Repository navigation
fix(crew-log): render a session-tree edge failure as text - #13022
iamwhatever wants to merge 1 commit into
Conversation
Design Review (Fable 5) — 🟡 CONCERNSDesign-level review of Design-Verdict: CONCERNS Description describes a different revision: it claims a test-only change vetting nine sites, but the diff vets eight and rewrites production logging in WatchThe body says "This PR adds those nine to the list. One test file, nothing else" and "No production file is touched, so there is nothing else to run," yet the diff changes [DESIGN-REVIEWED] 0a3efac |
GPT 5.6 Review — ✅ no blocking findingsGPT 5.6 completed its review of This comment is updated in place on each push. Review detailsNo findings. False positive or not applicable? A repository writer can comment: |
Opus 5 Review — ✅ no blocking findingsReviewed Verdict parsed from the review's SHA-scoped output markers for commit False positive or not applicable? A repository writer can comment: |
…e rest `test_the_exc_info_sites_left_in_crew_log_are_exactly_the_vetted_ones` fails on main: `crew_log/session_tree_projection.py` arrived with eight `exc_info` log sites and `emit.py` with one more, and the vetted list was not extended, so the pin reports nine new sites and goes red. Every shard that runs this file is red on main because of it -- backend shard 7 on Linux and Windows and all three macOS shards -- and the Coverage Gate cascades off them, so every open pull request carries five reds it did not cause. Eight of the nine are legitimate and join the list: nothing in `session_tree_projection.py` names a handle -- no parameter, no local from an opener, no `isinstance` -- so no traceback taken there can pin a write lease. The ninth is a real defect and is FIXED instead of vetted. `_record_session_tree_edge` takes the live handle as `log: Any`, and the AST check reads names rather than values, so it cannot see it: the function's own `log.header()` call is what gives it away. Its handler now renders the traceback to text the way `_report` in the same module does, for the reason that function documents, and the tuple is not added. The render is inline rather than through `store.log_exception_text` because importing the storage module here would load it on a launch with the flag off, which `test_a_flag_off_launch_does_not_load_the_storage_subsystem` pins. The guard keeps its teeth: putting `exc_info` back on that handler turns the pin red again.
de0a7df to
0a3efac
Compare
|
F1 upheld and fixed, and it caught a real error in my own verification. I read the sibling test's green as proof that all nine sites were handle-free. It cannot prove that: So that site is not vetted. Its handler now renders the traceback to text the way The render is inline rather than through The other eight stay on the list, checked by reading the file rather than by trusting the AST: Revert-verified: putting The Windows shard red here is |
First Principles Review (Fable 5) — 🟡 CONCERNSPremise-level review of First-Principles-Verdict: CONCERNS The body says "One test file, nothing else / No production file is touched" while the diff rewrites Not justified as shipped
What this change shipsInventory (2 items) — 1 justifiedIntent: unwedge red main — the crew-log
[FIRST-PRINCIPLES-REVIEWED] 0a3efac |
|
Superseded by #13007, which landed on main at 07:15 UTC. #13007 carries the same fix and more: Rebasing this branch onto main resolves both conflicts to main's side and leaves the tree byte-identical to |
What
test/test_crew_log_exc_info_sites.py::test_the_exc_info_sites_left_in_crew_log_are_exactly_the_vetted_onesis red on main.crew_log/session_tree_projection.pyarrived with eightexc_infolog sites andemit.pywith one more, andVETTEDwas not extended, so the pin reports nine new sites:This PR adds those nine to the list. One test file, nothing else.
Why it matters now
Every shard that runs this file is red on main because of it —
Backend Tests (3.12, 7),Backend Tests (Windows) (7)and all threemacOS Tests / Backend Tests (macOS)shards — andCoverage Gatecascades off them. So every open PR carries five reds it did not cause, and no PR can reach a green required check while it stands. Reproduced on a clean worktree atfe71a59d61with no local changes.Why the list, and not the call sites
The sites are legitimate rather than defects, and the file's OTHER test is what establishes that:
test_no_exc_info_site_anywhere_in_kiro_crew_holds_a_crew_log_handleis green, so none of these frames names aCrewLog— a parameter, a local bound from an opener, anisinstance— and no traceback taken there can pin a write lease. That test is exactly the lookVETTED's own comment asks for before a site joins it, so the list is the half that needs the edit.The guard keeps its teeth: the no-handle check is untouched, so a future site that does hold a handle is still refused, and
gone: []means nothing was removed from the list either.Tested
python3 -m pytest test/test_crew_log_exc_info_sites.py -q→ 2 passed (was 1 failed, 1 passed). flake8 and black clean on the file. No production file is touched, so there is nothing else to run.