Skip to content

Rename chief-of-staff to coordinator and keep programs moving - #183

Merged
sreejithraman merged 5 commits into
feat/coordinator-record-commandsfrom
refactor/coordinator-keep-moving
Nov 15, 2026
Merged

sreejithraman merged 5 commits into
feat/coordinator-record-commandsfrom
refactor/coordinator-keep-moving

Conversation

@sreejithraman

@sreejithraman sreejithraman commented Nov 15, 2026 •

Copy link
Copy Markdown
Owner

Summary

Chief of Staff is now Coordinator, and it keeps programs moving without being pushed. This is layer 2 of the coordinator stack, on top of #182 (feat/coordinator-record-commands), which added the record commands and status. The page (PR 3) and the Inbox prototype (PR 4) will stack above this one.

A review of six real chief programs found these problems: chiefs that idled until pushed, autonomy granted one sentence at a time, idle owners nobody noticed, lost or bundled asks, chiefs running outdated rules, and times shown in UTC. This layer addresses each one.

Light by default. For one request across threads, the coordinator answers it, reuses a running owner, or starts one from an Orchestration brief. Nothing else is needed: no record, page, Inbox, grant, or heartbeat. A request that touches a program's work goes to that program's coordinator. The program layer applies only to two or more outcomes, or when the user asks for one.

Keep moving. The status verdict decides how each turn ends:

runnable  -> never end the turn; act on each line
working   -> end the turn; owner messages and the heartbeat wake you
waiting   -> end with the one `say:` line status prints, then stay quiet
done      -> delete the heartbeat, hand off, clean up
every turn ends with `status --turn-end`

A decision whose recommended option is reversible is built on provisionally instead of stopping work. The coordinator re-reads its SKILL.md on resume and on each heartbeat wake. Times shown to the user are in their local zone.

Grant. Each program has one grant, set at the start and changed only on the user's word. It has five parts: start (grant.start), land (the existing landingAuthority), resolve (grant.resolve plus the existing approvals), what stays with the user (grant.user), and the heartbeat (grant.heartbeat). Every kickoff carries it. Owners decide reversible method choices themselves. Only scope, irreversible, and taste calls come up.

Heartbeat. In T3, a schedule_task interval trigger (every 10 minutes, bindToCurrentThread: true) runs record.py status --heartbeat and follows its first line. A tick commits nothing. It stays quiet while the program is waiting or working with unchanged facts. It wakes the coordinator on:

  • runnable work, such as an owner that went idle while it owes work;
  • a changed digest;
  • a forge note it hasn't seen yet: a base closing under provisional work, a failed read, or a row that goes stale because reads keep failing.

It says done only when the committed record is done too. What it last saw lives in the program's git directory, so the record stays clean.

Inbox. "Needs you" is now the Inbox, both on the page and in prose. The record key stays decisions. Each item asks one question, with context, two or more options with their effects, one recommendation, optional visuals, what it blocks, what continues meanwhile, and when it was asked. Program asks never use a blocking question tool unless the user is in a live back-and-forth.

Also:

  • Owner thread titles start with the outcome ID.
  • IDs are never reused.
  • Orchestration gains a Coordinator model row (Claude Opus 5.5 High, with Codex GPT-6.1 Sol High as fallback).
  • Orchestration now notices idle children and routes decision requests to a coordinator's Inbox when one exists.
  • Implement's release gate applies only in a program.

Compatibility.

  • The chief record key and the Sites wire names (chief-version.json, chief-refresh.js) stay, so running records validate and open Sites keep refreshing.
  • grant is optional.
  • Every reference to the old name is updated: other skills, README, SOURCES, notices, and scripts/check.py. Host sync finds skill folders by name, so it removes the old link and adds the new one.

Running programs

After this stack merges and the main checkout moves, skills/chief-of-staff disappears from both hosts. Send each running chief thread this message:

The chief-of-staff skill is now coordinator. Re-read skills/coordinator/SKILL.md, then run python3 scripts/record.py status --state <your program.json> from that folder, and follow its migration steps: write the grant from what I already granted, set up the heartbeat, and republish the page.

Evidence

  • python3 scripts/check.py on the head passes: 393 tests plus the skill and link checker.
  • New test_record.py cases cover:
    • the say: line;
    • heartbeat quiet, wake, and done;
    • a passing turn-end as the baseline;
    • a running↔waiting owner staying quiet;
    • a base closing under provisional work waking once;
    • a last merge found by the heartbeat waking instead of done;
    • a failed read waking once, then again when its row goes stale;
    • dependency-only waits;
    • proposed work at its stop letting the program finish.
  • New validator rules cover grant.
  • End to end, I ran the real script against a scratch program in a temp git repo on the final head:
    • first tick wakes, the next is quiet;
    • --turn-end fails until the page is published, then passes and prints say: Waiting on Inbox A-D1 (8h13m). Nothing is runnable.;
    • the row going stale wakes once, then the heartbeat is quiet again;
    • the record stays clean;
    • the page shows Inbox and Coordinator.
  • Trigger check: an agent given only the 62 skill descriptions routed 8 of 9 coordinator requests to coordinator. The ninth, a five-feature spec, went to implement, which offers Coordinator for that case. All 6 near misses went elsewhere.
  • Review: review-fix-loop ran wave 1 (Primary, Ponytail, Thermo, plus Gemini 3.8 Flash), then Adversarial and a repair pass (Primary, Thermo, Ponytail), then a final Primary repair and integrated pass. Every accepted finding is fixed with a regression test where it is code.
  • Not tested:
    • a live T3 scheduled tick in a real thread;
    • a real running program's migration;
    • firing through a real host skill listing.

Merge Danger

Mostly a two-way door, contained to skills and their scripts. A revert restores chief-of-staff. Programs migrated in the meantime keep validating under the old validator, except for a grant key, which the old validator rejects. Drop grant from program.json before rolling back, and delete each program's heartbeat scheduled task. Running chief threads need the message above once the merge goes live. Until then, they keep working on the old path.

🤖 Generated with Claude Code

sreejithraman and others added 5 commits November 15, 2026 12:45
The skill is now Coordinator, light by default: one request across
threads is answered, routed to a running owner, or given a new one, with
no record. Programs add a grant (start, land, resolve, what stays with
the user, heartbeat), a T3 heartbeat that runs status --heartbeat, and
an async Inbox in place of "Needs you".

record.py gains status --heartbeat, which observes without committing
and stays quiet while the digest holds, and a say: line for waiting
turns. The renderer accepts grant and labels the tab Inbox. The chief
record key stays, so running programs keep validating.

Orchestration gets a Coordinator model row, idle-child supervision, and
Inbox routing for decision requests.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A tick now wakes on a forge note it has not seen, such as a base closing
under provisional work, and says done only when the committed record is
done. A live owner's sub-status stays out of the digest, a dependency
names itself in the waiting line, and the saved prompt only runs the
command. Light starts use an Orchestration brief, not the program kickoff.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A failed read wakes the heartbeat once and again when its row goes
stale. Proposed work at its stop no longer holds a finished program in
waiting. Implement's release gate and the one-coordinator check now
apply to program work, so light owners land on their own finish line.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@sreejithraman
sreejithraman added this pull request to stack #184 November 15, 2026 18:22
@sreejithraman sreejithraman changed the title refactor/coordinator keep moving Rename chief-of-staff to coordinator and keep programs moving Nov 15, 2026
@sreejithraman
sreejithraman merged commit 247c6ee into main Nov 15, 2026
3 checks passed
@sreejithraman
sreejithraman deleted the refactor/coordinator-keep-moving branch November 15, 2026 23:47
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant