rejects.toThrow(/jump backward/i); + + expect(listP2pRuns().map((run) => run.id)).toEqual(beforeIds); + }); + + it('launches and tears down clone-mode helper sessions', async () => { + getSessionMock.mockImplementation((name: string) => { + if (name === 'deck_proj_brain') return { + agentType: 'claude-code-sdk', + runtimeType: 'transport', + projectDir: tempProjectDir, + parentSession: undefined, + label: 'brain', + projectName: 'proj', + requestedModel: 'claude-sonnet', + activeModel: 'claude-sonnet', + transportConfig: { baseUrl: 'http://localhost:1234' }, + effort: 'high', + }; + if (name === 'deck_sub_helper') return { agentType: 'qwen', projectDir: tempProjectDir, parentSession: 'deck_proj_brain', label: 'helper' }; + return null; + }); + + let helperSessionName = ''; + launchTransportSessionMock.mockImplementation(async (opts: any) => { + helperSessionName = opts.name; + return undefined; + }); + + sendKeysDelayedEnterMock.mockImplementation(async (session: string, prompt: string) => { + const filePath = [...prompt.matchAll(/\/\S+?\.md/g)].at(-1)?.[0] ?? pathFromPrompt(prompt); + const heading = headingFromPrompt(prompt); + await appendFile(filePath, `\n## ${heading}\n\nOutput from ${session}.\n`, 'utf8'); + setTimeout(() => notifySessionIdle(session), 20); + }); + + const run = await startP2pRun({ + initiatorSession: 'deck_proj_brain', + targets: [{ session: 'deck_sub_helper', mode: 'audit' }], + userText: 'x'.repeat(40_000), + fileContents: [], + serverLink: serverLinkMock as any, + advancedRounds: [ + { + id: 'implementation', + title: 'Implementation', + preset: 'implementation', + executionMode: 'single_main', + permissionScope: 'implementation', + }, + ], + contextReducer: { + mode: 'clone_sdk_session', + templateSession: 'deck_proj_brain', + }, + }); + + const done = await waitForStatus(run.id, ['completed'], 15_000); + expect(done.status).toBe('completed'); + expect(launchTransportSessionMock).toHaveBeenCalledWith(expect.objectContaining({ + agentType: 'claude-code-sdk', + projectDir: tempProjectDir, + requestedModel: 'claude-sonnet', + transportConfig: { baseUrl: 'http://localhost:1234' }, + skipStore: true, + fresh: true, + })); + expect(stopTransportRuntimeSessionMock).toHaveBeenCalledWith(helperSessionName); + }, 15_000); + + it('treats missing advanced audit verdicts as rework and records jump history', async () => { + getSessionMock.mockImplementation((name: string) => { + if (name === 'deck_proj_brain') return { agentType: 'claude-code-sdk', projectDir: tempProjectDir, parentSession: undefined, label: 'brain' }; + return null; + }); + + sendKeysDelayedEnterMock.mockImplementation(async (session: string, prompt: string) => { + const filePath = pathFromPrompt(prompt); + const heading = headingFromPrompt(prompt); + const body = heading.includes('Implementation Audit') + ? 'Audit completed without verdict marker.' + : `Output from ${session}.`; + await appendFile(filePath, `\n## ${heading}\n\n${body}\n`, 'utf8'); + setTimeout(() => notifySessionIdle(session), 20); + }); + + const run = await startP2pRun({ + initiatorSession: 'deck_proj_brain', + targets: [], + userText: 'loop until the audit stabilizes', + fileContents: [], + serverLink: serverLinkMock as any, + advancedRounds: [ + { + id: 'implementation', + title: 'Implementation', + preset: 'implementation', + executionMode: 'single_main', + permissionScope: 'implementation', + }, + { + id: 'implementation_audit', + title: 'Implementation Audit', + preset: 'implementation_audit', + executionMode: 'single_main', + permissionScope: 'analysis_only', + verdictPolicy: 'smart_gate', + jumpRule: { + targetRoundId: 'implementation', + marker: 'REWORK', + minTriggers: 0, + maxTriggers: 2, + }, + }, + ], + }); + + const done = await waitForStatus(run.id, ['completed'], 15_000); + expect(done.status).toBe('completed'); + expect(done.roundJumpCounts.implementation_audit).toBe(2); + expect(done.routingHistory).toHaveLength(2); + expect(done.helperDiagnostics.filter((entry) => entry.code === 'P2P_VERDICT_MISSING').length).toBeGreaterThanOrEqual(1); + expect(done.resultSummary).toContain('Final Summary'); + }); + + it('forces the minimum rework loops before handing off to smart-gate evaluation', async () => { + getSessionMock.mockImplementation((name: string) => { + if (name === 'deck_proj_brain') return { agentType: 'claude-code-sdk', projectDir: tempProjectDir, parentSession: undefined, label: 'brain' }; + return null; + }); + + let auditCount = 0; + sendKeysDelayedEnterMock.mockImplementation(async (session: string, prompt: string) => { + const filePath = pathFromPrompt(prompt); + const heading = headingFromPrompt(prompt); + let body = `Output from ${session}.`; + if (heading.includes('Implementation Audit')) { + auditCount += 1; + body = `Audit pass ${auditCount}.\n`; + } + await appendFile(filePath, `\n## ${heading}\n\n${body}\n`, 'utf8'); + setTimeout(() => notifySessionIdle(session), 20); + }); + + const run = await startP2pRun({ + initiatorSession: 'deck_proj_brain', + targets: [], + userText: 'forced rework should ignore PASS until minTriggers is satisfied', + fileContents: [], + serverLink: serverLinkMock as any, + advancedRounds: [ + { + id: 'implementation', + title: 'Implementation', + preset: 'implementation', + executionMode: 'single_main', + permissionScope: 'implementation', + }, + { + id: 'implementation_audit', + title: 'Implementation Audit', + preset: 'implementation_audit', + executionMode: 'single_main', + permissionScope: 'analysis_only', + verdictPolicy: 'forced_rework', + jumpRule: { + targetRoundId: 'implementation', + marker: 'REWORK', + minTriggers: 2, + maxTriggers: 3, + }, + }, + ], + }); + + const done = await waitForStatus(run.id, ['completed'], 15_000); + expect(done.status).toBe('completed'); + expect(auditCount).toBe(3); + expect(done.roundJumpCounts.implementation_audit).toBe(2); + expect(done.routingHistory).toHaveLength(2); + expect(done.routingHistory.every((entry) => entry.trigger === 'PASS')).toBe(true); + }); + + it('hands off forced_rework rounds to smart-gate behavior after minTriggers is satisfied', async () => { + getSessionMock.mockImplementation((name: string) => { + if (name === 'deck_proj_brain') return { agentType: 'claude-code-sdk', projectDir: tempProjectDir, parentSession: undefined, label: 'brain' }; + return null; + }); + + sendKeysDelayedEnterMock.mockImplementation(async (session: string, prompt: string) => { + const filePath = pathFromPrompt(prompt); + const heading = headingFromPrompt(prompt); + const body = heading.includes('Implementation Audit') + ? '\nAudit says pass.' + : `Implementation output ${session}.`; + await appendFile(filePath, `\n## ${heading}\n\n${body}\n`, 'utf8'); + setTimeout(() => notifySessionIdle(session), 20); + }); + + const run = await startP2pRun({ + initiatorSession: 'deck_proj_brain', + targets: [], + userText: 'forced rework should stop looping once the minimum rework count is satisfied and audit passes', + fileContents: [], + serverLink: serverLinkMock as any, + advancedRounds: [ + { + id: 'implementation', + title: 'Implementation', + preset: 'implementation', + executionMode: 'single_main', + permissionScope: 'implementation', + }, + { + id: 'implementation_audit', + title: 'Implementation Audit', + preset: 'implementation_audit', + executionMode: 'single_main', + permissionScope: 'analysis_only', + verdictPolicy: 'forced_rework', + jumpRule: { + targetRoundId: 'implementation', + marker: 'REWORK', + minTriggers: 2, + maxTriggers: 4, + }, + }, + ], + }); + + const done = await waitForStatus(run.id, ['completed'], 15_000); + expect(done.status).toBe('completed'); + expect(done.roundAttemptCounts.implementation).toBe(3); + expect(done.roundAttemptCounts.implementation_audit).toBe(3); + expect(done.roundJumpCounts.implementation_audit).toBe(2); + expect(done.routingHistory).toHaveLength(2); + }, 20_000); + + it('continues forced_rework routing on REWORK after minTriggers until maxTriggers is exhausted', async () => { + getSessionMock.mockImplementation((name: string) => { + if (name === 'deck_proj_brain') return { agentType: 'claude-code-sdk', projectDir: tempProjectDir, parentSession: undefined, label: 'brain' }; + return null; + }); + + let auditCount = 0; + sendKeysDelayedEnterMock.mockImplementation(async (session: string, prompt: string) => { + const filePath = pathFromPrompt(prompt); + const heading = headingFromPrompt(prompt); + let body = `Implementation output ${session}.`; + if (heading.includes('Implementation Audit')) { + auditCount += 1; + body = `Audit says rework ${auditCount}.\n`; + } + await appendFile(filePath, `\n## ${heading}\n\n${body}\n`, 'utf8'); + setTimeout(() => notifySessionIdle(session), 20); + }); + + const run = await startP2pRun({ + initiatorSession: 'deck_proj_brain', + targets: [], + userText: 'keep reworking until the forced loop budget is exhausted', + fileContents: [], + serverLink: serverLinkMock as any, + advancedRounds: [ + { + id: 'implementation', + title: 'Implementation', + preset: 'implementation', + executionMode: 'single_main', + permissionScope: 'implementation', + }, + { + id: 'implementation_audit', + title: 'Implementation Audit', + preset: 'implementation_audit', + executionMode: 'single_main', + permissionScope: 'analysis_only', + verdictPolicy: 'forced_rework', + jumpRule: { + targetRoundId: 'implementation', + marker: 'REWORK', + minTriggers: 1, + maxTriggers: 2, + }, + }, + ], + }); + + const done = await waitForStatus(run.id, ['completed'], 15_000); + expect(done.status).toBe('completed'); + expect(auditCount).toBe(3); + expect(done.roundAttemptCounts.implementation).toBe(3); + expect(done.roundAttemptCounts.implementation_audit).toBe(3); + expect(done.roundJumpCounts.implementation_audit).toBe(2); + expect(done.routingHistory).toHaveLength(2); + expect(done.routingHistory.every((entry) => entry.trigger === 'REWORK')).toBe(true); + }, 20_000); + + it('uses initiator synthesis as the authoritative verdict source for multi-dispatch audit rounds', async () => { + getSessionMock.mockImplementation((name: string) => { + if (name === 'deck_proj_brain') return { agentType: 'claude-code-sdk', projectDir: tempProjectDir, parentSession: undefined, label: 'brain' }; + if (name === 'deck_proj_w1') return { agentType: 'claude-code-sdk', projectDir: tempProjectDir, parentSession: undefined, label: 'w1' }; + return null; + }); + + sendKeysDelayedEnterMock.mockImplementation(async (session: string, prompt: string) => { + const filePath = pathFromPrompt(prompt); + const heading = headingFromPrompt(prompt); + let body = `Output from ${session}.`; + if (heading.includes('Implementation Audit Synthesis')) { + body = 'Synthesis accepts the implementation.\n'; + } else if (heading.includes('Implementation Audit (hop')) { + body = 'Worker thinks this still needs work.\n'; + } + await appendFile(filePath, `\n## ${heading}\n\n${body}\n`, 'utf8'); + setTimeout(() => notifySessionIdle(session), 20); + }); + + const run = await startP2pRun({ + initiatorSession: 'deck_proj_brain', + targets: [{ session: 'deck_proj_w1', mode: 'audit' }], + userText: 'worker verdicts must not override synthesis', + fileContents: [], + serverLink: serverLinkMock as any, + advancedRounds: [ + { + id: 'implementation', + title: 'Implementation', + preset: 'implementation', + executionMode: 'single_main', + permissionScope: 'implementation', + }, + { + id: 'implementation_audit', + title: 'Implementation Audit', + preset: 'implementation_audit', + executionMode: 'multi_dispatch', + permissionScope: 'analysis_only', + verdictPolicy: 'smart_gate', + jumpRule: { + targetRoundId: 'implementation', + marker: 'REWORK', + minTriggers: 0, + maxTriggers: 2, + }, + }, + ], + }); + + const done = await waitForStatus(run.id, ['completed'], 15_000); + expect(done.status).toBe('completed'); + expect(done.routingHistory).toEqual([]); + expect(done.roundJumpCounts.implementation_audit).toBeUndefined(); + }, 20_000); + + it('rejects invalid advanced configs before creating or running a daemon run', async () => { + await expect(startP2pRun({ + initiatorSession: 'deck_proj_brain', + targets: [], + userText: 'invalid advanced config', + fileContents: [], + serverLink: serverLinkMock as any, + advancedRounds: [ + { + id: 'implementation', + title: 'Implementation', + preset: 'implementation', + executionMode: 'single_main', + permissionScope: 'implementation', + }, + { + id: 'implementation_audit', + title: 'Implementation Audit', + preset: 'implementation_audit', + executionMode: 'single_main', + permissionScope: 'analysis_only', + verdictPolicy: 'smart_gate', + jumpRule: { + targetRoundId: 'missing_round', + marker: 'REWORK', + minTriggers: 0, + maxTriggers: 1, + }, + }, + ], + })).rejects.toThrow(/unknown target/i); + + expect(listP2pRuns()).toHaveLength(0); + expect(serverLinkMock.send).not.toHaveBeenCalled(); + }); + + it('times out if the whole-run deadline expires during final summary dispatch', async () => { + let activeRunId = ''; + sendKeysDelayedEnterMock.mockImplementation(async (session: string, prompt: string) => { + const filePath = pathFromPrompt(prompt); + const heading = headingFromPrompt(prompt); + await appendFile(filePath, `\n## ${heading}\n\nOutput from ${session}.\n`, 'utf8'); + if (heading.includes('Final Summary')) { + setTimeout(() => { + const current = getP2pRun(activeRunId); + if (current) current.deadlineAt = Date.now() - 1; + }, 0); + setTimeout(() => { + const current = getP2pRun(activeRunId); + if (current) current.deadlineAt = Date.now() - 1; + notifySessionIdle(session); + }, 50); + return; + } + setTimeout(() => notifySessionIdle(session), 20); + }); + + const run = await startP2pRun({ + initiatorSession: 'deck_proj_brain', + targets: [], + userText: 'summary timeout boundary', + fileContents: [], + serverLink: serverLinkMock as any, + advancedRunTimeoutMs: 60_000, + advancedRounds: [ + { + id: 'implementation', + title: 'Implementation', + preset: 'implementation', + executionMode: 'single_main', + permissionScope: 'implementation', + }, + ], + }); + activeRunId = run.id; + + const done = await waitForStatus(run.id, ['timed_out'], 15_000); + const payload = serializeP2pRun(done); + + expect(payload.status).toBe('timed_out'); + expect(payload.run_phase).toBe('summarizing'); + expect(payload.summary_phase).toBe('failed'); + expect(payload.error).toContain('advanced_run_timeout'); + }, 20_000); + + it('fails artifact_generation rounds when declared outputs are left stale', async () => { + const stalePath = join(tempProjectDir, 'docs', 'plan.md'); + await mkdir(join(tempProjectDir, 'docs'), { recursive: true }); + await writeFile(stalePath, 'stale artifact\n', 'utf8'); + + sendKeysDelayedEnterMock.mockImplementation(async (session: string, prompt: string) => { + const filePath = pathFromPrompt(prompt); + const heading = headingFromPrompt(prompt); + await appendFile(filePath, `\n## ${heading}\n\nOutput from ${session}.\n`, 'utf8'); + setTimeout(() => notifySessionIdle(session), 20); + }); + + const run = await startP2pRun({ + initiatorSession: 'deck_proj_brain', + targets: [], + userText: 'artifact round must update the file', + fileContents: [], + serverLink: serverLinkMock as any, + advancedRounds: [ + { + id: 'custom_artifact', + title: 'Custom Artifact', + preset: 'custom', + executionMode: 'single_main', + permissionScope: 'artifact_generation', + artifactOutputs: ['docs/plan.md'], + }, + ], + }); + + const done = await waitForStatus(run.id, ['failed'], 15_000); + expect(done.error).toContain('Expected artifact not observably updated'); + expect(done.error).toContain('docs/plan.md'); + }, 20_000); + + it('completes the openspec preset after proposal artifacts are created and audit eventually passes', async () => { + getSessionMock.mockImplementation((name: string) => { + if (name === 'deck_proj_brain') return { agentType: 'claude-code-sdk', projectDir: tempProjectDir, parentSession: undefined, label: 'brain' }; + if (name === 'deck_proj_w1') return { agentType: 'claude-code-sdk', projectDir: tempProjectDir, parentSession: undefined, label: 'w1' }; + return null; + }); + + let auditCount = 0; + sendKeysDelayedEnterMock.mockImplementation(async (session: string, prompt: string) => { + const filePath = pathFromPrompt(prompt); + const heading = headingFromPrompt(prompt); + let body = `Output from ${session}.`; + if (heading.includes('OpenSpec Propose')) { + const changeDir = join(tempProjectDir, 'openspec', 'changes', 'smart-p2p-upgrade-e2e'); + await mkdir(join(changeDir, 'specs', 'smart-p2p-rounds'), { recursive: true }); + await writeFile(join(changeDir, 'proposal.md'), '# proposal\n', 'utf8'); + await writeFile(join(changeDir, 'design.md'), '# design\n', 'utf8'); + await writeFile(join(changeDir, 'tasks.md'), '# tasks\n', 'utf8'); + await writeFile(join(changeDir, 'specs', 'smart-p2p-rounds', 'spec.md'), '# spec\n', 'utf8'); + body = 'OpenSpec artifacts written.'; + } else if (heading.includes('Implementation Audit Synthesis') || heading.includes('Implementation Audit')) { + auditCount += 1; + body = auditCount === 1 + ? 'Audit requests another pass.\n' + : 'Audit accepts the implementation.\n'; + } + await appendFile(filePath, `\n## ${heading}\n\n${body}\n`, 'utf8'); + setTimeout(() => notifySessionIdle(session), 20); + }); + + const run = await startP2pRun({ + initiatorSession: 'deck_proj_brain', + targets: [{ session: 'deck_proj_w1', mode: 'audit' }], + userText: 'run the openspec preset end to end', + fileContents: [], + serverLink: serverLinkMock as any, + advancedPresetKey: 'openspec', + }); + + const done = await waitForStatus(run.id, ['completed'], 20_000); + expect(done.status).toBe('completed'); + expect(auditCount).toBe(2); + expect(done.roundJumpCounts.implementation_audit).toBe(1); + await expect(access(join(tempProjectDir, 'openspec', 'changes', 'smart-p2p-upgrade-e2e', 'proposal.md'))).resolves.toBeUndefined(); + }, 25_000); + + it('cleans up loop-generated hop artifacts after repeated advanced attempts settle', async () => { + _setRoundHopCleanupDelayMs(20); + + getSessionMock.mockImplementation((name: string) => { + if (name === 'deck_proj_brain') return { agentType: 'claude-code-sdk', projectDir: tempProjectDir, parentSession: undefined, label: 'brain' }; + if (name === 'deck_proj_w1') return { agentType: 'claude-code-sdk', projectDir: tempProjectDir, parentSession: undefined, label: 'w1' }; + return null; + }); + + let auditCount = 0; + sendKeysDelayedEnterMock.mockImplementation(async (session: string, prompt: string) => { + const filePath = pathFromPrompt(prompt); + const heading = headingFromPrompt(prompt); + let body = `Output from ${session}.`; + if (heading.includes('Implementation Audit')) { + auditCount += 1; + body = auditCount < 3 + ? `Audit iteration ${auditCount}.\n` + : 'Audit accepts the latest attempt.\n'; + } + await appendFile(filePath, `\n## ${heading}\n\n${body}\n`, 'utf8'); + setTimeout(() => notifySessionIdle(session), 20); + }); + + const run = await startP2pRun({ + initiatorSession: 'deck_proj_brain', + targets: [{ session: 'deck_proj_w1', mode: 'audit' }], + userText: 'loop cleanup regression', + fileContents: [], + serverLink: serverLinkMock as any, + advancedRounds: [ + { + id: 'implementation', + title: 'Implementation', + preset: 'implementation', + executionMode: 'multi_dispatch', + permissionScope: 'implementation', + }, + { + id: 'implementation_audit', + title: 'Implementation Audit', + preset: 'implementation_audit', + executionMode: 'single_main', + permissionScope: 'analysis_only', + verdictPolicy: 'smart_gate', + jumpRule: { + targetRoundId: 'implementation', + marker: 'REWORK', + minTriggers: 0, + maxTriggers: 3, + }, + }, + ], + }); + + const done = await waitForStatus(run.id, ['completed'], 20_000); + expect(done.roundJumpCounts.implementation_audit).toBe(2); + + await new Promise((resolve) => setTimeout(resolve, 120)); + + const discussionsDir = join(tempProjectDir, '.imc', 'discussions'); + const files = await readdir(discussionsDir); + expect(files).toContain(`${done.id}.md`); + expect(files.filter((file) => file.includes('.round'))).toEqual([]); + }, 25_000); + + it('applies per-round timeout budgets for advanced rounds', async () => { + sendKeysDelayedEnterMock.mockImplementation(async (session: string, prompt: string) => { + const filePath = pathFromPrompt(prompt); + const heading = headingFromPrompt(prompt); + await appendFile(filePath, `\n## ${heading}\n\nSlow output from ${session}.\n`, 'utf8'); + setTimeout(() => notifySessionIdle(session), 50); + }); + + const run = await startP2pRun({ + initiatorSession: 'deck_proj_brain', + targets: [], + userText: 'per-round timeout should override the default hop timeout', + fileContents: [], + serverLink: serverLinkMock as any, + advancedRunTimeoutMs: 60_000, + advancedRounds: [ + { + id: 'implementation', + title: 'Implementation', + preset: 'implementation', + executionMode: 'single_main', + permissionScope: 'implementation', + timeoutMinutes: 0, + }, + ], + }); + + const done = await waitForStatus(run.id, ['timed_out'], 15_000); + expect(done.status).toBe('timed_out'); + expect(done.runPhase).toBe('round_execution'); + expect(done.error).toContain('timed_out'); + }, 20_000); + + it('initializes advanced runs with deterministic bookkeeping fields', async () => { + getSessionMock.mockImplementation((name: string) => { + if (name === 'deck_proj_brain') return { agentType: 'claude-code-sdk', projectDir: tempProjectDir, parentSession: undefined, label: 'brain' }; + if (name === 'deck_sub_helper') return { agentType: 'qwen', projectDir: tempProjectDir, parentSession: 'deck_proj_brain', label: 'helper' }; + return null; + }); + + const run = await startP2pRun({ + initiatorSession: 'deck_proj_brain', + targets: [{ session: 'deck_sub_helper', mode: 'audit' }], + userText: 'bookkeeping init', + fileContents: [], + serverLink: serverLinkMock as any, + advancedPresetKey: 'openspec', + contextReducer: { + mode: 'clone_sdk_session', + templateSession: 'deck_proj_brain', + }, + advancedRunTimeoutMs: 120_000, + }); + + expect(run.advancedP2pEnabled).toBe(true); + expect(run.currentRoundId).toBe('discussion'); + expect(run.currentExecutionStep).toBe(1); + expect(run.currentRoundAttempt).toBe(1); + expect(run.roundAttemptCounts).toEqual({ discussion: 1 }); + expect(run.roundJumpCounts).toEqual({}); + expect(run.routingHistory).toEqual([]); + expect(run.helperEligibleSnapshot).toEqual([ + { sessionName: 'deck_proj_brain', agentType: 'claude-code-sdk', parentSession: null }, + { sessionName: 'deck_sub_helper', agentType: 'qwen', parentSession: 'deck_proj_brain' }, + ]); + expect(run.deadlineAt).toBeTypeOf('number'); + expect(run.deadlineAt).toBeGreaterThan(Date.now()); + + await cancelP2pRun(run.id, serverLinkMock as any); + }); + + it('keeps advanced loop bookkeeping deterministic while legacy projections remain compatibility-only', async () => { + getSessionMock.mockImplementation((name: string) => { + if (name === 'deck_proj_brain') return { agentType: 'claude-code-sdk', projectDir: tempProjectDir, parentSession: undefined, label: 'brain' }; + if (name === 'deck_proj_w1') return { agentType: 'claude-code-sdk', projectDir: tempProjectDir, parentSession: undefined, label: 'w1' }; + return null; + }); + + let auditCount = 0; + let finalSummaryCount = 0; + sendKeysDelayedEnterMock.mockImplementation(async (session: string, prompt: string) => { + const filePath = pathFromPrompt(prompt); + const heading = headingFromPrompt(prompt); + let body = `Output from ${session}.`; + if (heading.includes('Implementation Audit')) { + auditCount += 1; + body = auditCount === 1 + ? 'Needs one more pass.\n' + : 'Looks good now.\n'; + } + if (heading.includes('Final Summary')) finalSummaryCount += 1; + await appendFile(filePath, `\n## ${heading}\n\n${body}\n`, 'utf8'); + setTimeout(() => notifySessionIdle(session), 20); + }); + + const run = await startP2pRun({ + initiatorSession: 'deck_proj_brain', + targets: [{ session: 'deck_proj_w1', mode: 'audit' }], + userText: 'compatibility bookkeeping under loop-back', + fileContents: [], + serverLink: serverLinkMock as any, + advancedRounds: [ + { + id: 'implementation', + title: 'Implementation', + preset: 'implementation', + executionMode: 'multi_dispatch', + permissionScope: 'implementation', + }, + { + id: 'implementation_audit', + title: 'Implementation Audit', + preset: 'implementation_audit', + executionMode: 'single_main', + permissionScope: 'analysis_only', + verdictPolicy: 'smart_gate', + jumpRule: { + targetRoundId: 'implementation', + marker: 'REWORK', + minTriggers: 0, + maxTriggers: 2, + }, + }, + ], + }); + + const done = await waitForStatus(run.id, ['completed'], 20_000); + const payload = serializeP2pRun(done); + + expect(finalSummaryCount).toBe(1); + expect(done.roundAttemptCounts).toMatchObject({ + implementation: 2, + implementation_audit: 2, + }); + expect(done.currentExecutionStep).toBe(4); + expect(done.routingHistory).toHaveLength(1); + expect(done.remainingTargets).toEqual([]); + expect(payload.remaining_count).toBe(0); + expect(payload.completed_hops_count).toBe(1); + expect(payload.hop_counts?.completed).toBe(1); + expect(payload.advanced_p2p_enabled).toBe(true); + expect(payload.advanced_nodes?.find((node) => node.id === 'implementation')?.attempt).toBe(2); + expect(payload.all_nodes?.length).toBeGreaterThan(0); + }, 25_000); + + it('injects reducer summaries into later loop prompts and keeps the helper prompt focused on the latest attempt context', async () => { + getSessionMock.mockImplementation((name: string) => { + if (name === 'deck_proj_brain') return { agentType: 'claude-code-sdk', projectDir: tempProjectDir, parentSession: undefined, label: 'brain' }; + return null; + }); + + const helperPrompts: string[] = []; + const implementationPrompts: string[] = []; + let auditCount = 0; + sendKeysDelayedEnterMock.mockImplementation(async (session: string, prompt: string) => { + const helperPath = [...prompt.matchAll(/\/\S+?\.md/g)].at(-1)?.[0]; + const heading = headingFromPrompt(prompt); + if (prompt.includes('[P2P Helper Task')) { + helperPrompts.push(prompt); + if (!helperPath) throw new Error('missing helper summary path'); + await appendFile(helperPath, `\n## ${heading}\n\nREDUCED-CONTEXT-LATEST\n`, 'utf8'); + setTimeout(() => notifySessionIdle(session), 20); + return; + } + if (heading.includes('Implementation (attempt')) implementationPrompts.push(prompt); + const filePath = pathFromPrompt(prompt); + let body = `Output from ${session}.`; + if (heading.includes('Implementation Audit')) { + auditCount += 1; + body = auditCount === 1 + ? 'Try once more.\n' + : 'Pass now.\n'; + } + await appendFile(filePath, `\n## ${heading}\n\n${body}\n`, 'utf8'); + setTimeout(() => notifySessionIdle(session), 20); + }); + + const run = await startP2pRun({ + initiatorSession: 'deck_proj_brain', + targets: [], + userText: `latest-attempt reducer focus ${'x'.repeat(40_000)}`, + fileContents: [], + serverLink: serverLinkMock as any, + advancedRounds: [ + { + id: 'implementation', + title: 'Implementation', + preset: 'implementation', + executionMode: 'single_main', + permissionScope: 'implementation', + }, + { + id: 'implementation_audit', + title: 'Implementation Audit', + preset: 'implementation_audit', + executionMode: 'single_main', + permissionScope: 'analysis_only', + verdictPolicy: 'smart_gate', + jumpRule: { + targetRoundId: 'implementation', + marker: 'REWORK', + minTriggers: 0, + maxTriggers: 2, + }, + }, + ], + contextReducer: { + mode: 'reuse_existing_session', + sessionName: 'deck_proj_brain', + }, + }); + + const done = await waitForStatus(run.id, ['completed'], 20_000); + expect(done.status).toBe('completed'); + expect(helperPrompts.length).toBeGreaterThan(0); + expect(helperPrompts[0]).toContain('Focus on: latest implementation attempt, latest audit findings, declared artifact targets'); + const laterImplementationPrompt = implementationPrompts.find((entry) => entry.includes('attempt 2')); + expect(laterImplementationPrompt).toContain('Reduced context for this attempt:'); + expect(laterImplementationPrompt).toContain('REDUCED-CONTEXT-LATEST'); + }, 25_000); + + it('cleans worker-hop artifacts after repeated loop attempts', async () => { + getSessionMock.mockImplementation((name: string) => { + if (name === 'deck_proj_brain') return { agentType: 'claude-code-sdk', projectDir: tempProjectDir, parentSession: undefined, label: 'brain' }; + if (name === 'deck_proj_w1') return { agentType: 'claude-code-sdk', projectDir: tempProjectDir, parentSession: undefined, label: 'w1' }; + return null; + }); + + let auditCount = 0; + sendKeysDelayedEnterMock.mockImplementation(async (session: string, prompt: string) => { + const filePath = pathFromPrompt(prompt); + const heading = headingFromPrompt(prompt); + let body = `Output from ${session}.`; + if (heading.includes('Implementation Audit')) { + auditCount += 1; + body = auditCount < 3 + ? `Needs more work ${auditCount}.\n` + : 'Approved.\n'; + } + await appendFile(filePath, `\n## ${heading}\n\n${body}\n`, 'utf8'); + setTimeout(() => notifySessionIdle(session), 20); + }); + + const run = await startP2pRun({ + initiatorSession: 'deck_proj_brain', + targets: [{ session: 'deck_proj_w1', mode: 'audit' }], + userText: 'artifact cleanup across repeated loop attempts', + fileContents: [], + serverLink: serverLinkMock as any, + advancedRounds: [ + { + id: 'implementation', + title: 'Implementation', + preset: 'implementation', + executionMode: 'multi_dispatch', + permissionScope: 'implementation', + }, + { + id: 'implementation_audit', + title: 'Implementation Audit', + preset: 'implementation_audit', + executionMode: 'single_main', + permissionScope: 'analysis_only', + verdictPolicy: 'smart_gate', + jumpRule: { + targetRoundId: 'implementation', + marker: 'REWORK', + minTriggers: 0, + maxTriggers: 3, + }, + }, + ], + }); + + const done = await waitForStatus(run.id, ['completed'], 25_000); + expect(done.roundAttemptCounts.implementation).toBe(3); + await waitForNoRoundHopArtifacts(tempProjectDir, done.id); + }, 30_000); + + it('cleans worker-hop artifacts even when advanced synthesis fails before the run completes', async () => { + getSessionMock.mockImplementation((name: string) => { + if (name === 'deck_proj_brain') return { agentType: 'claude-code-sdk', projectDir: tempProjectDir, parentSession: undefined, label: 'brain' }; + if (name === 'deck_proj_w1') return { agentType: 'claude-code-sdk', projectDir: tempProjectDir, parentSession: undefined, label: 'w1' }; + return null; + }); + + sendKeysDelayedEnterMock.mockImplementation(async (session: string, prompt: string) => { + const filePath = pathFromPrompt(prompt); + const heading = headingFromPrompt(prompt); + if (heading.includes('Implementation Synthesis')) { + throw new Error('synthesis unavailable'); + } + await appendFile(filePath, `\n## ${heading}\n\nOutput from ${session}.\n`, 'utf8'); + setTimeout(() => notifySessionIdle(session), 20); + }); + + const run = await startP2pRun({ + initiatorSession: 'deck_proj_brain', + targets: [{ session: 'deck_proj_w1', mode: 'audit' }], + userText: 'cleanup worker artifacts on synthesis failure', + fileContents: [], + serverLink: serverLinkMock as any, + advancedRounds: [ + { + id: 'implementation', + title: 'Implementation', + preset: 'implementation', + executionMode: 'multi_dispatch', + permissionScope: 'implementation', + }, + ], + }); + + const done = await waitForStatus(run.id, ['failed'], 20_000); + expect(done.error).toContain('synthesis unavailable'); + await waitForNoRoundHopArtifacts(tempProjectDir, done.id); + }, 25_000); }); diff --git a/test/daemon/p2p-parser.test.ts b/test/daemon/p2p-parser.test.ts index b3e71ddb..916fe880 100644 --- a/test/daemon/p2p-parser.test.ts +++ b/test/daemon/p2p-parser.test.ts @@ -264,7 +264,7 @@ describe('structured P2P routing via WS fields', () => { await new Promise((resolve) => setTimeout(resolve, 0)); expect(startP2pRun).toHaveBeenCalledTimes(1); - const [_initiator, targets] = (startP2pRun as ReturnType).mock.calls[0]; + const [{ targets }] = (startP2pRun as ReturnType).mock.calls[0]; expect(targets).toEqual([ { session: 'deck_proj_w1', mode: 'brainstorm>discuss' }, ]); @@ -286,7 +286,7 @@ describe('structured P2P routing via WS fields', () => { await new Promise((resolve) => setTimeout(resolve, 0)); expect(startP2pRun).toHaveBeenCalledTimes(1); - const [_initiator, targets] = (startP2pRun as ReturnType).mock.calls[0]; + const [{ targets }] = (startP2pRun as ReturnType).mock.calls[0]; expect(targets).toEqual([ { session: 'deck_proj_w1', mode: 'audit' }, { session: 'deck_proj_w2', mode: 'review' }, @@ -306,7 +306,8 @@ describe('structured P2P routing via WS fields', () => { await new Promise((r) => setTimeout(r, 100)); expect(startP2pRun).toHaveBeenCalledOnce(); - const [_initiator, targets, cleanText] = (startP2pRun as ReturnType).mock.calls[0]; + expect((startP2pRun as ReturnType).mock.calls[0]).toHaveLength(1); + const [{ targets, userText: cleanText }] = (startP2pRun as ReturnType).mock.calls[0]; expect(targets.length).toBeGreaterThan(0); expect(targets.every((t: any) => t.mode === 'audit')).toBe(true); // Text should be clean — no @@tokens @@ -326,7 +327,7 @@ describe('structured P2P routing via WS fields', () => { await new Promise((r) => setTimeout(r, 100)); expect(startP2pRun).toHaveBeenCalledOnce(); - const [_initiator, targets, cleanText] = (startP2pRun as ReturnType).mock.calls[0]; + const [{ targets, userText: cleanText }] = (startP2pRun as ReturnType).mock.calls[0]; expect(targets).toEqual([ { session: 'deck_proj_w1', mode: 'review' }, ]); @@ -348,10 +349,55 @@ describe('structured P2P routing via WS fields', () => { await new Promise((r) => setTimeout(r, 100)); expect(startP2pRun).toHaveBeenCalledOnce(); - const extraPrompt = (startP2pRun as ReturnType).mock.calls[0][6]; + expect((startP2pRun as ReturnType).mock.calls[0]).toHaveLength(1); + const [{ extraPrompt }] = (startP2pRun as ReturnType).mock.calls[0]; expect(extraPrompt).toContain("Use the user's selected i18n language (Chinese (Simplified)) for the discussion."); }); + it('forwards advanced p2p options through the structured session.send path', async () => { + const advancedRounds = [ + { + id: 'implementation', + title: 'Implementation', + preset: 'implementation', + executionMode: 'single_main', + permissionScope: 'implementation', + }, + ]; + + handleWebCommand({ + type: 'session.send', + sessionName: 'deck_proj_brain', + text: 'run advanced p2p', + commandId: 'cmd-advanced-1', + p2pAtTargets: [{ session: 'deck_proj_w1', mode: 'audit' }], + p2pAdvancedPresetKey: 'openspec', + p2pAdvancedRounds: advancedRounds as any, + p2pAdvancedRunTimeoutMinutes: 45, + p2pContextReducer: { + mode: 'clone_sdk_session', + templateSession: 'deck_proj_brain', + }, + }, mockServerLink as any); + + await new Promise((r) => setTimeout(r, 100)); + + expect(startP2pRun).toHaveBeenCalledOnce(); + expect((startP2pRun as ReturnType).mock.calls[0]).toHaveLength(1); + expect((startP2pRun as ReturnType).mock.calls[0]?.[0]).toMatchObject({ + initiatorSession: 'deck_proj_brain', + targets: [{ session: 'deck_proj_w1', mode: 'audit' }], + userText: 'run advanced p2p', + advancedPresetKey: 'openspec', + advancedRounds, + advancedRunTimeoutMs: 45 * 60_000, + contextReducer: { + mode: 'clone_sdk_session', + templateSession: 'deck_proj_brain', + }, + }); + }); + it('structured p2pAtTargets stays authoritative for single-target P2P runs', async () => { handleWebCommand({ type: 'session.send', @@ -364,7 +410,7 @@ describe('structured P2P routing via WS fields', () => { await new Promise((r) => setTimeout(r, 100)); expect(startP2pRun).toHaveBeenCalledOnce(); - const [_initiator, targets, cleanText] = (startP2pRun as ReturnType).mock.calls[0]; + const [{ targets, userText: cleanText }] = (startP2pRun as ReturnType).mock.calls[0]; expect(targets).toEqual([ { session: 'deck_proj_w1', mode: 'review' }, ]); @@ -387,7 +433,7 @@ describe('structured P2P routing via WS fields', () => { await new Promise((r) => setTimeout(r, 100)); expect(startP2pRun).toHaveBeenCalledOnce(); - const [_initiator, targets, cleanText] = (startP2pRun as ReturnType).mock.calls[0]; + const [{ targets, userText: cleanText }] = (startP2pRun as ReturnType).mock.calls[0]; expect(targets).toEqual([ { session: 'deck_proj_w2', mode: 'discuss' }, { session: 'deck_proj_w1', mode: 'audit' }, @@ -407,7 +453,7 @@ describe('structured P2P routing via WS fields', () => { await new Promise((r) => setTimeout(r, 100)); expect(startP2pRun).toHaveBeenCalledOnce(); - const [_initiator, targets, cleanText] = (startP2pRun as ReturnType).mock.calls[0]; + const [{ targets, userText: cleanText }] = (startP2pRun as ReturnType).mock.calls[0]; expect(targets.length).toBeGreaterThan(0); expect(targets.every((t: any) => t.mode === 'brainstorm')).toBe(true); expect(cleanText).toBe('brainstorm ideas'); @@ -440,7 +486,7 @@ describe('structured P2P routing via WS fields', () => { await new Promise((r) => setTimeout(r, 100)); expect(startP2pRun).toHaveBeenCalledOnce(); - const [_initiator, targets, cleanText] = (startP2pRun as ReturnType).mock.calls[0]; + const [{ targets, userText: cleanText }] = (startP2pRun as ReturnType).mock.calls[0]; expect(targets).toEqual([ { session: 'deck_proj_w1', mode: 'audit' }, ]); @@ -469,7 +515,7 @@ describe('structured P2P routing via WS fields', () => { await new Promise((r) => setTimeout(r, 100)); expect(startP2pRun).toHaveBeenCalledOnce(); - const [_initiator, targets] = (startP2pRun as ReturnType).mock.calls[0]; + const [{ targets }] = (startP2pRun as ReturnType).mock.calls[0]; expect(targets.length).toBeGreaterThan(0); expect(targets.every((t: any) => t.mode === 'audit')).toBe(true); }); @@ -485,7 +531,7 @@ describe('structured P2P routing via WS fields', () => { await new Promise((r) => setTimeout(r, 100)); expect(startP2pRun).toHaveBeenCalledOnce(); - const call = (startP2pRun as ReturnType).mock.calls[0]; - expect(call[5]).toBe(2); + const [{ rounds }] = (startP2pRun as ReturnType).mock.calls[0]; + expect(rounds).toBe(2); }); }); diff --git a/test/shared/p2p-advanced.test.ts b/test/shared/p2p-advanced.test.ts new file mode 100644 index 00000000..04271ab8 --- /dev/null +++ b/test/shared/p2p-advanced.test.ts @@ -0,0 +1,184 @@ +import { describe, expect, it } from 'vitest'; + +import { + BUILT_IN_ADVANCED_PRESETS, + resolveP2pRoundPlan, + type P2pAdvancedRound, +} from '../../shared/p2p-advanced.js'; + +describe('resolveP2pRoundPlan', () => { + it('uses the shared default hop timeout for legacy runs when no override is provided', () => { + const plan = resolveP2pRoundPlan({ + modeOverride: 'audit', + }); + + expect(plan.advanced).toBe(false); + expect(plan.rounds).toHaveLength(1); + expect(plan.rounds[0]?.timeoutMinutes).toBe(8); + }); + + it('preserves legacy combo behavior when advanced config is absent', () => { + const plan = resolveP2pRoundPlan({ + modeOverride: 'brainstorm>discuss', + roundsOverride: 2, + hopTimeoutMinutes: 8, + }); + + expect(plan.advanced).toBe(false); + expect(plan.rounds).toHaveLength(2); + expect(plan.rounds.map((round) => round.modeKey)).toEqual(['brainstorm', 'discuss']); + expect(plan.rounds.every((round) => round.timeoutMinutes === 8)).toBe(true); + }); + + it('resolves the openspec preset and freezes sdk helper eligibility', () => { + const plan = resolveP2pRoundPlan({ + advancedPresetKey: 'openspec', + advancedRunTimeoutMinutes: 45, + contextReducer: { + mode: 'clone_sdk_session', + templateSession: 'deck_proj_brain', + }, + participants: [ + { sessionName: 'deck_proj_brain', agentType: 'claude-code-sdk' }, + { sessionName: 'deck_sub_worker', agentType: 'qwen', parentSession: 'deck_proj_brain' }, + { sessionName: 'deck_proj_cli', agentType: 'codex' }, + ], + }); + + expect(plan.advanced).toBe(true); + expect(plan.rounds.map((round) => round.id)).toEqual(BUILT_IN_ADVANCED_PRESETS.openspec.map((round) => round.id)); + expect(plan.contextReducer).toEqual({ + mode: 'clone_sdk_session', + templateSession: 'deck_proj_brain', + }); + expect(plan.helperEligibleSnapshot).toEqual([ + { sessionName: 'deck_proj_brain', agentType: 'claude-code-sdk' }, + { sessionName: 'deck_sub_worker', agentType: 'qwen', parentSession: 'deck_proj_brain' }, + ]); + expect(plan.overallRunTimeoutMinutes).toBe(45); + }); + + it('rejects forward jumps in advanced rounds', () => { + const rounds: P2pAdvancedRound[] = [ + { + id: 'implement', + title: 'Implement', + preset: 'implementation', + executionMode: 'multi_dispatch', + permissionScope: 'implementation', + }, + { + id: 'audit', + title: 'Audit', + preset: 'implementation_audit', + executionMode: 'single_main', + permissionScope: 'analysis_only', + verdictPolicy: 'smart_gate', + jumpRule: { + targetRoundId: 'audit', + marker: 'REWORK', + minTriggers: 0, + maxTriggers: 2, + }, + }, + ]; + + expect(() => resolveP2pRoundPlan({ advancedRounds: rounds })).toThrow(/jump backward/i); + }); + + it('rejects artifact-generation rounds without declared outputs', () => { + const rounds: P2pAdvancedRound[] = [ + { + id: 'custom_artifact', + title: 'Custom Artifact', + preset: 'custom', + executionMode: 'single_main', + permissionScope: 'artifact_generation', + }, + ]; + + expect(() => resolveP2pRoundPlan({ advancedRounds: rounds })).toThrow(/must declare artifact outputs/i); + }); + + it('rejects forced_rework rounds whose maxTriggers are below minTriggers', () => { + const rounds: P2pAdvancedRound[] = [ + { + id: 'implementation', + title: 'Implementation', + preset: 'implementation', + executionMode: 'single_main', + permissionScope: 'implementation', + }, + { + id: 'implementation_audit', + title: 'Implementation Audit', + preset: 'implementation_audit', + executionMode: 'single_main', + permissionScope: 'analysis_only', + verdictPolicy: 'forced_rework', + jumpRule: { + targetRoundId: 'implementation', + marker: 'REWORK', + minTriggers: 2, + maxTriggers: 1, + }, + }, + ]; + + expect(() => resolveP2pRoundPlan({ advancedRounds: rounds })).toThrow(/invalid maxTriggers/i); + }); + + it('rejects proposal_audit rounds that attempt to drive routing', () => { + const rounds: P2pAdvancedRound[] = [ + { + id: 'discussion', + title: 'Discussion', + preset: 'discussion', + executionMode: 'multi_dispatch', + permissionScope: 'analysis_only', + }, + { + id: 'proposal_audit', + title: 'Proposal Audit', + preset: 'proposal_audit', + executionMode: 'single_main', + permissionScope: 'analysis_only', + verdictPolicy: 'smart_gate', + jumpRule: { + targetRoundId: 'discussion', + marker: 'REWORK', + minTriggers: 0, + maxTriggers: 2, + }, + }, + ]; + + expect(() => resolveP2pRoundPlan({ advancedRounds: rounds })).toThrow(/proposal_audit cannot drive routing/i); + }); + + it('rejects reducer sessions that are not sdk-backed participants', () => { + expect(() => resolveP2pRoundPlan({ + advancedPresetKey: 'openspec', + contextReducer: { + mode: 'reuse_existing_session', + sessionName: 'deck_proj_cli', + }, + participants: [ + { sessionName: 'deck_proj_cli', agentType: 'codex' }, + ], + })).toThrow(/eligible SDK-backed participant/i); + }); + + it('rejects clone-mode reducer templates that are not sdk-backed participants', () => { + expect(() => resolveP2pRoundPlan({ + advancedPresetKey: 'openspec', + contextReducer: { + mode: 'clone_sdk_session', + templateSession: 'deck_proj_cli', + }, + participants: [ + { sessionName: 'deck_proj_cli', agentType: 'codex' }, + ], + })).toThrow(/eligible SDK-backed participant/i); + }); +}); diff --git a/web/src/app.tsx b/web/src/app.tsx index 9eacd124..bda87eae 100644 --- a/web/src/app.tsx +++ b/web/src/app.tsx @@ -5,66 +5,8 @@ import { type FileBrowserPreviewState, type FileBrowserPreviewUpdate, } from './components/FileBrowser.js'; -import { mapP2pStatusToUiState, type P2pActivePhase, type P2pProgressNodeStatus } from '@shared/p2p-status.js'; import { DAEMON_MSG } from '@shared/daemon-events.js'; - -function mapP2pRunToDiscussion(r: Record) { - const rawSnapshot = r.progress_snapshot; - const snapshot = typeof rawSnapshot === 'string' - ? (() => { try { return JSON.parse(rawSnapshot) as Record; } catch { return {}; } })() - : (rawSnapshot ?? {}); - const source = { ...r, ...snapshot } as Record; - const id = `p2p_${source.id}`; - const status = String(source.status ?? ''); - const state = mapP2pStatusToUiState(status); - const mode = source.mode_key ?? 'discuss'; - const currentRoundMode = source.current_round_mode ?? mode; - const initiatorLabel = source.initiator_label ?? 'brain'; - const currentTarget = source.current_target_label ?? (source.current_target_session ? String(source.current_target_session).split('_').pop() : undefined); - const totalCount = source.total_count ?? 3; - const totalHops = source.total_hops ?? Math.max(0, totalCount - 2); - const nodes = Array.isArray(source.all_nodes) ? source.all_nodes.map((n: any) => ({ - session: typeof n.session === 'string' ? n.session : undefined, - label: String(n.label ?? ''), - displayLabel: String(n.displayLabel ?? n.display_label ?? n.label ?? ''), - agentType: String(n.agentType ?? ''), - ccPreset: n.ccPreset ?? n.cc_preset ?? null, - mode: typeof n.mode === 'string' ? n.mode : undefined, - phase: typeof n.phase === 'string' ? n.phase as 'initial' | 'hop' | 'summary' : undefined, - status: String(n.status ?? 'pending') as P2pProgressNodeStatus, - })) : undefined; - const hopStates = Array.isArray(source.hop_states) ? source.hop_states.map((hop: any) => ({ - hopIndex: Number(hop.hop_index ?? 0), - roundIndex: Number(hop.round_index ?? 0), - session: typeof hop.session === 'string' ? hop.session : undefined, - mode: typeof hop.mode === 'string' ? hop.mode : undefined, - status: String(hop.status ?? 'queued') as 'queued' | 'dispatched' | 'running' | 'completed' | 'timed_out' | 'failed' | 'cancelled', - })) : undefined; - return { - id, - topic: `P2P ${mode} · ${initiatorLabel}`, - state, - modeKey: currentRoundMode, - currentRound: source.current_round ?? 1, - maxRounds: source.total_rounds ?? 1, - completedHops: source.completed_hops_count ?? 0, - completedRoundHops: typeof source.completed_round_hops_count === 'number' ? source.completed_round_hops_count : undefined, - totalHops, - activeHop: source.active_hop_number ?? null, - activeRoundHop: source.active_round_hop_number ?? null, - activePhase: (typeof source.active_phase === 'string' ? source.active_phase : 'queued') as P2pActivePhase, - initiatorLabel, - currentSpeaker: currentTarget, - conclusion: state === 'done' ? (source.result_summary ?? undefined) : undefined, - error: state === 'failed' ? (source.error ?? undefined) : undefined, - filePath: undefined, - fileId: source.discussion_id ? String(source.discussion_id) : source.id, - startedAt: source.created_at ? new Date(source.created_at).getTime() : undefined, - hopStartedAt: typeof source.hop_started_at === 'number' ? source.hop_started_at : undefined, - nodes, - hopStates, - }; -} +import { mapP2pRunToDiscussion } from './p2p-run-mapping.js'; import { useTranslation } from 'react-i18next'; import { ErrorBoundary } from './components/ErrorBoundary.js'; import { LanguageSwitcher } from './components/LanguageSwitcher.js'; diff --git a/web/src/components/P2pConfigPanel.tsx b/web/src/components/P2pConfigPanel.tsx index 61734c1c..f0943901 100644 --- a/web/src/components/P2pConfigPanel.tsx +++ b/web/src/components/P2pConfigPanel.tsx @@ -2,12 +2,19 @@ * P2pConfigPanel — modal settings panel for P2P config mode. * Lets the user configure per-session participation and modes, plus round count. */ -import { useState, useEffect } from 'preact/hooks'; +import { useState, useEffect, useMemo } from 'preact/hooks'; import { useTranslation } from 'react-i18next'; import { getUserPref, saveUserPref } from '../api.js'; import { P2pComboManager } from './P2pComboManager.js'; import { useP2pCustomCombos } from './p2p-combos.js'; import type { P2pSavedConfig, P2pSessionConfig } from '@shared/p2p-modes.js'; +import { BUILT_IN_ADVANCED_PRESETS } from '@shared/p2p-advanced.js'; +import type { + P2pAdvancedPresetKey, + P2pAdvancedRound, + P2pContextReducerConfig, + P2pContextReducerMode, +} from '@shared/p2p-advanced.js'; interface SessionRow { name: string; @@ -156,6 +163,25 @@ const sectionCardStyle: Record = { padding: 14, }; +const fieldLabelStyle: Record = { + display: 'flex', + flexDirection: 'column', + gap: 6, + fontSize: 12, + color: '#cbd5e1', +}; + +const fieldInputStyle: Record = { + width: '100%', + background: '#0f172a', + border: '1px solid #334155', + borderRadius: 6, + color: '#e2e8f0', + fontSize: 13, + padding: '7px 9px', + outline: 'none', +}; + const roundsBtnStyle = (active: boolean): Record => ({ padding: '4px 12px', borderRadius: 6, @@ -253,9 +279,24 @@ export function P2pConfigPanel({ const [rounds, setRounds] = useState(3); const [hopTimeoutMinutes, setHopTimeoutMinutes] = useState(8); const [extraPrompt, setExtraPrompt] = useState(''); + const [advancedExpanded, setAdvancedExpanded] = useState(false); + const [advancedPresetKey, setAdvancedPresetKey] = useState(''); + const [advancedRounds, setAdvancedRounds] = useState(undefined); + const [advancedRunTimeoutMinutes, setAdvancedRunTimeoutMinutes] = useState(30); + const [contextReducerMode, setContextReducerMode] = useState(''); + const [contextReducerSession, setContextReducerSession] = useState(''); + const [contextReducerTemplate, setContextReducerTemplate] = useState(''); const [loading, setLoading] = useState(true); const [saving, setSaving] = useState(false); + const enabledSdkParticipants = useMemo( + () => allEligible.filter((entry) => entry.flavor === 'sdk').filter((entry) => { + const cfg = sessionCfg[entry.key]; + return !!cfg?.enabled && cfg.mode !== 'skip'; + }), + [allEligible, sessionCfg], + ); + // Config key uses the main session (sub-sessions follow parent config) const configKey = scopeSession ? `p2p_session_config:${scopeSession}` : null; @@ -271,6 +312,13 @@ export function P2pConfigPanel({ setRounds(parsed.rounds ?? 3); setHopTimeoutMinutes(parsed.hopTimeoutMinutes ?? 8); setExtraPrompt(parsed.extraPrompt ?? ''); + setAdvancedPresetKey(parsed.advancedPresetKey ?? ''); + setAdvancedRounds(parsed.advancedRounds); + setAdvancedRunTimeoutMinutes(parsed.advancedRunTimeoutMinutes ?? 30); + setContextReducerMode(parsed.contextReducer?.mode ?? ''); + setContextReducerSession(parsed.contextReducer?.sessionName ?? ''); + setContextReducerTemplate(parsed.contextReducer?.templateSession ?? ''); + setAdvancedExpanded(Boolean(parsed.advancedPresetKey || parsed.contextReducer || parsed.advancedRunTimeoutMinutes != null)); } catch { /* start fresh */ } } setLoading(false); @@ -287,6 +335,21 @@ export function P2pConfigPanel({ }); }, [configKey]); + useEffect(() => { + if (contextReducerMode === 'reuse_existing_session') { + const stillEligible = enabledSdkParticipants.some((entry) => entry.key === contextReducerSession); + if (!stillEligible) setContextReducerSession(enabledSdkParticipants[0]?.key ?? ''); + return; + } + if (contextReducerMode === 'clone_sdk_session') { + const stillEligible = enabledSdkParticipants.some((entry) => entry.key === contextReducerTemplate); + if (!stillEligible) setContextReducerTemplate(enabledSdkParticipants[0]?.key ?? ''); + return; + } + setContextReducerSession(''); + setContextReducerTemplate(''); + }, [contextReducerMode, contextReducerSession, contextReducerTemplate, enabledSdkParticipants]); + useEffect(() => { if (typeof window === 'undefined') return; const handleResize = () => setIsMobile(window.innerWidth < 768); @@ -317,7 +380,25 @@ export function P2pConfigPanel({ for (const e of allEligible) { merged[e.key] = sessionCfg[e.key] ?? { enabled: false, mode: 'audit' }; } - const cfg: P2pSavedConfig = { sessions: merged, rounds, hopTimeoutMinutes, extraPrompt: extraPrompt.trim() || undefined }; + let contextReducer: P2pContextReducerConfig | undefined; + if (advancedPresetKey && contextReducerMode === 'reuse_existing_session' && contextReducerSession) { + contextReducer = { mode: 'reuse_existing_session', sessionName: contextReducerSession }; + } else if (advancedPresetKey && contextReducerMode === 'clone_sdk_session' && contextReducerTemplate) { + contextReducer = { mode: 'clone_sdk_session', templateSession: contextReducerTemplate }; + } + const resolvedAdvancedRounds = advancedPresetKey + ? (advancedRounds ? JSON.parse(JSON.stringify(advancedRounds)) as P2pAdvancedRound[] : JSON.parse(JSON.stringify(BUILT_IN_ADVANCED_PRESETS[advancedPresetKey])) as P2pAdvancedRound[]) + : undefined; + const cfg: P2pSavedConfig = { + sessions: merged, + rounds, + hopTimeoutMinutes, + extraPrompt: extraPrompt.trim() || undefined, + advancedPresetKey: advancedPresetKey || undefined, + advancedRounds: resolvedAdvancedRounds, + advancedRunTimeoutMinutes: advancedPresetKey ? advancedRunTimeoutMinutes : undefined, + contextReducer, + }; try { if (configKey) await saveUserPref(configKey, JSON.stringify(cfg)); onSave(cfg); @@ -327,6 +408,26 @@ export function P2pConfigPanel({ }; const getEntry = (key: string) => sessionCfg[key] ?? { enabled: false, mode: 'audit' }; + const handleAdvancedPresetChange = (value: string) => { + const nextPreset = value as P2pAdvancedPresetKey | ''; + setAdvancedPresetKey(nextPreset); + if (!nextPreset) { + setAdvancedRounds(undefined); + setContextReducerMode(''); + setContextReducerSession(''); + setContextReducerTemplate(''); + return; + } + setAdvancedRounds((prev) => prev ?? (JSON.parse(JSON.stringify(BUILT_IN_ADVANCED_PRESETS[nextPreset])) as P2pAdvancedRound[])); + setAdvancedRunTimeoutMinutes((prev) => (prev > 0 ? prev : 30)); + }; + + const updateAdvancedRound = (roundId: string, updater: (round: P2pAdvancedRound) => P2pAdvancedRound) => { + setAdvancedRounds((prev) => { + if (!prev) return prev; + return prev.map((round) => (round.id === roundId ? updater(round) : round)); + }); + }; const overlayStyle: Record = { position: 'fixed', inset: 0, @@ -505,6 +606,317 @@ export function P2pConfigPanel({ }} />
+ +
+ + {advancedExpanded && ( +
+ + + {advancedPresetKey && ( + <> + + + + + {contextReducerMode === 'reuse_existing_session' && ( + + )} + + {contextReducerMode === 'clone_sdk_session' && ( + + )} + + {advancedRounds && advancedRounds.length > 0 && ( +
+
+ {t('p2p.settings_advanced_rounds', 'Advanced rounds')} +
+ {advancedRounds.map((round) => ( +
+
+
+
{round.title}
+
{round.id} · {round.preset}
+
+ +
+ +
+ + +
+ + {(round.verdictPolicy === 'smart_gate' || round.verdictPolicy === 'forced_rework' || round.jumpRule) && ( +
+ + + + +
+ )} + + {(round.permissionScope === 'artifact_generation' || (round.artifactOutputs?.length ?? 0) > 0) && ( + + )} + +