Task grain and review cadence: largest-ownable tasks, frontier review, resume-based fix loop, sde-telemetry (v7.59.0) - #81
Conversation
…ontier review, resume-based fix loop (v1)
…e-based fix loop, opus worker text; land scripts/sde-telemetry
…lines, reviewed-clean frontier, stale-resume context, sonnet→opus escalation, acceptance fixes)
…v1.1 re-test observations
… by message id and reports active time (codex P1+P2s)
…lector v2 classification)
…, frontier review, resume-based fix loop, sde-telemetry)
…sk-prefixed noun descriptions (codex round 2)
…routes through the frontier; reviewer template task-head/checkout-head (codex round 3)
…ut, and worktree inputs (codex round 4)
…e shipped (v7.59.0)
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 1c558ea2f7
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| (re.compile(r"^(implement|execute|run) task|^task \d+:" | ||
| r"|^t\d+ (implementer|executor)", re.I), "executor"), | ||
| (re.compile(r"^t\d+ fixer|\bfix", re.I), "fixer"), |
There was a problem hiding this comment.
Task N: dispatch as an executor
When the fresh-fixer fallback is described naturally as Task 2: fix pagination (or a reviewer as Task 2: review pagination), the broad ^task \d+: alternative matches first, so role_of never reaches the fixer or reviewer pattern. This misreports the role counts used to monitor whether the new execution cadence reduces fixer/reviewer dispatches; restrict this alternative to executor-shaped descriptions or prioritize explicit fix/review descriptions.
Useful? React with 👍 / 👎.
| if t - previous > IDLE_GAP: | ||
| active += previous - segment_start | ||
| segment_start = t |
There was a problem hiding this comment.
For any test, build, or model/tool operation that legitimately takes more than 15 minutes, the transcript contains a gap over IDLE_GAP and this branch discards the entire interval as idle. Controlled-track executors and reviewers can run commands this long, so the reported active duration—and therefore comparisons against the monitoring baseline—is understated; identify actual wait/resume boundaries from transcript events instead of treating every long gap as an idle resume.
Useful? React with 👍 / 👎.
| Your review is read-only on this checkout. Do not mutate the working | ||
| tree, the index, HEAD, or branch state in any way. | ||
| tree, the index, HEAD, or branch state in any way. A check that must | ||
| see this task's own tree — a focused test, a named risk — runs in the | ||
| detached worktree the controller names, not in this checkout, which | ||
| may sit past the task's head. |
There was a problem hiding this comment.
When an immediate review runs at the task head, the controller omits the Checkout line and does not create a detached worktree, yet this instruction still requires every focused test or named-risk check needing the task tree to run in a detached worktree that the controller names. Likewise, during a deferred review the reviewer may discover the need for such a check only after reading the diff, after the controller has already decided whether to provision one. In either case the reviewer can be left unable to perform an otherwise warranted focused check; make this conditional on a named worktree actually being supplied, allow the reviewer to request/create one, or provision one for every affected review.
Useful? React with 👍 / 👎.
Historical-spec before/after (human-proposed, 2026-08-21)Eight shipped controlled-track specs (2026-08-03..20) re-cut from their pre-execution spec text at fork-point checkouts: 2 opus planners on the NEW writing-plans text + 1 opus planner on the OLD text per spec (breakdowns with Interfaces only). The shipped plan + its real run (telemetry/ledger) is the before.
Totals over 8 specs: shipped 72 tasks; controls (n=8) 81; new-text medians 49.5; new frontier review points 37. Read: shipped 72 tasks → new-text medians 49.5 (−31%; −39% vs same-model old-text controls, which cut finer than the frontier-authored shipped plans in 5/8). Frontier review points from the new breakdowns' |
…24 breakdowns) recorded
…worktree sentence conditional on Checkout (GitHub Codex P2s)
1 participant
What
Re-tunes the controlled track's unit economics — the doctrine change the closed experiment spec (PR #72) was designed to test, adopted now with telemetry monitoring instead of a paired double-run (human's call, 2026-08-21).
Interfaces: Consumesnow names the producing task. "Bite-Sized Task Granularity" → "Bite-Sized Steps" (content unchanged).executedlines (base, handles) so compaction never re-executes an executed-but-unreviewed task. Model text aligned to the standing directive: opus/high default, sonnet for simple tasks, never fable.scripts/sde-telemetry: per-session dispatch counts by role, span and active time, token totals deduplicated by message id, start-time-sorted dispatch list.docs/doperpowers/specs/2026-08-21-task-grain-and-review-cadence-design.md: quotes the skill-bound text verbatim, carries the four-run baseline (2026-08-18..20) and the monitoring protocol (next three controlled-track features; reopen criteria), Decision Log, and the smoke-check results.Evidence
Consumesnaming its producer 3/3; new SDE loop on the PR Board client reads the paged surface (read-surface program, patch 2) #74 plan → frontier schedule with batched siblings and concurrent reviewers 5/5 reps (v1:{1}|{2,3,4,5}|{6}|{7}×3; v1.1: one split at the Task 3→4 shared-file overlap, one batched), ledgerexecutedlines and no re-execution after simulated compaction.tests/claude-code/run-skill-tests.sh --test test-subagent-driven-execution.shPASS against the worktree skill (viaclaude --plugin-dirshim).Monitoring
After each of the next three controlled-track features: run the collector, append a row to the spec's Surprises, compare against the baseline; reopen on the spec's §3 criteria.
v7.59.0 — version collision: PR #80 also bumps to 7.59.0; whichever merges second rebumps on the manifest conflict (no other file overlap with main).