Repository navigation
docs(lessons): FIX-2152 epic wrap, cycle 29 ledger - #4257
Conversation
Rows for the epic spec, its three amendments, sixteen issue specs and the implementation PRs. Three classes: a check over a committed generated table (fixed by #4235), an interleaving tail patched one pass at a time (cycle 27's class, again), and main fixes ported into eight open branches. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiH4wEhbNtsBnU8xobe
A chain of different cases on one structure passes the separate-race test each time, so 10.3's trigger never fires on it (#4239, #4085 via #4124, cycle 27's #4013). From the fourth pass with a finding on a structure an earlier pass changed, the worker answers the primitive question on the PR before fixing. Proposal for owner review; drop this commit to decline it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiH4wEhbNtsBnU8xobe
|
The latest updates on your projects. Learn more about Vercel for GitHub.
3 Skipped Deployments
|
|
@codex review |
|
Codex Review: Didn't find any major issues. More of your lovely PRs please. Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
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". |
There was a problem hiding this comment.
/simplify review
This PR does two jobs: a FIX-2152 cycle-29 ledger (measurement + baselines) and a repo-side enforcement tweak (issue-implement 10.3). The measurement is strong — especially the honest sampling method, the inventory-table (a) post-mortem, and tying the interleaving tail to why cycle-27’s epic-wake dispatch would not have fired on coordinator-attributed PRs. I would not ask for a rewrite of the epic analysis.
Where the approach adds complexity is mostly packaging and repetition, not the underlying insight.
Went further than necessary
-
Ledger + skill change in one PR — Cycle 27 explicitly kept orchestration out of the lessons PR (“not built in a lessons PR”, cycle-27 enforcement (a)). Cycle 29’s own Proposal says the 10.3 line is a separate commit the owner can take or drop, but it still ships in the same review surface. That forces reviewers to accept/reject policy and history together and makes it harder to merge the ledger while debating the fork.
- Cleaner: ledger-only PR (index +
cycle-29.md); follow-up forissue-implement(+ optionaldistill-lessonsrule row / coordinator brief). Or keep one PR but make the skill hunk clearly optional in the PR description and drop duplicated enforcement prose from the ledger.
- Cleaner: ledger-only PR (index +
-
The count trigger is narrated ~4× in
cycle-29.md— class (b) (37–40), Proposal (62–63),## (b) A count, not a judgment, Enforcement (b), and Accepted. Cycle 27 also has overlap, but this entry is longer and the epic-specific tables already carry the evidence.- Cleaner: one canonical section (
## (b)+ table); class bullet = one line + anchor; fold Proposal into the lede or Accepted only.
- Cleaner: one canonical section (
-
Two “Accepted” mechanisms for the same class without a supersession line — Cycle 27 Accepted
epic-wakeround-4 dispatch (not built). Cycle 29 Accepted new 10.3 prose and Rejected building cycle-27’s step. That pivot may be right given coordinator attribution, but readers grepping the ledger will see conflicting Accepted rows.- Cleaner: one explicit sentence in cycle-29 Accepted/Rejected: “For project-coordinator PRs, cycle-27’s epic-wake step remains owner-side; repo-side mirror is 10.3 pass-4 sentence.” Optionally a forward pointer in cycle-27 (if you amend old cycles).
-
issue-implement10.3: second bold block — Main already has the long “second race / separate race / name primitive” paragraph (cycle-20/27). The new paragraph repeats “answer on the PR before fixing” and introduces a second trigger agents must reconcile. Cycle 27 rejected rung-5 prose because 10.3 already said it; this adds prose for a different failure mode (different-case chains), which is fair, but the shape can be simpler.- Cleaner: one appended sentence on the existing paragraph (or a single Count override sentence), and tie “fourth review pass” to existing 10.1/10.7 round counting (
prFeedbackRounds/ cap language) so standalone and coordinator paths share one definition.
- Cleaner: one appended sentence on the existing paragraph (or a single Count override sentence), and tie “fourth review pass” to existing 10.1/10.7 round counting (
-
Method block placement — Cycles 25–27 put Method: scope (and round definitions) before the big tables. Here it lands after Proposal (65–75), so a 219-line doc explains sampling late.
- Cleaner: move Method (+ out-of-sample + jhoffner note) immediately after the opening summary.
-
Duplicate cycle-27 rows in the (b) table — #4013, #3795, and the rung-3 rejection are already in cycle-27. Keeping them helps scoring continuity but bloats cycle-29.
- Cleaner: epic-only rows in the table + link: “Mechanism / rung analysis: cycle-27 § Enforcement (a).”
Possibly the right call (uncertain — owner judgment)
- Count at pass 4 vs judgment at pass 2 — The empirical case that “separate race” never fires on different-case chains is convincing for #4124/#4239. I’m less sure a pass count alone is the right enforcer:
#4145(grep) hits ~4 passes on one structure with no primitive needed; cycle-29 itself notes rung-3 automation misfires. The doc acknowledges rung-5 prose + coordinator brief — that may be the ceiling until something outside the repo counts “same structure.” - Epic row without job audit — Job audit deferred is honest; the
ways per jobrow (139–147) may still read as audited. Mark provisional or shorten until the coordinator runs audit.
What I would not block on
- Cycle 29 while 28 is an open draft — the index note is enough if intentional.
- Class letter (a) = inventory here vs interleaving in cycle 27 — fine if cross-linked once at first (b) mention.
- Heavy (a) commit tally — high signal for a one-off class; could move to a footnote for scanability only.
Verdict
Not approving yet — not because the thesis is wrong, but because the combined diff is harder than it needs to be (bundled policy + repeated sections + cycle-27 Accepted tension). I’d be comfortable approving a ledger-only slice or a deduped single PR once supersession and 10.3 shape are tightened.
Skipped recommendations (larger than this PR)
- Implement cycle-27’s
epic-wakeround-4 conditional (still the only in-repo counter today). - Merge queue / branch-up-to-date (owner ask, third cycle — agree it’s not more ledger text).
distill-lessonsrule→enforcer row for pass-4 answer (cycle-26 pattern); optional if 10.3 stays prose-only.
Sent by Cursor Automation: Review For Best Approach
…-yzt8o2 # Conflicts: # docs/internal/cycle-ledger.md Co-authored-by: jhoffner <579835+jhoffner@users.noreply.github.com>


Requested by Jake · project thread
What's asked of you
Two things, both small:
issue-implement10.3 (the second commit). From a PR's fourth review pass with a finding on a structure an earlier pass already changed, the worker answers the "is this a missing primitive?" question on the PR before fixing. Drop the commit to decline it. The ledger stands either way.mainwent red.epic-wake, so cycle 27's proposedepic-wakestep would not have reached them.Draft on purpose: lessons go in after your review.
What the ledger found (cycle 29, FIX-2152)
The new entry is
docs/internal/cycle-ledger/cycle-29.md. It is numbered 29 because the open draft #4245 already claims 28.main, three restore-and-revert pairs, and one redmain. Remove the lab README file inventory #4235 deleted it. Already fixed; nothing new proposed.mainred, ports everywhere. Theuser_inputrename and test(workforce): bound the next host's shutdown drain in the call crash tests (FIX-2269) #4204's drain were ported into four PRs each.The owner's own use of the labs after wiring filed about twenty children, seven of them package bugs no goal check had caught. That is epic D1 working, so it is not scored as rework.
Not done here
inboxItems, from Cursor's note on FIX-2157: wire the Claude Code-style lab's threads mode to the real packages #4094.Evidence
mainfor every merged PR.node scripts/check-retired-terms.mjs: PASS.node scripts/check-mechanism-anchors.mjs: PASS.Generated by Claude Code