CalvinBackup/gstack
1
0
Fork 0
mirror of https://github.com/garrytan/gstack.git synced 2026-09-21 04:10:47 +02:00
Code Issues Packages Projects Releases Wiki Activity

docs(TODOS): two of three first-execution findings fixed (skillify family, context-restore)

Browse Source 
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
Test
2026-08-29 15:04:46 +00:00
co-authored by Claude Fable 5
parent f239a2ed4c
commit 8f81c6a671
1 changed files with 13 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files
TODOS.md
+13 -12
View File
@@ -557,18 +557,19 @@ coverage fill. Remaining, in rough priority order:
- **P3 — codex-e2e-plan-format's testIfSelected names have no map keys** - **P3 — codex-e2e-plan-format's testIfSelected names have no map keys**
(run-all only today) + 15 E2E / 2 judge PHANTOM touchfiles keys select (run-all only today) + 15 E2E / 2 judge PHANTOM touchfiles keys select
tests that exist nowhere — add keys or delete, one sweep. Effort S. tests that exist nowhere — add keys or delete, one sweep. Effort S.
- **P2 — first-execution rot found by the sliced lane's first live run** - **P3 — first-execution rot from the sliced lane's first live runs: 2 of 3
(PR #2721): (a) `skill-e2e-skillify` "provenance-refusal" fails FIXED** (PR #2721): (a) ✅ skillify family — root cause was HOME==cwd
deterministically (CI x2 + local) — its child reports "The skillify skill making claude treat <cwd>/.claude/skills as the PERSONAL dir (project
is not registered in this environment"; the test's `HOME: workDir` env skills never registered); all three tests now use a fresh HOME subdir,
override breaks project-skill registration under the hermetic env while the refusal test gained a not-registered tripwire + assistant-text-only
its 5 sibling tests (same fixture, no HOME override) load skills fine. matching (the skill body echo could pass vacuously), and the siblings now
Start by diffing the passing siblings' env. (b) genuinely exercise the Skill-tool path (verified paid, 5/5).
`skill-e2e-session-intelligence` "context-restore loads latest" failed (b) ✅ session-intelligence context-restore — assertion was prose-matching
twice in CI but passes locally (4/4) — CI-container/stochastic shape; over stochastic wording; now verbatim RESTORED-marker + tool-call
watch across runs. (c) `tpa-apple-ban` failed only on retry attempt 2 — corroboration with a stronger older-file negative (3/3 paid green).
flake watch. All three are the coverage contract WORKING (these files ran (c) `tpa-apple-ban` failed only on retry attempt 2 once — flake watch
in no CI lane before). Effort S each. only. The lane finding these on first execution is the coverage contract
working.
- **P2 — make-pdf image promotion is per-render nondeterministic on CI**: - **P2 — make-pdf image promotion is per-render nondeterministic on CI**:
two renders of the same fixture SECONDS apart in one CI job produced 2 vs two renders of the same fixture SECONDS apart in one CI job produced 2 vs
3 landscape pages (an image's promotion depends on load timing at render). 3 landscape pages (an image's promotion depends on load timing at render).