plat 134

← Pulse platform issue index

PLAT-134 — direct product chat still entered through the legacy multi-agent orchestrator and constructed capabilities the product explicitly forbade

Coordination	Value
Assigned agent	Codex
Ticket state	partially_implemented — direct chat runtime, durable product-chat rotation, and hybrid tool contract are covered; compatibility naming and unreachable legacy implementation remain to be extracted/deleted
Last synchronized	2026-08-22

• Priority: P1 — ordinary and product chat are simple conversational sessions, but their runtime was assembled as a generic multi-agent orchestrator. That made delegation, schedule ownership, tier selection and background-agent lifecycle appear to be part of every chat even when a product profile explicitly disabled them.
• Owner: direct chat request/runtime construction (cmd/server/server.go, cmd/server/agent_profile_routes.go), chat UI model/config projection (frontend/src/components/ChatArea.tsx, ChatInput.tsx)

Intended product contract

Video Studio and Dominion need one reusable product-chat core:

1. a server-owned system prompt;
2. attached skills;
3. an explicit allowlisted tool surface;
4. a server-owned model and credential binding;
5. a server-owned project/workspace binding; and
6. a durable conversation key that resumes the same conversation.

They do not use chat-level delegation. Finance keeps its dashboard UI but does not expose chat for now. Workflow Builder's run_in_background and call_sub_agent are a separate automation-authoring capability and are not in scope for removal.

Root cause

The historical name agent_mode="multi-agent" became both a UI category and a runtime policy. handleQuery treated every ordinary/product chat as an orchestrator and constructed session-specific wrappers for:

• delegate, query_agent, terminate_agent, and list_agents;
• chat-owned schedule create/update/delete/trigger tools;
• async delegated-task execution and its background registry;
• delegation-tier model selection and per-tab reasoning controls; and
• a large delegation guidance surface.

Product profiles eventually filtered forbidden tools out, but only after the server had constructed and registered the broad machinery. The public browser request could also carry a large QueryRequest, even though prompt, model, tools, skills, permissions and workspace are product-owned authority.

This is the wrong dependency direction: a product declared a small capability set, then the generic chat runtime built a larger orchestrator and tried to subtract capabilities from it.

Refactor shipped

The live runtime now follows the direct-chat contract:

• Product chat accepts only message and optional conversation_key; the server resolves the profile and authors the internal request.
• Ordinary chat uses GetAgentWorksChatInstructionsWithUser, which tells the agent to work directly with attached skills/tools and not create sub-agents or own schedules.
• handleQuery no longer calls CreateDelegationTools, CreateDelegationToolExecutors, delegated-task wrappers, or chat schedule tool factories.
• The chat-specific schedule tool file and pre-registration allowlist were removed. Scheduler default helpers that are genuinely used by workflow scheduling were retained in a small neutral file.
• Direct chat no longer lets delegation tiers silently replace the selected primary model. Product profiles retain server-owned model authority.
• Video Studio is currently pinned to its server-owned Claude Code model; its provider options and frontend provider dropdown were removed so the UI cannot claim a choice the minimal product-chat request no longer sends.
• Video Studio also declares credential_scope: global; its per-project coding agent token button/dialog were removed, and the runtime no longer layers old project token overrides over the shared Claude login.
• The frontend no longer eagerly loads delegation tiers for chat and no longer carries the unused delegated-task reasoning popup.
• The delegation reference entry/template was removed from the chat guidance surface.
• Product mode now installs one shared ProductChatSurface automatically. Existing agent_error, conversation_error, failed completion, and cancel carriers normalize into a single product failure state; retryable failures can replay the last human turn and raw provider text remains collapsed under technical details. New products receive this behavior through inputVariant="product" without product-specific event parsing.

The focused backend contract asserts that the direct-chat path contains no delegation or schedule registration. go build ./..., the focused backend tests, TypeScript, and all 632 frontend tests pass.

Boundary deliberately preserved

Workflow Builder and scheduled workflow execution still need background-agent tracking for run_in_background, call_sub_agent, message-sequence children, Pulse reviewers/Fixer, execution-tree liveness, and auto-notifications. That is workflow orchestration, not product chat delegation. Deleting the shared background registry would break those features and is explicitly outside this ticket's direct-chat removal.

The bot connector also still owns a legacy model-profile store currently named delegation_tier_config. Direct/product chat no longer loads it; it must be renamed or migrated independently before the store/API can be deleted.

2026-08-22 — product chat New Chat and browser tool boundary

Two product-facing details were found while exercising Video Studio. They are shared ProductChatSurface/runtime behaviour, not a Video Studio-only UI patch.

New Chat must rotate the server-owned durable conversation

The old New Chat action called resetTabChat, which cleared the tab and minted a browser-local random session id. A keyed product project, however, resolves its authoritative session from product.json, and the product-conversation registry returns that same canonical session on the next query. The result was a misleading blank screen followed by the old durable transcript returning as soon as the user said “hi”.

The client now calls POST /api/agent-profiles/{id}/conversation/new before it clears a product tab. The server mints a fresh conversation and session, replaces the registry entry, and, for a keyed project, updates that project's product.json session_id. The production/artifact workspace is retained and the prior transcript is not deleted; it simply ceases to be the active chat. If rotation fails, the UI keeps the old conversation rather than pretending a new one exists. Singleton products use the same endpoint but only rotate their registry binding.

Browser product uses the guarded MCP surface

Video Studio's runtime.agent_tools.mode: mcp_only exposes the guarded MCP bridge for all workspace and product work, including execute_shell_command for product HTTP APIs. This retains the session's project-scoped Folder Guard and structured tool progress/results for the web client. It does not claim that native CLI tools have the same guard: they are not enabled for this browser product.

Native terminal questions cannot be used by a browser product

Testing native tools exposed a second boundary: Claude Code's native default tool surface can open an in-terminal multiple-choice question. The web client cannot render or answer that terminal UI, so the provider waits forever and ProductChat continues to show Working. This is not an LLM latency issue and it is not a Folder Guard decision.

Browser products must therefore default to mcp_only until native interactive questions are either bridged into ProductChat or denied by a provider-specific, tested --disallowed-tools configuration. Prompt guidance alone is insufficient.

Separately, mcpagent must preserve provider-projected .claude/skills and its managed CLAUDE.md across per-turn Agent.Close() calls for persistent native sessions. A web turn creates a short-lived Agent, but its tmux CLI session is intentionally retained; deleting the files at the former boundary left the still-live session with no inspectable project setup.

Dominion AI uses the same transport: auto + agent_tools.mode: mcp_only profile shape as Video Studio. The transport keeps the Claude session and its project projection warm; MCP-only keeps terminal-native tools out of the web chat path.

Remaining compatibility cleanup

The behavior is simplified, but the source tree is not yet fully reduced:

1. "multi-agent" remains the persisted/internal compatibility identifier for ordinary chat. It should become a neutral chat runtime kind through a versioned state migration, not a blind string replacement.
2. Product chat still forwards a server-authored QueryRequest into handleQuery. Extracting a shared direct-turn runner would remove the broad request type from the product path without another frontend migration.
3. Unreachable chat-delegation factories/executors remain in virtual-tools/delegation_tools.go, delegation.go, and background_agents.go. Those files co-locate workflow notifier/lifecycle types that are still live. Extract the workflow-owned types first, then delete the dead chat-only functions and tests.
4. Rename/migrate the bot connector's tier configuration so chat terminology no longer leaks into unrelated connector model selection.

Acceptance

• A normal chat and a product-profile chat register zero chat-level delegation or schedule-management tools.
• Product clients can send only message plus an optional durable conversation_key; model, tools, skills, credentials and workspace cannot be overridden by the browser.
• Video Studio and Dominion resume the same product conversation and remain constrained to their declared profile capabilities.
• New Chat rotates the durable product conversation server-side; a subsequent product query cannot restore the transcript it just replaced.
• Browser-product documentation and prompts require the guarded MCP surface and explain why provider-native terminal tools are unavailable there.
• Chat startup performs no delegation-tier API request and has no delegated-task reasoning control.
• Every product chat stops its working state and renders a normalized failure for both explicit error events and terminal completion text such as all LLMs failed; retry does not require product-specific code.
• Workflow Builder run_in_background / call_sub_agent, scheduled workflow children and Pulse background reviewers continue working unchanged.
• A later source cleanup can delete the unreachable chat delegation implementation without changing runtime behavior.

Verification status

Code-level and automated verification are complete. A rebuilt server still needs one live direct-chat turn each for ordinary Chat, Video Studio and Dominion, plus one Workflow Builder background-agent run proving the preserved boundary.