Skip to content

handoff.md

Latest commit

 

History

History
90 lines (79 loc) · 5.79 KB

handoff.md

File metadata and controls

90 lines (79 loc) · 5.79 KB
Raw

Handoff — mt#3824

Done

  • Root-caused and fixed .minsky/hooks/guard-feedback-shape.test.ts's size-ceiling failure. calibration-review-cadence-detector's formatCadenceWarning rendered one line per review-due calibration log with no cap, so its size scaled with due-log count — driven both by real repo activity and by wall-clock time alone (every registry entry with a liveSinceDate ages toward "never-fired" as its reviewByDays window closes, independent of file content, including inside the guard's own isolated canary).
  • Fix: capped formatCadenceWarning at MAX_DUE_LOGS_LISTED = 2 named lines + cappedEvidenceLines's "…and N more" tail (mt#3705's shared bound), trimmed the never-fired reason clause, and re-declared registry.ts's denialMessageSizeChars from 450 → 800 (the true measured worst case, proven identical at 2 vs 20 due logs once capped).
  • Corrected guard-feedback-shape.test.ts's stale FEEDBACK_SHAPE comment (claimed a nonexistent slice(0, 8) cap).
  • Added hermetic formatCadenceWarning tests in .minsky/hooks/calibration-review-cadence-detector.test.ts proving the ceiling holds identically at 1/3/8 due logs (pure function, no fs/wall-clock).
  • Verified BOTH directions of success criterion 2 live: guard-feedback-shape.test.ts green with a real log due (current repo state); run() returns null in an isolated temp workspace with every registered log's watermark freshly advanced (none due).
  • Negative control: git stash reproduces the original "639 chars > declared 450" failure; git stash pop restores green.
  • validate_typecheck / validate_lint: both clean.
  • Investigated but did NOT reproduce the spec's ## CORRECTION section's suspected second cause (duplicate-signature-scan dropping out of the producing set) — its canary is expects: "calibration" by design (recorded skip, no DB in canary) and it isn't in the current producing array at all. Not a live issue; not addressed.
  • Created PR #2701 (#2701), commit f66cb0962.
  • Reviewer round 1: CHANGES_REQUESTED, one BLOCKING finding — .minsky/hooks/calibration-review-cadence-detector.test.ts:470 hardcoded DECLARED_CEILING = 800 duplicating the registry's value (drift risk).
  • Fixed locally: read the ceiling from GUARD_REGISTRY directly instead (.minsky/hooks/calibration-review-cadence-detector.test.ts, import GUARD_REGISTRY from ./registry). Committed locally as 20f4c0b3e — verified green (bun test --preload ./tests/setup.ts --timeout=15000 ./.minsky/hooks/calibration-review-cadence-detector.test.ts ./.minsky/hooks/guard-feedback-shape.test.ts → 54 pass, 0 fail), typecheck/lint clean.
  • PR body updated via session_pr_edit to describe the R1 fix (title/body only — does not push code).

In progress / blocked

Commit 20f4c0b3e is NOT pushed to the PR branch yet. Confirmed via mcp__minsky__git_log(ref: "origin/task/mt-3824") — remote HEAD is still f66cb0962.

Every push mechanism tried fails deterministically with the SAME unrelated failure: mcp__minsky__git_push (plain and force: true), mcp__minsky__session_commit, mcp__minsky__session_update (plain and force: true) — all fail-close on scripts/run-tests-gated.ts's pre-push hook, which is red on src/adapters/shared/commands/calibration.test.ts:345 (observability.calibration-review — --ack covers all review-due legs (mt#2878) > still skips a newly-covered-leg log whose ask is open when no askId is supplied — Expected: false / Received: true).

This is mt#3818's bug, not mine — confirmed via mcp__minsky__git_diff(from: "main", to: "HEAD", nameOnly: true): my diff touches 6 files, none of them src/adapters/shared/commands/calibration.test.ts. mt#3818's own spec (read directly) diagnoses this exact failure as a calendar time-bomb: the knowledge-acquisition registry entry's liveSinceDate: "2026-07-23" + reviewByDays: 14 crossed today (2026-08-06), flipping an unrelated process-global watermarkAdvanced flag the test asserts on instead of the per-path property it should. mt#3818 is IN-REVIEW with PR #2699 as of this writing (checked repeatedly, still IN-REVIEW, not yet merged).

One push DID succeed once — session_pr_create's own internal push (used to create PR #2701 and land f66cb0962) — but that mechanism isn't independently invocable for an existing PR; every tool I could find that updates an EXISTING PR's code (git_push, session_commit, session_update) goes through the hook-enforced path and fails identically. Do not keep retrying these blindly — 4 attempts already made, all identical.

Remaining

  1. Poll mcp__minsky__tasks_status_get(taskId: "mt#3818") until it reaches DONE (or check PR #2699 merged).
  2. Once merged, push commit 20f4c0b3e (try mcp__minsky__session_update first — a rebase onto the now-fixed main should make the gated suite green again; fall back to mcp__minsky__git_push if needed).
  3. Re-run mcp__minsky__session_pr_wait-for-review(reviewer: "minsky-reviewer[bot]", since: "2026-08-06T19:13:37Z") (the R1 review's submittedAt) to pick up the bot's re-review of the R1 fix.
  4. On APPROVE: do NOT merge — the dispatching task explicitly said "Do not merge." Report status and stop. On further CHANGES_REQUESTED: iterate per the standard convergence loop.
  5. Confirm CI's build job green on the PR (success criterion 4) — check via mcp__minsky__session_pr_checks or mcp__minsky__forge_check_runs_list once pushed.

Known issues

  • None beyond the push blocker above. All local verification (both test files, typecheck, lint, both directions of the state-independence claim, negative control) is solid and unaffected by the push blocker — it's purely a remote-sync problem caused by an unrelated in-flight task.