Repository navigation
lessons: FIX-1649 epic wrap (cycle 21) - #2716
Conversation
…ic, proposal skipped Forty-seven rows: the epic PR and three amendments, fourteen specs and spec amendments, and twenty-nine implementation PRs, the closure #2709 among them. No direction artifact took a second round, and no implementation PR took more than two fix waves. The closure's runs found three reused controls WRONG in cycle 19's three shapes, three child checks broken by later merges, and a runner that passed with no verdicts. Each was found on its designed path, so no grounding change is proposed. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Cnm5DbhFmkarpesBpPQ9Dq
|
The latest updates on your projects. Learn more about Vercel for GitHub. 3 Skipped Deployments
|
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
Bugbot couldn't run - usage limit reachedBugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit. A user or team admin can review and increase usage limits in the Cursor dashboard. (requestId: 6710b6de-a793-41f3-a764-bc4174823e7a) |
There was a problem hiding this comment.
Simplify review (cycle 21 ledger)
This PR is documentation-only: an index row plus a new cycle-21.md epic-wrap entry. I compared it to cycles 19–20, the distill-lessons skill, and the cycle-ledger header contract.
Verdict
Approve. The shape matches the instrument you have been using since roughly cycle 15: narrative lead → Method → per-PR table → Load → closure narrative → proposal decision → candidates / filed → forward claims → Finding map with checksum. For an epic with 47 scored artifacts, a wide table is expected, not accidental scope creep. The substantive call—proposal skipped because the recurring closure-control shapes are only visible on the closure’s first run—is argued in The class and Candidates considered and dropped, and it lines up with cycle 19’s precedent.
I do not think this entry invented a heavier doc species than cycle 19/20; it is longer mainly because the sample is larger (29 implementation PRs vs 15 in cycle 20).
Where complexity is real (but mostly inherited)
| Area | Take |
|---|---|
| 47-row table + Finding map | Duplication between table cells and map bullets is established convention (thread-id audit trail + footer sums). The map is not gratuitous; it is how prior cycles stay rebuildable. |
| Dense table cells | Each row packs bot limits, POC refs, and multi-finding prose. That drives token cost more than row count. A slimmer cell convention (counts in table, thread detail only in map) would be a process change for future cycles, not a blocker here. |
| Manual REST harvest | Method says every thread was read via GitHub REST. That is how distill-lessons is written today; there is no in-repo generator. A structured sidecar (JSON per PR) or a small verify script would simplify future wraps; this entry is consistent with current practice. |
Optional simplifications (non-blocking)
These would shrink maintenance surface next time; I would not hold this PR for them unless you want to tighten the template repo-wide.
- “Proposal skipped” twice — opening paragraph and
## Proposal skippedlargely repeat. One place (lead + short cross-ref) would be enough. - Method inventory — listing every PR number before the table duplicates the table itself; a shorter scope paragraph (“47 rows below; rules follow cycle 20”) would age better.
- Per-row “Merged head … unread” — already summarized in Load / claims; repeating on ~25 rows is low signal (cycle 20 does the same; could be one Method bullet).
- Filed block length — operational notes (Linear lag, #2642 open, FIX-1735 on #2678) are useful but dilute the closure-control thesis; cycle 20’s filed section is similarly wide—worth a template discussion, not a rework of this file alone.
Cleaner alternative approach (bigger than this PR)
If the goal is less human duplication, the lever is upstream of this markdown:
- Generate table rows + Finding map + footer from cached API payloads + a small human classification file (classes, felt-off, upstream-fix column).
- Keep hand-written only judgment sections (lead, closure runs, The class, candidates, filed narrative).
- Add a 20-line verifier that parses class counts per PR and asserts footer = sum(rows)—footer arithmetic on this entry checks out today (104 impl + 62 direction = 166), but there is no CI guard.
That is a distill-lessons / tooling change, not something this PR should absorb mid-wrap.
What I would not change
- Dropping the mega-table (ledger contract).
- Dropping the Finding map without replacing the thread-index + checksum role.
- Forcing an upstream proposal when the entry correctly records skipped with evidence.
No changeset needed (internal docs).
Sent by Cursor Automation: Review For Best Approach
| - **The closure is where reuse is graded.** Run 1 found all three, and run 2 confirmed all six | ||
| controls RIGHT on the next `main`. That cost one fix PR (#2641, two waves) and one re-run. | ||
|
|
||
| ## Proposal skipped |
There was a problem hiding this comment.
Simplify (optional): **Proposal skipped.** already appears in the opening (line 17), and this section mostly restates that verdict plus one orchestration.md quote. Collapsing to a single location (or a one-line ## Proposal skipped with “see lead”) would cut ~6 lines without losing the instrument.
| 3. **Unread merged heads.** Baseline: 25 of 29. | ||
| 4. **Changeset length**, once more. Baseline: 0 of 7 fragments that Codex read. | ||
|
|
||
| ## Finding map |
There was a problem hiding this comment.
Simplify (process, not a blocker): The Finding map mirrors class×counts already in the table; its unique value is T/SL positions and the footer checksum (lines 315–318). That split is established in cycles 15–21. If you ever want a leaner doc, generate map+footer from the table (or vice versa) rather than maintaining both by hand—this file is doing the right thing for today’s template.
| finding this entry records. It is not yet a proposal, because the earliest point that can see each | ||
| one is the closure's own control run (see *Candidates considered and dropped*). **Proposal skipped.** | ||
|
|
||
| **Method — scope.** Forty-seven reviewed artifacts. They were found from FIX-1649's 35 Linear |
There was a problem hiding this comment.
Simplify (optional): Method pre-lists almost every PR number that appears again in the table (lines 52–100). Cycle 20 does this at smaller scale; here it is a second inventory. A shorter scope block (“47 artifacts; epic + amendments + specs + impl including closure #2709”) would be easier to maintain when numbers shift.
There was a problem hiding this comment.
Code Snob: already tight. No POC.
I read docs/internal/cycle-ledger/cycle-21.md, the new index row in docs/internal/cycle-ledger.md, and cycle 20's finding map, which is the shape this entry matches. distill-lessons asks for one row per artifact, the class counts, the prior-cycle scores, and the next claims. That is the 47-row table, the closure class, the dropped candidates, the filed items, and the finding-map checksum (104 implementation + 62 direction).
The lead and ## Proposal skipped state the same decision, and the method lists PR numbers the table lists again. Cutting both is about ten lines. The rows, the three control shapes, and the drops stay. The unread-head SHAs and the no-Codex marks sit inside rows that remain one line either way, and several name the head a bot did read (49fd744 on #2709, 2008bbb on #2641). Those SHAs are what the cycle 19 score cites. Moving the finding map into the cells adds the thread index to every row.
The PR body repeats the lead and the three filed decisions. That text is the ask. The cycle file is the record.
Head 9c1009e75.
Sent by Cursor Automation: Code Snob
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 9c1009e75a
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".
| - the epic PR #2421 and three amendments: #2423, #2619, and #2648 (FIX-1737's spec, which also | ||
| amends the epic); |
There was a problem hiding this comment.
Include the reviewed epic-wrap amendment in the sample
The declared exhaustive scope omits PR #2714 (spec/FIX-1649-wrap), even though its branch matches the stated epic-ID search and it merged roughly seven hours before this ledger commit. That PR contains review-fix commits for both Cursor (daa7b797) and Codex (5a51b7bb), making it a fourth reviewed epic amendment; excluding it leaves the 47-artifact sample and all downstream direction/finding totals incomplete.
Useful? React with 👍 / 👎.


What's asked of you
Nothing to approve in the grounding. This PR adds ledger rows only. The proposal pass is skipped, and the reason is below. Read the rows and the filed items. Merge when the rows look correct.
Why no proposal. Review cost was low on every measure. No direction artifact took a second round. No implementation PR took more than two fix waves (29 waves over 29 PRs; cycle 20 had 57 over 15). The expensive rework came from the closure runs, and the closure is the designed place to find it:
main(FIX-1737 B, C and D, and FIX-1718). Closure item 3 exists to catch this.GOAL_PART=bogusprinted PASS with 0 verdicts. BP-003 already covers this, and one wave fixed it.What would change my mind. A third closure where the reused controls break in the same three shapes, and the change that broke them merged before the closure plan. Then the plan could have seen it, and the closure plan's control table should get one line. The cycle file records this as claim 1.
Your decisions, filed and not proposed
relatedlinks to FIX-1663. The epic wrapped with all four open. Re-parent or close them.1d1e1cb). None of those findings has an issue. The rule says each finding becomes a child.What's in it
docs/internal/cycle-ledger/cycle-21.md: the new cycle file. It has 47 rows, the closure's runs, the control class, the dropped candidates, the filed items, the scores for cycles 18–20, and a finding map.docs/internal/cycle-ledger.md: the index row for cycle 21.Other filed items are in the cycle file: #2434 merged 5 minutes after it opened, before Codex; Linear states lag the merges; POC #2642 is still open; a zero-check guard for
goals/lib/verdict.mts.Scores: cycle 19's control-unit fix holds on one unconfounded instance (#2709 had no control-unit finding), so that claim closes. Cycle 20's race-primitive trigger had no qualifying PR. No changeset needed: internal docs only.
🤖 Generated with Claude Code
https://claude.ai/code/session_01Cnm5DbhFmkarpesBpPQ9Dq
Generated by Claude Code
Note
Low Risk
Internal documentation only; no runtime, auth, or product code changes.
Overview
Adds cycle 21 to the internal development-loop ledger for the Shift Manager lab shell epic wrap (FIX-1649, 2026-10-04).
docs/internal/cycle-ledger/cycle-21.mdis a new ~318-line entry: 47 scored PRs (epic, specs, 29 implementation including closure #2709), load metrics vs cycle 20, closure run timeline, the recurring reused-control failure class (three WRONG controls on first closure run), skipped proposal rationale, filed follow-ups (reviewer exhaustion at 16 artifacts, open children, closure branch fixes without issues), scoring of cycles 18–20 claims, and claims for the next cycle.docs/internal/cycle-ledger.mdgains one index row pointing at that file.Reviewed by Cursor Bugbot for commit 9c1009e. Configure here.