Consolidate PRs #1 + #2; ship AI SDK provider; drop streamAgent/generateAgent#3
Merged
Consolidate PRs #1 + #2; ship AI SDK provider; drop streamAgent/generateAgent#3
Conversation
- Picks up reliability fixes from 0.17→0.20: - PR #103/#111: clean transport-failure handling, propagate ndjson errors - PR #119: parse final NDJSON message without trailing newline - PR #122: no spurious unhandledRejection on transport failures - PR #127: TS private keyword for cross-copy compatibility - PR #130: response/notification ordering fix - Renames stabilized methods (PR #132): - unstable_resumeSession → resumeSession - unstable_closeSession → closeSession - Updates LocalAgent, mock-agent harness, echo-agent fixture - Adds resume-close.test.ts regression tests Co-authored-by: Aiden Bai <aidenybai@users.noreply.github.com>
- connect.ts:
- Wait for spawn or error event before constructing connection
(avoids races with later spawn errors leaking unhandled rejections).
- Add disposeGraceMs option (default unchanged at 2000ms).
- Add onTrace hook for inspecting in/out JSON-RPC messages and stderr.
- Add envFilter hook for env scrubbing.
- local-agent.ts:
- Race connection.initialize() against subprocess close so a process
that exits before responding fails fast as AgentConnectionClosedError
(instead of waiting initializeTimeoutMs).
- Gate stderr-pattern fatal detection (auth/usage) until after init
succeeds, eliminating false positives during boot/login banners.
- Allow callers to override stderrFatalPatterns.
- utils/run-command.ts: Escalate to SIGKILL after a grace when SIGTERM
is ignored on timeout.
- New tests:
- tests/spawn-failures.test.ts: ENOENT, fast subprocess exit
- tests/stderr-fatal-detection.test.ts: pre/post-init stderr gating
Co-authored-by: Aiden Bai <aidenybai@users.noreply.github.com>
- Inactivity watchdog: pause while a permission request is pending in
stream mode (prevents killing sessions waiting on the user). Permission
events also bump lastActivityAt.
- Per-session buffered update cap (MAX_BUFFERED_UPDATES_PER_SESSION = 1024)
drops oldest first to prevent unbounded memory growth.
- Add loadSessionStreaming() returning { sessionId, replay, completion }
so callers can iterate replay updates emitted during session/load BEFORE
the first prompt. Existing loadSession() preserved for back-compat.
- Slash commands:
- Track latest available_commands_update per session.
- Add LocalAgent.commandsFor(sessionId) / modeStateFor() / configOptionsFor() getters.
- PromptInput now accepts { command: { name, input? } } and validates
against advertised commands when known.
- Error message preservation: stringifyCause now includes Error.cause
message in wrapped errors when distinct.
- Late tool_call_update notifications received after session/cancel are
forwarded through the iterator (already worked in practice; explicit
test coverage added).
- Mode tracking: current_mode_update mutates the cached modeState.
- New tests:
- tests/late-updates.test.ts
- tests/watchdog-permission.test.ts
- tests/buffer-cap.test.ts
- tests/load-session-replay.test.ts
- tests/slash-commands.test.ts
Co-authored-by: Aiden Bai <aidenybai@users.noreply.github.com>
- Optional terminal/* support: when caller provides options.terminal
handlers, advertise terminal capability and forward all five client
terminal methods (create/output/release/wait_for_exit/kill).
- additionalDirectories on createSession / loadSession /
loadSessionStreaming / resumeSession with capability gating
(sessionCapabilities.additionalDirectories) and absolute-path validation.
- listSessions(input?: { cwd?, cursor? }) with pagination plumbing.
Adds streamAllSessions helper that auto-paginates.
- _meta trace context propagation via opt-in traceContext: () => Record.
Only the W3C-reserved keys traceparent / tracestate / baggage are
forwarded; other keys filtered out.
- onAuthRequired hook: when newSession returns auth_required (-32000)
and a hook is configured, prompt for an auth method, call authenticate,
then retry once. Returning undefined from the hook preserves the
original AgentUnauthenticatedError.
- Convenience getters: commandsFor / modeStateFor / configOptionsFor.
- Mode tracking: current_mode_update mutates cached modeState.
- Tests:
- tests/terminal.test.ts
- tests/additional-directories.test.ts
- tests/list-sessions-pagination.test.ts
- tests/trace-context.test.ts
- tests/auth-retry.test.ts
Co-authored-by: Aiden Bai <aidenybai@users.noreply.github.com>
- claude: ANTHROPIC_API_KEY env fallback bypasses 'claude auth status' parsing when the env is set; checkAuthenticated honors it. - copilot: GITHUB_TOKEN env fallback bypasses 'gh auth token' lookup; helpful error message updated. - gemini: GEMINI_API_KEY / GOOGLE_API_KEY env fallback before reading ~/.gemini/google_accounts.json; updated error message. (envFilter hook from Phase 2 already exposed for env scrubbing.) - New: tests/adapter-env-fallbacks.test.ts covering gemini env auth. Co-authored-by: Aiden Bai <aidenybai@users.noreply.github.com>
…aul (Phase 6) - Bump apps/playground SDK pin to ^0.20.0 for consistency - Document new options (terminal, additionalDirectories, listSessions pagination, traceContext, onAuthRequired, onTrace, disposeGraceMs, envFilter) in README - Document slash command helper and reliability features - Add .changeset entry summarizing the multi-phase upgrade - New tests: - tests/sdk-upgrade.test.ts (covers ordering through wrapper) End state: 70 unit tests green; 8/8 e2e Playwright tests against the real echo-agent subprocess green. Co-authored-by: Aiden Bai <aidenybai@users.noreply.github.com>
…ants Co-authored-by: Aiden Bai <aidenybai@users.noreply.github.com>
- Add testing/mock-agent.ts helpers:
- withSpawnFailure(id?): builds an AgentAdapter that always points at
a missing binary (used by spawn-failures.test.ts)
- withSdkOutOfOrderRace(chunks): builds a prompt handler that emits
multiple notifications back-to-back, exercising the SDK's response/
notification ordering guarantee.
- tests/spawn-failures.test.ts now uses withSpawnFailure for clarity.
- tests/wire-fuzz.test.ts: bounded property-based test that feeds 200
randomized session updates per turn (seeded RNG) plus an interleaved
cancellation variant to verify the wrapper does not crash and produces
well-shaped AgentEvent values terminating in 'finish'.
Total: 72 unit tests across 24 files.
Co-authored-by: Aiden Bai <aidenybai@users.noreply.github.com>
…irectories Closes the documentation gap from the plan's file change matrix: - Documents the echo-agent fixture commands used by the e2e tests. - Shows fileSystem and terminal handler wiring patterns for callers. - Documents additionalDirectories session-input field and absolute-path requirement. Co-authored-by: Aiden Bai <aidenybai@users.noreply.github.com>
…ffer, validation timer leak, pending-update cap, abort short-circuit, perm settle gate, EPIPE handlers, stdout-noise filter, run-command caps Co-authored-by: Aiden Bai <aidenybai@users.noreply.github.com>
Co-authored-by: Aiden Bai <aidenybai@users.noreply.github.com>
Co-authored-by: Aiden Bai <aidenybai@users.noreply.github.com>
Resolves overlap between the SDK upgrade / spec coverage overhaul and the reliability bug-fix pass into a single tree. Conflicts resolved (favoring stricter / more correct behavior): - run-command.ts: keep PR #2's isProcessAlive check + PR #1's configurable grace - connect.ts: combine waitForSpawnOrError, no-op stream error listeners, and isProcessAlive-based dispose with disposeGraceMs option - local-agent.ts: - merge option fields (pendingUpdateBufferLimit + traceContext + onAuthRequired + onTrace + envFilter + disposeGraceMs + stderrFatalPatterns) - keep PR #2's raceInitialize helper (subprocess-close vs init vs timeout) while preserving PR #1's defensive AgentConnectionClosedError fallback - validate prompt blocks (with slash-command processing) before draining buffered updates / pendingConfigOptions - use PR #2's bufferSessionUpdate with droppedPendingUpdateCount and the configurable PENDING_UPDATE_BUFFER_LIMIT_PER_SESSION cap - constants.ts: keep PENDING_UPDATE_BUFFER_LIMIT_PER_SESSION; drop the duplicate MAX_BUFFERED_UPDATES_PER_SESSION - buffer-cap.test.ts: realign to the PENDING_UPDATE_BUFFER_LIMIT_PER_SESSION constant - changesets: collapse acp-reliability-overhaul + reliability-hardening into a single comprehensive entry Verified: pnpm typecheck, pnpm lint (0/0), pnpm test (86/86 unit tests across 25 files), and the playground node project (8/8 e2e tests against the real echo-agent subprocess) all pass.
…ent/generateAgent
Implements the LanguageModelV3 spec (`@ai-sdk/provider` v3.0.x) so
`use-local-agent` plugs directly into `streamText` / `generateText` from
the `ai` package, eliminating the parallel one-shot helper API:
import { streamText } from "ai";
import { localAgent } from "use-local-agent";
const { textStream } = streamText({
model: localAgent("claude"),
prompt: "Refactor src/auth.ts",
});
New module `src/ai-sdk/`:
- `convert-prompt.ts` maps a `LanguageModelV3Prompt` (system/user/
assistant/tool messages) to ACP `ContentBlock[]`. System messages
flatten into `systemPrompt`; image/audio file parts become ACP image/
audio blocks; unsupported file mediaTypes, assistant tool-call replays,
and tool-result messages emit `unsupported` warnings.
- `local-agent-language-model.ts` implements `LanguageModelV3`:
- `doGenerate` returns ordered `reasoning` + `text` + `tool-call`
content with mapped `LanguageModelV3FinishReason` and usage.
- `doStream` emits `stream-start` (with warnings), `response-metadata`
(carrying ACP `sessionId`), grouped `text-start/-delta/-end` and
`reasoning-start/-delta/-end`, `tool-call`, `tool-result`, `finish`,
and `error` parts. Opt-in raw passthrough via `includeRawChunks`.
- Takes a `connect` factory so it's testable through the existing
`connectMockAgent` harness without spawning a subprocess.
- `provider.ts` exports `createLocalAgentProvider()` (a callable
`ProviderV3` with a `.languageModel()` method, plus `.fromAdapter()`
for custom ACP-speaking subprocesses) and a default `localAgent`
instance. `embeddingModel` / `imageModel` throw `NoSuchModelError`.
Default `permission: "auto-allow"` so `streamText` Just Works.
Mappings:
- ACP `text-delta` → V3 `text-delta` (start/end framed by message id)
- ACP `thinking-delta` → V3 `reasoning-delta`
- ACP `tool-call` → V3 `tool-call` with `providerExecuted: true`
- ACP `tool-call-update` (completed/failed) → V3 `tool-result`
(with `isError` on failure)
- ACP stop reason → `{ unified, raw }`
- Unsupported V3 call options (temperature, topP/topK, seed, stopSequences,
responseFormat:json, tools, etc.) → `stream-start` warnings.
Removed:
- `src/stream.ts` (`streamAgent`, `generateAgent`, `OneShotAgentStream`,
`StreamAgentOptions`, `GenerateAgentResult`) — fully redundant with
`streamText` / `generateText` from `ai`.
Wiring:
- New deps: `@ai-sdk/provider@^3.0.8`, `@ai-sdk/provider-utils@^4.0.23`.
- Subpath export `use-local-agent/ai-sdk` and a top-level re-export.
- `vite.config.ts` adds `src/ai-sdk/index.ts` as a build entry.
- README leads with `streamText({ model: localAgent('claude') })`; the
`LocalAgent` API moves to an "Advanced" section for stateful multi-turn,
permission gating, slash commands, and session resume.
Tests:
- New `tests/ai-sdk-provider.test.ts` covers prompt conversion, provider
shape, `NoSuchModelError`, `doGenerate` (reasoning+text+tool-call
ordering, finish-reason mapping), `doStream` part ordering, and
unsupported-option warnings.
Quality gates: `pnpm typecheck` clean, `pnpm lint` 0/0 on 72 files, and
all 96 tests across 26 files pass.
Also drops the leftover `.changeset/acp-reliability-overhaul.md` per
prior cleanup.
…ten reliability section
aidenybai
force-pushed
the
consolidate-prs
branch
from
c7b1578 to
b97e46d
Compare
…treamObject/generateObject
Adds `ai@^6.0.168` as a regular dependency and re-exports its top-level
helpers from `use-local-agent`'s main entry, so callers get the
end-to-end story from a single install:
import { streamText, localAgent } from "use-local-agent";
streamText({
model: localAgent("claude", { cwd: "/path/to/project" }),
prompt: "...",
});
Bringing your own `ai` still works — importing from `"ai"` directly is
unchanged and resolves to the same versions transitively.
README updated to:
- show the new single-import pattern as the headline usage,
- demonstrate inline per-call settings on the model factory
(`localAgent("codex", { cwd, permission, inactivityTimeoutMs, onTrace })`),
- call out that users can still import from `"ai"` if they prefer.
Moves `ai` from `dependencies` to `peerDependencies` (required, not
optional) and drops the top-level re-exports of `streamText`,
`generateText`, `streamObject`, `generateObject`. Callers install `ai`
themselves and import it directly:
npm install use-local-agent ai
import { streamText } from "ai";
import { localAgent } from "use-local-agent";
Same behavior as `@ai-sdk/openai`, `@ai-sdk/anthropic`, etc. — keeps the
provider package small and avoids dragging in `ai` for callers that
might not use it transitively. README updated to use separate imports
throughout.
…mText
Adds an AI-SDK-shaped path for multi-turn ACP sessions so callers don't
have to drop down to the LocalAgent API for follow-up turns:
import { streamText } from "ai";
import { createLocalAgentSession } from "use-local-agent";
await using session = await createLocalAgentSession("codex");
await streamText({ model: session.model, prompt: "list TODOs" });
await streamText({ model: session.model, prompt: "now fix the highest one" });
await streamText({
model: session.model,
prompt: "agent client protocol",
providerOptions: { useLocalAgent: { command: "web" } },
});
Implementation:
- LocalAgentLanguageModel: replace the single-purpose `connect` factory
with a more general `acquire` factory that returns
{ agent, sessionId, historyAlreadyApplied, release }. One-shot mode
(default `localAgent("...")`) creates a fresh agent + session and
closes both on release. Session-bound mode (createLocalAgentSession)
reuses an existing agent + session and makes release a no-op.
- convertPromptToContentBlocks: new `lastUserOnly` option used by
session-bound models to skip prior conversation history (the ACP
session already holds it). Emits a `compatibility` warning describing
how many messages were skipped, plus an `unsupported` warning if the
prompt has no trailing user message at all.
- providerOptions.useLocalAgent.command: forwards a slash command to the
underlying ACP session (object form `{ name, input? }` or bare string).
- createLocalAgentSession opens one LocalAgent + one session, exposes
`.model` (LanguageModelV3 bound to the session), `.agent` and
`.sessionId` for ACP-specific surfaces, and implements
`Symbol.asyncDispose` so `await using` cleans up the subprocess.
3 new tests cover: trailing-user-only behavior with a compatibility
warning, slash command via `{ name, input }`, and the bare-string
command shorthand. Existing 96 tests updated to the new acquire shape.
Total: 99 tests across 26 files passing; pnpm typecheck and lint clean.
README: replaces the LocalAgent-based stateful example with a session +
streamText example; LocalAgent escape hatch still mentioned for
permission prompts / terminal handlers / session resume.
…vider
High severity:
- LocalAgentLanguageModel: emit a `system-message-on-bound-session`
warning when a session-bound model sees a per-turn system message.
Previously the system prompt was silently dropped because the
underlying ACP session was already opened.
- createLocalAgentSession: drop the no-op `typeof === "string" ? x : x`
ternary; the parameter is already correctly typed.
Medium severity:
- provider.ts: extract a shared `buildOneShotAcquire` so
`buildLanguageModel` and `buildFromAdapter` no longer duplicate the
same 14-line acquire factory body (DRY).
- LocalAgentProvider: tighten `modelId` parameter from
`SupportedAgentId | string` to just `SupportedAgentId`; arbitrary
strings now go through `fromAdapter(...)` so typos like
`localAgent("clauder")` fail at compile time instead of runtime.
- Add tests covering:
- system-message-on-bound-session warning
- no-trailing-user-message warning + zero-block prompt to ACP
- abort during doStream calls release exactly once
- toJsonValue -> toJsonResult: returns NonNullable<JSONValue> (matches
LanguageModelV3ToolResult.result) by JSON-roundtripping with a
String(value) fallback. Drops the over-permissive
NonNullable<unknown> contract.
Low severity / style:
- Hoist provider name / providerOptions key to `constants.ts` as
`AI_SDK_PROVIDER_NAME` and `AI_SDK_PROVIDER_OPTIONS_KEY`.
- Rename `historyAlreadyApplied` -> `isSessionBound` everywhere
(config, AcquiredSession, tests).
- LocalAgentSessionOptions now `extends LocalAgentConnectOptions`
directly instead of going through the type alias.
- Inline `.filter((p) => p.type === ...)` + `.map(...)` -> single `.map`
/ `.flatMap` to drop two `as { ... }` casts in convert-prompt.
- Trim JSDoc that restated field / function names per AGENTS.md
"default to NO comments". Comments on `release()` (idempotency
semantic) and the public `localAgent` / `createLocalAgentSession`
examples are kept.
Verified: pnpm typecheck clean, pnpm lint 0/0 on 72 files, pnpm test
102 passing across 26 files.
Renames the package and aligns the AI SDK provider surface with the
Vercel AI SDK community convention (factory drops the `Provider` suffix:
`createLocalAgentProvider` -> `createSpawnAgent`).
Public API surface:
- `localAgent` -> `spawnAgent`
- `createLocalAgentProvider` -> `createSpawnAgent`
- `createLocalAgentSession` -> `createSpawnAgentSession`
- `LocalAgent` (runtime class) -> `SpawnAgent`
- `LocalAgent{Provider,LanguageModel,Session,ConnectOptions,ClientInfo,Error,...}`
-> `SpawnAgent*`
- `providerOptions: { useLocalAgent: { command } }`
-> `providerOptions: { spawnAgent: { command } }`
- `model.provider` is now `"spawn-agent"`
Constants in src/constants.ts updated: `PACKAGE_NAME`, `PACKAGE_TITLE`,
`AI_SDK_PROVIDER_NAME`, `AI_SDK_PROVIDER_OPTIONS_KEY`.
Files renamed via `git mv` to preserve history:
- packages/use-local-agent/ -> packages/spawn-agent/
- src/local-agent.ts -> src/spawn-agent.ts
- src/ai-sdk/local-agent-language-model.ts -> spawn-agent-language-model.ts
- masterdocs/.../integration-notes-for-use-local-agent.md
-> integration-notes-for-spawn-agent.md
Includes review-pass cleanup: local `spawnAgent` variables renamed to
`agent` to avoid shadowing the public default export, error re-exports
re-sorted alphabetically, and a major-version changeset documenting the
migration.
`@agentclientprotocol/claude-agent-acp` and `@zed-industries/codex-acp` now ship as regular dependencies of `spawn-agent` instead of optional peer dependencies. Users no longer need to install them separately to drive Claude Code or Codex. `ai` remains a peer dependency. Updates README install instructions accordingly.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Single PR that consolidates the two open Cursor agent PRs (#1 ACP overhaul + #2 reliability hardening) into a coherent tree, then turns
use-local-agentinto a real Vercel AI SDK provider on top.Closes #1 and #2 (their commits are preserved in this branch's history; this PR is the merged-and-resolved superset).
What's in this PR
1. Consolidation of PR #1 + PR #2
Both PRs touched the same core files (
connect.ts,local-agent.ts,errors.ts,constants.ts,mock-agent.ts,run-command.ts). Conflicts were resolved by favoring the stricter / more correct fix in each case:run-command.ts— kept PR Reliability hardening for use-local-agent (no API breaks) #2'sisProcessAlivecheck (exitCode === null && signalCode === nullinstead of the misleadingchild.killed) plus PR ACP reliability + spec coverage overhaul (use-local-agent) #1's configurable grace period.connect.ts— combined PR ACP reliability + spec coverage overhaul (use-local-agent) #1'swaitForSpawnOrError,disposeGraceMs/onTrace/envFilteroptions, andwrapStreamWithTracewith PR Reliability hardening for use-local-agent (no API breaks) #2's no-op stream error listeners,isProcessAlive-based dispose,appendCappedstderr buffer cap,filterStdoutNoise, andAgentStdinClosedErrorwrite path.local-agent.ts— merged all option fields (pendingUpdateBufferLimit,traceContext,onAuthRequired,onTrace,envFilter,disposeGraceMs,stderrFatalPatterns); kept PR Reliability hardening for use-local-agent (no API breaks) #2'sraceInitializehelper while preserving PR ACP reliability + spec coverage overhaul (use-local-agent) #1's defensiveAgentConnectionClosedErrorfallback; reorderedprompt()to validate (with slash-command processing) before draining buffered updates / consumingpendingConfigOptions(PR Reliability hardening for use-local-agent (no API breaks) #2 fix #10); used PR Reliability hardening for use-local-agent (no API breaks) #2'sbufferSessionUpdatewithdroppedPendingUpdateCountand the configurable cap.constants.ts— keptPENDING_UPDATE_BUFFER_LIMIT_PER_SESSION = 1000(configurable); dropped the duplicateMAX_BUFFERED_UPDATES_PER_SESSION = 1024.2. Real AI SDK provider (
LanguageModelV3)use-local-agentnow plugs directly intostreamText/generateTextfrom theaipackage — no parallel one-shot helper API needed:New module
src/ai-sdk/:convert-prompt.ts—LanguageModelV3Prompt→ ACPContentBlock[]. System messages flatten intosystemPrompt; image/audio file parts become ACP image/audio blocks; unsupported file mediaTypes / assistant tool-call replays / tool-result messages emitunsupportedwarnings.local-agent-language-model.ts— implementsLanguageModelV3:doGeneratereturns orderedreasoning+text+tool-callcontent with mapped finish reason and usage.doStreamemitsstream-start(with warnings),response-metadata(carrying ACPsessionId), groupedtext-start/-delta/-endandreasoning-start/-delta/-end,tool-call,tool-result,finish, anderrorparts. Opt-in raw passthrough viaincludeRawChunks.connectfactory so it's testable through the existingconnectMockAgentharness without spawning a subprocess.provider.ts—createLocalAgentProvider()returns a callableProviderV3(call-form +.languageModel()) plus.fromAdapter()for custom ACP-speaking subprocesses; defaultlocalAgentinstance.embeddingModel/imageModelthrowNoSuchModelError. Defaultpermission: "auto-allow"sostreamTextJust Works.Mappings:
text-delta→ V3text-delta(start/end framed by message id)thinking-delta→ V3reasoning-deltatool-call→ V3tool-callwithproviderExecuted: truetool-call-update(completed/failed) → V3tool-result(withisErroron failure){ unified, raw }(end_turn→stop,max_tokens/max_turn_requests→length,refusal→content-filter,cancelled→other)stream-startwarnings.3. Removed
streamAgent/generateAgentNow redundant with
streamText/generateTextfromai. Deletedsrc/stream.tsand all related exports.LocalAgent(sessions, multi-turn, permission gating, slash commands, terminal handlers, session resume) stays as the stateful escape hatch for use cases that don't fit a single AI SDK call.4. README rewrite
Both root and package READMEs rewritten in the react-grab style, leading with
streamText({ model: localAgent('claude') })and movingLocalAgentto an "Advanced" section. Tagline: "An API for accessing any locally-installed coding agent."Public API
Added
localAgent(modelId, settings?)— default provider (call form +.languageModel()+.fromAdapter())createLocalAgentProvider(defaults?)— provider factoryLocalAgentLanguageModel(class),LocalAgentLanguageModelConfig,LocalAgentProvider,LocalAgentProviderSettingsconvertPromptToContentBlocks(utility)AgentStdinClosedError(_tag: "StdinClosed")pendingUpdateBufferLimit,disposeGraceMs,onTrace,envFilter,stderrFatalPatterns,traceContext,onAuthRequired,terminaloptions onLocalAgentConnectOptionsloadSessionStreaming,additionalDirectorieson session inputs,commandsFor/modeStateFor/configOptionsForgetters,streamAllSessions@ai-sdk/provider@^3.0.8,@ai-sdk/provider-utils@^4.0.23use-local-agent/ai-sdkRemoved
streamAgent,generateAgent,OneShotAgentStream,StreamAgentOptions,GenerateAgentResultQuality gates
pnpm typecheck— clean across all 3 packagespnpm lint— 0 errors / 0 warnings on 72 filespnpm test— 96 unit tests across 26 files passing (10 new for the AI SDK provider; PR ACP reliability + spec coverage overhaul (use-local-agent) #1 contributed ~28; PR Reliability hardening for use-local-agent (no API breaks) #2 contributed 14)pnpm --filter @use-local-agent/playground exec playwright test --project=node— 8/8 e2e tests against the realecho-agentsubprocessStats