Skip to content

chore(tars): add a coordinator trial kit and outcome comparison - #1022

Merged
astyfx merged 10 commits into
mainfrom
chore/tars-coordinator-trial
Nov 19, 2026
Merged

astyfx merged 10 commits into
mainfrom
chore/tars-coordinator-trial

Conversation

@astyfx

@astyfx astyfx commented Nov 19, 2026

Copy link
Copy Markdown
Member

Summary

Stacked on the PR-base fix (base fix/tars-pr-base-and-agent-start). Trial kit for synthesis #13, with no product code: two reviews concluded Tars\x27s planner should become an always-on Task that uses tools; the synthesis approved a one-week side-by-side trial, not a replacement.

  • docs/developer/tars-coordinator-trial.md: the five capabilities mapped to existing tools (no gap blocks the trial), a coordinator agent definition the owner can create as a saved Agent, the provider/worktree setup under which an independent review really runs (an Auto coordinator; a Fixed one cannot call the other provider), the one-week protocol and the decision rule.
  • scripts/measure-coordinator-outcomes.ts: the same outcome numbers as the Tars script, side by side, for a date range in local time. PR evidence is scoped to the delegation that produced it; failed, blocked and taken-over work is never counted as unattended.
  • The Tars script now opens a copied database.

Review and verification

Written and reviewed by different providers: each branch was implemented by one model, reviewed independently by a model of the other provider, and fixed until the reviewer's blocking findings were resolved. Lead checks on the rebased head: bun run typecheck, bun run lint, bun run check:i18n, bun run check:doc-paths, bun run check:max-lines-ratchet, bun run build:pages, git diff --check, and every changed unit test file run one per process: all exit 0. New regression tests fail on the previous code and pass now. Not verified: the installed Electron app, live providers, real macOS sleep, real GitHub calls.

Part of the design-review follow-up (synthesis priorities 1–14).

🤖 Generated with Claude Code

@astyfx
astyfx requested a review from a team as a code owner November 19, 2026 07:52
@astyfx
astyfx requested review from heath-s and png-1105 and removed request for a team November 19, 2026 07:52
@upwind-code-us

upwind-code-us Bot commented Nov 19, 2026 •

Copy link
Copy Markdown

Upwind Code Scan

Upwind code scan | ⏭️ Skipped

No scans ran: this PR changes nothing Upwind scans.

Last updated: 4a89ca2 · 2026-11-19 11:06 UTC

@astyfx
astyfx force-pushed the chore/tars-coordinator-trial branch from 54a39ac to 15506af Compare November 19, 2026 08:19

@heath-s heath-s left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Approved

Base automatically changed from fix/tars-pr-base-and-agent-start to main November 19, 2026 11:03
astyfx and others added 10 commits November 19, 2026 20:04
A copy of the app's WAL-mode database could not be opened read only, so the
documented way to measure a copy failed. Open such a copy as immutable, add an
exclusive --until bound, and report failed work separately.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Read the writing delegations a coordinator agent's tasks started from the run
ledger and their child runs, map them onto the Tars outcome definitions, and
print both summaries for the same range.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Map the five coordinator capabilities to existing Local MCP and shell tools
with the limits that shape the setup, give a saved Agent definition to create
in a minute, and write the one-week protocol, decision rule and limits.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ants

--since/--until were compared to stored UTC times as strings, so a bare
date closed at UTC midnight (09:00 in Seoul) and an offset was ignored.
A bare date is now that day's midnight on this machine's clock, a time
with an offset is that instant, and stored times are compared as instants.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… request

Any GitHub pull request link in a child's report became the unit's
publishedAt, and the shared outcome let that win over a failed state, so
a failed unit whose blocked report quoted some PR counted as a pull
request and as unattended. A unit now reaches a pull request only when
the child gave the link in a completed report or its answer, the
coordinator's completed report lists it as verified with the unit's
branch as head, and it is in the repository the work ran in; never for a
failed unit or one whose latest report is blocked. The shared outcome
also keeps a failed row failed. Tars's numbers are unchanged.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A task that ever ran as the coordinator had all its delegations in the
period counted, including those made while it ran as another agent. Each
delegation now belongs to the agent its task ran as when it was made: the
task's newest assignment by then, unless that one had ended.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A Fixed Claude coordinator cannot start a Codex reviewer (a pinned
model keeps every helper on the run's provider), and a default read-only
review runs in the coordinator's workspace, not at the owner's commit,
so the Reviewer stops on the mismatch. The coordinator now runs on Auto
preferring Claude, names the reviewer's provider, and reviews in a new
worktree started at the commit. The doc also states the pull request,
attribution and date-range rules the measuring script now applies, and
the product change a pinned coordinator would need.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…never for taken-over work

The coordinator numbers looked for verified pull requests across every report
of the parent task, so a later unit on a reused branch borrowed an earlier
report, and a follow-up delegation in the same child task was counted with the
first one's pull request. Work a person took over was a pull request reached
without them.

A unit now counts only its own records: the events of a shared agent run go
to the delegation that was the task's newest when they happened, a pull
request belongs to the delegation that first gave its link, and the
coordinator's report verifies it only when written after that link and under
the assignment that delegated the unit. In the shared definition, taken-over
work is taken over whatever was recorded for it, and a takeover that was
handed back or work the person finished counts as a human decision.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
A pull request was keyed by its link as written, so a follow-up that quoted
the same pull request with other letter case in owner/repo became a second
owner and, once the coordinator listed both spellings, a second success. An
http link was not read at all, which moved the pull request to a later
delegation.

Read every link as one identity (host, owner and repository lower-cased, and
the number) for the unit's claims, the first-to-give-it rule and the
coordinator's verified list. measure-tars-outcomes.ts compares no links.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
@astyfx
astyfx force-pushed the chore/tars-coordinator-trial branch from 15506af to 4a89ca2 Compare November 19, 2026 11:06
@astyfx
astyfx merged commit 7a986ef into main Nov 19, 2026
14 checks passed
@astyfx
astyfx deleted the chore/tars-coordinator-trial branch November 19, 2026 11:21
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.

2 participants