Live proving Run on the eligible labeled task

[JacobStephens2]

Parent

Part of #78.

What to build

Live proof instead of suite-only proof: a real unattended Run dispatched on the eligible labeled career task, verified end to end — merge-clean branch tip, Proposal body and comments as specified, briefing visible in the window. Dispatched only after the instance-side preconditions (Target enrollment, operator allowlist, retry budget) are confirmed.

Acceptance criteria

• The Run is dispatched after the instance-side preconditions are confirmed.
• Its branch tip is merge-clean with evidence in history and comments.
• Its briefing is visible on the live card and in history.

Blocked by

• Run ends merge-clean: scaffolding removed before the Proposal #79, Proposal body points at history and comments #81, Route narrative in issue comments #82, Journal the Iteration briefing at Run start #80, Briefing visible on the live card and in history #83 (needs the behavior, the comments, and the visible briefing; style convergence is excluded — it does not gate the proof).

[JacobStephens2]

Live proof complete. Dispatched on JacobStephens2/career#83 after preconditions were confirmed (career enrolled in targets.toml, allowlist JacobStephens2, retry budget 2, review cap 20 with 0 awaiting, deployed tree updated to post-#88 main, career work/box checkouts created, grok auth + career token present, cycle unpaused). Also created the missing needs-info/awaiting-review labels in career, and temporarily peeled ready-for-agent off career#70/#81 so the cycle would pick #83 (restored afterwards).

Cycle 37 (pre-fix): attempts 1 on #83 and #84 both died in ~15s at boundary creation — grok.sh passed the image as the sbx agent name. Fixed by #90, merged, deployed (guardrail observed ref 1d44b1f, clean).

Cycle 58 (post-fix), scoped to JacobStephens2/career:

• Briefing visible on the live card and in history #83 attempt 2: 2 iterations, consecutive-noops (both iterations hit the 900s wall clock), exit 3. Draft PR JacobStephens2/career#104, branch tip merge-clean (removal commit b541813, history holds Seed/Iteration/Run-ended/Remove-scaffolding), Proposal body names task/area/bound/removal commit and points at history, notify comment posted, briefing journaled (run.briefing). Budget spent → given-up → ready-for-human with the evidence-carrying issue comment. No third dispatch possible.
• One shell across the window #84 attempt 2 (drain, same cycle): 4 iterations, consecutive-noops, exit 3. Draft PR JacobStephens2/career#105, tip merge-clean, same body/comment/briefing record, given-up → ready-for-human.
• Cycle halted none-eligible. Attempt-1 drafts #102/#103 were closed unmerged to unblock the retries; branches and records remain.

Acceptance: Run dispatched after preconditions confirmed — yes. Branch tip merge-clean with evidence in history and comments — yes (#104, #105). Briefing visible on live card and in history — journaled for all four Runs (run.briefing rows 46/53/67/74); render path is the deployed #83 code, covered by web/tests/test_run_briefing.py. Visual confirmation on the window behind the login is still yours.

Follow-ups worth tickets: grok iterations timing out at the wall clock (agent slowness, not a selector defect); failed-and-proposed issues can never retry while their draft stays open (proposal-open skip beats the retry route); box.unreachable/run.watch-failed advisories persist on Single-Host (pre-existing #60 gap).