[AbirAbbas]

Summary

The Workstreams plan, WP-10 (engine side): a workstream's own agent gets a short workstream section in its system prompt and six tools that reach the workstream store through a narrow door the app injects. Every tool enforces the AI rules in code; a plain agent's belt and page are unchanged.

• Door. session.WorkstreamDoor (internal/session/workstream_contract.go) is set by the app only on the conversation whose created-by is workstream (one line in engine.go beside cfg.Floor, one registration line in app.go). It resolves its workstream at each call by agent id (workstream.Store.ByAgent), because the agent is opened a moment before the workstream row exists. Task nodes, errands, quick tasks and question clarifications never carry it.
• Prompt. Names the workstream (name, why) and says: it proposes and the person keeps; it writes only its own pages; it never invents numbers (live blocks instead); it starts nothing.
• Tools (internal/session/workstream_tools.go):
  • workstream_read {what: workstream|pages|page|records|plan|goaheads, page_id?, kind?, state?}: wire-shaped JSON; numbers come from records and Factory reads, absent means unknown.
  • propose_decision / propose_question {text, source?}: a proposed record by agent; no path keeps it. Source defaults to the agent's chat; explicit sources are a member job, a page of this workstream or an http(s) URL.
  • propose_jobs {jobs: [{title, repo?, text?}], wave?, wave_title?, milestone_id?}: appends drafts only; existing waves, items and auto_start untouched; retries a version_conflict once, then reports. Never touches a Factory door.
  • propose_goahead {scope, factory?, job?, wave?, milestone_id?, cap_usd, ends_on, why}: the contract has no proposed go-ahead, so this is a proposed decision whose text begins Proposed go-ahead (not granted): ; it grants nothing.
  • write_agent_page {page_id?, title?, parent_id?, markdown, updated_from, source?}: creates or updates agent-owned pages only; a page the person owns is refused inside the store transaction and nothing changes; writes carry updated_by: agent, the store's time and updated_from.
• Unknown argument fields (kept, state, cost_usd, auto_start, …) are refused before the door is touched.
• Store. savePage refuses an agent write to a non-agent page in its transaction (ErrNotAgentPage); pagedoc.FromMarkdown imports markdown (live-block markers included) and the contract example round-trips byte for byte.
• Spend. spend_usd already counts the agent conversation's recorded spend (spawned tasks fold into it). New: a workstream agent's settled turn sends workstream.changed (workstream part) so the spend is read again.

Validation contract

item	tests
(a) a workstream agent gets the tools and prompt; a plain agent does not	internal/session/workstream_tools_test.go, cmd/codeaf/appworkstreams_agent_test.go
(b) proposals are proposed records / drafts, never kept	workstream_tools_test.go, appworkstreams_agent_test.go
(c) propose_jobs never calls the Factory	internal/session/workstream_factory_test.go, appworkstreams_agent_test.go
(d) write_agent_page refuses a person's page and changes nothing (including the read/save race)	internal/workstream/workstream_agent_test.go, workstream_tools_test.go
(e) agent pages carry the AI mark, time and source; history too	internal/workstream/workstream_agent_test.go, appworkstreams_agent_test.go
(f) plan conflict retried once, then reported	workstream_tools_test.go
(g) inputs cannot set counts, costs or states	internal/session/workstream_validation_test.go, workstream_tools_test.go
(h) manual and laws	TestTheManualMentionsTheWorkstreamAgentsTools, manual section in collections.md

Review and fix rounds

Two independent reviews. Fixed: the prompt's workstream brief is a non-blocking snapshot (no store read under the session lock); an agent's record and page writes check done inside their own transaction (the wire's unattributed record.propose keeps its old behaviour); imported list items always start with a paragraph; malformed inline markdown is parsed in linear time; repeated JSON keys are refused at any depth; the tools follow the session's argument laws (one decoder, marked refusals, no counting retry loop; cap_usd excused as dollars in its own file).

CI

PR gate: build, vet, gofmt, change entry and manual gates green. The remaining reds are the integration branch's baseline: internal/session complexity ledgers, prefix budgets, TestProgramHoldClassifiesEveryRegisteredChatTool, TestALandStartsNoAutomaticMaintenance; cmd/codeaf TestAppChatRowsR2SavedTurnTime, TestAppChatStartupR3QuietRelease, TestRecipeDoorWritesTheWorktreeWhenOnlyTheCheckoutIsHeld.

Contract gaps

• No proposed state or record kind for go-aheads: proposals are decision records with a fixed text prefix; the person grants the real go-ahead.
• No turn-level message source is reachable from inside the engine tool; the default source is the agent's conversation.

No wire shapes or handlers changed. Suggestions on the person's pages, autonomy-driven small changes and the watching agent are later work.

🤖 Generated with Claude Code