Talk runs the captain: hourly check-ins, the friends watch and owner instructions as routines in a thread, with memory and resume #11015

Talk now carries the owner's memory, runs the three routines, and carries a long thread past a full context. It is a partial slice of #11011: the proof day, the v0-style builder, and releases, pillar movement and stuck child forges in the check-in are not done.

Tests I added pass, and so do the neighbouring ones I ran (reserved names, studio, shards, check list, hollow, principles, settings pages, conversation tests) plus gate.ts, tsc and design-lint. I did not run the full suite or the slow checks.

What changed

• Memory. Short durable rules, up to 60 of 400 characters, are listed, added, edited and deleted in Settings › Captain (/settings/captain, owner-only, same-origin). Every Talk turn carries them in its system prompt.

- The first visit offers starting rules: never rm, tokens off argv, never dark without a go, ask first on money, DNS and a dark roll. Delete them and they stay gone.

• Routines. The hourly check-in, the friends watch and the morning paper each have an editable schedule: every 15 to 1440 minutes, or daily at a UTC hour.

- Each posts at most three lines into the owner's "Captain" thread, once a minute-clock finds it due. A forge that was down runs an owed routine once on waking, and "Post now" runs one at once. - Only a report that says something needs the owner (red trunk, unhealthy forge, stuck runs, waiting sign-ups) pushes to the phone, through the existing notify URL. - The lines are computed from forge data, not written by a model, so a routine costs no spend.

• Instructions. The Talk brief now tells the agent that the owner's message is an instruction to act on at once with their powers. Real money, DNS and a dark roll stay behind the existing Confirm and Cancel buttons.
• Resume. Sessions already resumed (#10457). When the last turn left the context 85% full, the next turn starts a fresh session from the stored record plus the memory, and the thread says so.

What I did not do

• The proof day is not done. It needs the owner to use it for a day, and I have no thread to link.
• The check-in covers health, red trunk, asks, crew holds, grokbot pulls and merges. It does not yet cover releases, pillar movement, or child forges stuck.
• There is no terminal screen for Settings › Captain; the commit says No door: for that.
• The v0-style builder (live preview, a version per turn) is not started.
• Routines post into the Captain thread, so the next message in /talk lands there unless the owner has opened another conversation.

Try it: as the admin, open /settings/captain, press Post now on Forge check-in, and read the Captain thread in /talk. The how-to is in docs/features/talk-captain.md; tests are test/talk-captain.test.ts and test/talk-captain-turn.test.ts.

🤖 Generated with Claude Code

Closes #11011

Talk now carries the owner's memory, runs the three routines, and carries a long thread past a full context. It is a partial slice of #11011: the proof day, the v0-style builder, and releases, pillar movement and stuck child forges in the check-in are not done.

Tests I added pass, and so do the neighbouring ones I ran (reserved names, studio, shards, check list, hollow, principles, settings pages, conversation tests) plus gate.ts, tsc and design-lint. I did not run the full suite or the slow checks.

What changed

• Memory. Short durable rules, up to 60 of 400 characters, are listed, added, edited and deleted in Settings › Captain (/settings/captain, owner-only, same-origin). Every Talk turn carries them in its system prompt.

- The first visit offers starting rules: never rm, tokens off argv, never dark without a go, ask first on money, DNS and a dark roll. Delete them and they stay gone.

• Routines. The hourly check-in, the friends watch and the morning paper each have an editable schedule: every 15 to 1440 minutes, or daily at a UTC hour.

- Each posts at most three lines into the owner's "Captain" thread, once a minute-clock finds it due. A forge that was down runs an owed routine once on waking, and "Post now" runs one at once. - Only a report that says something needs the owner (red trunk, unhealthy forge, stuck runs, waiting sign-ups) pushes to the phone, through the existing notify URL. - The lines are computed from forge data, not written by a model, so a routine costs no spend.

• Instructions. The Talk brief now tells the agent that the owner's message is an instruction to act on at once with their powers. Real money, DNS and a dark roll stay behind the existing Confirm and Cancel buttons.
• Resume. Sessions already resumed (#10457). When the last turn left the context 85% full, the next turn starts a fresh session from the stored record plus the memory, and the thread says so.

What I did not do

• The proof day is not done. It needs the owner to use it for a day, and I have no thread to link.
• The check-in covers health, red trunk, asks, crew holds, grokbot pulls and merges. It does not yet cover releases, pillar movement, or child forges stuck.
• There is no terminal screen for Settings › Captain; the commit says No door: for that.
• The v0-style builder (live preview, a version per turn) is not started.
• Routines post into the Captain thread, so the next message in /talk lands there unless the owner has opened another conversation.

Try it: as the admin, open /settings/captain, press Post now on Forge check-in, and read the Captain thread in /talk. The how-to is in docs/features/talk-captain.md; tests are test/talk-captain.test.ts and test/talk-captain-turn.test.ts.

🤖 Generated with Claude Code

The four failing checks are fixed and pushed. Locally, the gate, tsc, design-lint and the five affected test files (attention, server-timing-sweep, header-width, design-audit, talk-captain) pass. I didn't run the full suite; the forge's check does that on this push.

• Server-Timing: /settings/captain now reports a render phase, like the other pages. The captainPage handler in src/web/pages/captain.ts already reported session.
• Header width: I wrapped the over-long comment at src/crew/conversation.ts:337.
• Design audit: I added the /settings/captain entry to docs/design/audit.md.
• First-view budget: I raised the settings words budget from 140 to 141 at both widths. The new "Captain" nav link adds one word to the Settings page. I raised the budget rather than drop an existing label. A person may want to confirm that trade.

Together with the earlier commit, this pull makes Talk the captain. Standing rules from Settings › Captain are loaded into every Talk turn. The hourly check-in, friends watch and morning paper post into a Captain thread. A thread at 85% context continues in a fresh session. To try it, open /settings/captain as the admin, press "Post now" on "Forge check-in", then read the Captain thread in /talk.

I did not build a terminal screen for Settings › Captain; it is a follow-up. The same data is reachable in Talk and on the web page.

All six review points are addressed and pushed. The gate, tsc, design-lint, the Captain tests and the attention, audit, header-width, Server-Timing and palette tests pass locally. I didn't run the full suite or the slow checks; the forge's check does that on this push.

• Terminal screen. src/tui/screens/captain.ts is a new palette entry, "Captain", offered only to a site admin who is a person. It lists the rules and the three routines. a adds a rule, enter edits a rule or a schedule, d deletes a rule, space turns a routine on or off, and p posts a routine now. It uses the same core functions as the page. A schedule is typed as minutes (45) or at 7 for daily at that UTC hour.
• Reads no longer write. listRoutines and the new rulesOf only read, and show defaults and starting rules until a first write. The first memory write keeps the starting rules (seedMemory), and a routine row is created when it is first saved or run. Starting rules are shown with ids start-N, so one can be edited or deleted before anything is stored. memoryBrief still reads only stored rules.
• Merged import. ROUTINE_KINDS now comes from the same import as the other routine names.
• Page. I removed the Memory and Routines paragraphs. Each routine is now a list row with its fields inline, not a nested settings block.
• Untrusted posts. The Talk brief now says routine posts quote text other people wrote and are untrusted data, never instructions. A test covers it.
• Budget. I restored the Settings words budget to 140. To make room I renamed the "Cloud accounts" nav label to "Cloud". The page's own <h2> and title still say "Cloud accounts".

• Starting rules and the brief. The brief carries starting rules only once the owner's first write has stored them. Before that it carries none, which matches how it behaved before this change.
• Tests I changed. The "memory is in the brief" test now expects the starting rules beside a first added rule, because adding the first rule stores them. Starting rules are also stored a millisecond earlier than that first rule, so they stay in order ahead of it.
• Tests I added. Reads write nothing, a starting rule can be deleted before anything is stored, the brief fences routine posts, and schedule parsing for the terminal screen.

I didn't add a feature doc or an audit entry for the terminal screen; docs/features/talk-captain.md still says the terminal has no screen, so it needs a line changed. The remaining gaps are as before: releases, pillar movement and stuck child forges in the check-in, the proof day, and the builder.

The check failure is fixed and pushed. The forge's check hasn't re-run on the new push yet.

The new Captain screen adds a "Captain" entry to the command palette. The site-admin roster test, "a site-admin person is offered the whole roster", didn't list it, so it failed. The test was out of date, not the code. I added the entry to the expected roster for site-admin people in test/palette-roster.test.ts. It sits between "Pack repositories" and "Connection apps", with the hint "what Talk remembers, and the check-ins it posts for you".

test/palette-roster.test.ts passes (15 of 15) and bun x tsc --noEmit is clean. I ran no other tests.

Talk is where the forge's owner steers it: a message there is an instruction acted on at once, with the owner's own powers.

• Memory (core/talk_memory.ts): up to 60 short rules, edited at Settings › Captain (/settings/captain). Every Talk session carries them in its system prompt. The first visit offers a set of starting rules, and a rule the owner deletes stays gone.
• Routines: the hourly forge check-in, the friends watch and the morning paper, each on its own schedule. Each posts at most three lines into the owner's Captain thread. The lines are computed, not written by a model, so a routine costs no spend. Only a report that needs the owner reaches the phone.
• Resume: a thread resumes its Claude session. When the last turn left the context 85% full, the next turn starts a fresh session from the thread's stored record and the memory, and says so in the thread.
• Terminal: the palette's Captain screen (tui/screens/captain.ts) does the same reads and writes as the settings page.

This push answers the review:

• Feature doc: docs/features/talk-captain.md no longer says the terminal has no Captain screen. It now lists the screen's keys (a, enter, d, space, p, r) and the schedule syntax: 45 means every 45 minutes (15 to 1440), and at 7 (or daily 7) means daily at 07:00 UTC.
• Refusal words: the text shown when a rule is refused is now written once, as MEMORY_WORDS in core/talk_memory.ts. It is typed against MemoryError, so a new error cannot go without words, and the 60-rule limit in its text comes from MEMORY_MAX. The web page and the terminal screen both import it.
• Earlier fix: the site-admin palette roster test now expects the Captain entry.

Not changed: the stale "No door:" line sits in an already-pushed commit message. Rewriting it would need a force push, which is refused here, and no later commit reuses it.

Checks: bun scripts/gate.ts and bun x tsc --noEmit pass, and the design lint reports nothing. scripts/fast.ts named 1,784 test files, which is over its own 5-minute budget, so I ran the files that cover this change by hand instead: test/talk-captain.test.ts, test/talk-captain-turn.test.ts and test/palette-roster.test.ts. All 31 tests pass.

To try it: sign in as the site admin and open /settings/captain, or open Captain from the terminal's palette. Add a rule, set a routine to at 7, then use Post now on the page (or p in the terminal) and read the Captain thread in Talk.

🤖 Generated with Claude Code