test(kanban): pin that carrying a ref does not cost the bundle beside it by Kyzcreig · Pull Request #866 · ANG-Ventures/hermes-agent

Kyzcreig · 2026-09-22T07:56:25Z

Stacked on #845 (which is stacked on #842). Test-only.

The carded defect is already fixed — this closes the GATE, not the bug

t_e41d3dd4 was written against PR #842 head 3ca4b600a3, where the cleanup
carry-forward read only previous["refs"]. Two commits landed on #842 after that
base — a61dda5edf ("close FleetReview's 2xP1 on the cleanup relaxation") added
carried_bundles and gated the ref-completion branch on not bundles, and
7afe1b9e37 re-keyed the coverage test on absent.

Re-measured on pristine #845 head 7c3e0b6db1 (impl blob 370e5b30f1, byte-identical
to this branch's): both shapes the card measured — bundle-only partial loss and mixed
partial loss — already carry the bundle and record kind: "bundle". The bundle-only
shape is additionally pinned by the pre-existing
test_a_bundle_vouched_missing_repo_is_carried_into_the_rewritten_survivor.
No behaviour change is needed. Zero lines of hermes_cli/kanban_survivor.py in this diff.

What IS missing: the MIXED shape has no gate

carried (refs) and carried_bundles are two independent collections read out of the
SAME recorded survivor. Every pre-existing partial-loss test loses a repository vouched
for by exactly ONE of them, so nothing in the suite exercises "carry a ref and a bundle
at once".

Mutant M2 — keep the bundle carry-forward, but disable it whenever a ref is also carried:

carried_bundles = [] if carried else [b for b in (previous or {}).get("bundles") or () ...]

Under M2 the mixed shape silently drops the bundle and relabels the record kind: "ref",
telling an operator the work is pushed when its only copy is an orphaned bundle attachment.

Measured (my run, isolated `--basetemp`, four survivor files)

arm	result
HEAD, full four-file suite	96 passed (95 before this test)
M2 vs the 95 pre-existing tests (new test deselected)	95 passed, 0 failed — SURVIVES
M2 vs the new test alone	1 failed — killed only by this pin

Mutant applied in a throwaway detached worktree, reverted, worktree removed; both
checkouts git status --porcelain clean afterwards. ruff check clean on the test file
and on kanban_survivor.py.

Verify:

.venv/bin/python -m pytest \
  tests/hermes_cli/test_kanban_survivor.py \
  tests/hermes_cli/test_kanban_survivor_stale_bases.py \
  tests/hermes_cli/test_kanban_survivor_authority.py \
  tests/hermes_cli/test_kanban_external_survivor.py \
  --basetemp=/tmp/<isolated> -q

Notes

HOTSPOT: hermes_cli/kanban_survivor.py is a six-way collision surface
(kanban: make the operator survivor reachable when a recorded repo vanishes (t_49c1dec1) #837/t_49c1dec1, kanban: reclamation consults the RECORDED survivor instead of re-deriving from stale bases #842/t_a49e8a28, fix(kanban): preserve home-clone work via verified canonical live trees #796/t_f31d9d5a, t_de2e348e, test(kanban): pin the cleanup relaxation's coverage decision per input shape #845/t_1bcd8aec, this).
This PR's only conflict surface is tests/hermes_cli/test_kanban_survivor_stale_bases.py
— it appends to the end of the file.
Land order: kanban: reclamation consults the RECORDED survivor instead of re-deriving from stale bases #842 -> test(kanban): pin the cleanup relaxation's coverage decision per input shape #845 -> this. Re-measure the baseline on the MERGED tree;
t_62f0eb15 has a tests/conftest.py change in flight.
Fork-only subsystem; no upstream PR (kanban_survivor.py does not exist in
NousResearch/main).

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

Kyzcreig · 2026-09-22T08:07:37Z

FleetReview

Reviewed with 2 of 3 model families — openai unavailable.

Confidence: 4/5

Findings

P1 hermes_cli/kanban_survivor.py:496 — Carried refs are merged with previous["refs"] again on the elif claimed: arm, duplicating recovery-index entries when repos exist but capture nothing
P3 tests/hermes_cli/test_kanban_survivor_stale_bases.py:1046 — Set-based refs assertion in the new test cannot see a duplicated carried ref

FleetReview provenance · models: C=claude-code-opus-5, D=grok-4.6, G=grok-4.6 · cost: $4.23 · duration: 11m 05s · rounds: 1 · files examined: 1

…t shape `missing <= _vouched_repositories(previous)` is the whole of what lets a reaper DELETE a workspace whose recorded repository is gone. It was pinned for ONE cell of its input matrix: a ref-shaped survivor under total loss. Three mutations of the other cells left all 78 survivor tests green. Tests only; no production change (`git diff hermes_cli/` is empty). ref/partial -- `if repos: carried = [...]` (kanban_survivor.py:380-386) was reached by exactly one PRE-EXISTING test, which reads the COMPLETION snapshot and passes with the branch disabled. Disabled, the surviving repo satisfies the completion alone and the RECORDED ref for the vanished repo is silently dropped -- preserve() SUCCEEDS with the lost work pointed at nothing. Now driven through the CLEANUP path, asserting both refs survive. patch/any -- `_vouched_repositories()` excludes `repositories` because a patch is keyed by base SHA against a checkout. That was prose in a docstring. Widening the helper to collect `repositories` turns a fail-CLOSED hold into a reapable workspace whose only survivor is a patch against a base SHA that exists nowhere. Now pinned, with a direct shape assertion as teeth. bundle/total -- no test exercised a bundle-shaped recorded survivor at cleanup, so dropping `bundles` from the helper converts every bundle-backed reclamation into a permanent HOLD. Verified: that mutant passes 78/78 on the unmodified suite. Now pinned. Verified (CPython 3.11.15, pytest 9.1.1, isolated --basetemp, four survivor files): baseline #842 head 3ca4b60 : 78 passed with these tests : 81 passed mutant if repos -> if False : 1 failed, 80 passed (ref/partial test) mutant +repositories : 1 failed, 80 passed (patch test) mutant -bundles : 1 failed, 80 passed (bundle test) mutant -bundles, no new tests: 78 passed <- the gap was real Each mutant is caught by exactly its own test and nothing else. ruff clean. Scope note: enumerating the matrix surfaced a LIVE defect, not a test gap -- bundle-shaped survivor under PARTIAL loss. `carried` collects only `previous["refs"]` while `_vouched_repositories()` accepts bundles too, so the bundle passes the coverage test, carries nothing, gets erased from the record, and the hold clears. Measured on pristine code; carded as t_e41d3dd4 rather than fixed here, and deliberately NOT pinned -- pinning it would freeze the bug. Refs: t_1bcd8aec (found by argus on t_a49e8a28 / PR #842 round 1) (cherry picked from commit 8272cd4)

…exits Round-2 class sweep on the cleanup relaxation, re-derived over THREE axes (survivor shape x loss extent x workspace content) after #842 r2 landed both FleetReview P1 fixes. The omitted axis in round 1 was workspace content. Measured, not inferred: a 30-cell probe over {ref,bundle,patch,mixed,none} x {total,partial,partial-dirty} x {clean,loose} at 3afa908 found one unpinned family -- the `_loose_files()` guard is exercised only by REF-shaped survivors. Waiving it for `kind == "bundle"` at the relaxation's own exit flips four cells (bundle|partial|loose, mixed|partial|loose, and both partial-dirty variants) from fail-closed HOLD to a reclaim that rmtree's `qa-output/verdict.md`, and leaves all 84 tests GREEN. Two tests added: test_partial_loss_holds_for_loose_evidence_under_a_bundle_shaped_survivor test_a_bundle_shaped_survivor_does_not_buy_a_delete_for_loose_evidence Verified: 86 passed across the four survivor files. Mutation gate (mutation_gate_r2.sh) -- M4 (relaxation exit waived for bundles) kills the partial test; M5 (the `elif claimed:` arm waived) is an EQUIVALENT mutant, traced with sys.settrace: the total-loss bundle+loose case raises at the first site and never reaches the second, so it flips zero cells alone. The combined M4+M5 mutant kills both tests. Source restored pristine after every arm. Test-only: zero lines of hermes_cli/kanban_survivor.py changed. (cherry picked from commit 7d7f223)

…ping `missing <= vouched` is the whole of what lets a reaper delete a workspace whose recorded repository is gone. Every existing test loses exactly ONE repository, where the subset test and an intersection test agree — so relaxing it to `missing & vouched` (or `missing & vouched or missing <= vouched`) left all 86 tests on the four survivor files, and all 151 on the wider survivor surface, GREEN while flipping two cells from fail-closed HOLD to reapable. Adds the discriminating shape: two recorded repos gone, the recorded survivor vouching for only one. The unvouched repo has no ref, no bundle and no patch anywhere, so under the mutant its unpushed work is pointed at nothing after the reap. Pinned on both the total and partial loss arms, plus a full-coverage teeth case proving the assertion keys on TOTAL coverage rather than on multi-repo loss itself. Measured (mutation_gate_r3.sh, source restored + `git diff --quiet` verified after every arm): P1 `missing & vouched or missing <= vouched` -> 2 failed, 1 passed P2 `missing & vouched` -> 2 failed, 1 passed P0 pristine -> 3 passed Four survivor files: 89 passed. Wider surface (7 files): 154 passed. ruff clean. Zero lines of kanban_survivor.py changed — behaviour untouched. (cherry picked from commit 7c3e0b6)

The carded defect (t_e41d3dd4) is ALREADY FIXED: #842 r2 (a61dda5) added `carried_bundles` and gated the ref branch on `not bundles`. Both shapes the card measured -- bundle-only partial loss and mixed partial loss -- were driven through `preserve(cleanup=True)` on pristine #845 head and both already carry the bundle and record `kind: "bundle"`. No behaviour change is needed. What IS missing is a gate. `carried` and `carried_bundles` are independent collections read from the same recorded survivor, and every existing partial-loss test loses a repository vouched for by exactly ONE of them. So a mutant that keeps the bundle carry-forward but disables it whenever a ref is also carried -- `carried_bundles = [] if carried else [...]` -- leaves all 30 tests on this file GREEN while the MIXED shape drops the bundle and relabels the record `kind: "ref"`, telling an operator the work is pushed when its only copy is an orphaned bundle attachment. Measured on this tree, pre-existing suite only: M1 `carried_bundles = []` -> 3 failed / 28 passed (already gated) M2 `[] if carried else [...]` -> 0 failed / 30 passed SURVIVES With this test: M1 3 failed, M2 1 failed (killed only by the new test). Test-only: zero lines of hermes_cli/kanban_survivor.py; impl byte-identical to the pristine base. Four survivor files at this head: 95 passed before, 96 after. (cherry picked from commit 1e6339b)

Kyzcreig · 2026-09-23T10:42:45Z

apollo/merge-pass 2026-09-23: rebased onto current main (clean, own commits only, no content change) so CI runs against the post-#915 main; landing via fleet-merge on the FR-pause lane once green.

Kyzcreig · 2026-09-23T10:54:54Z

🤖 merged-by: apollo · lane: fr-pause-0922 · gate: BYPASS: FR PAUSED by Ace ruling 2026-09-22 (state/fleetreview-pause-20260922.md); t_e41d3dd4 approved · why: argus APPROVED (card done); rebased clean onto main; FR paused; CI green; Apollo merge pass 2026-09-22

Kyzcreig force-pushed the daedalus/t_e41d3dd4-mixed-carry-pin branch from 1e6339b to f14afd1 Compare September 22, 2026 17:33

Kyzcreig changed the base branch from survivor/t_1bcd8aec-cleanup-coverage to main September 22, 2026 17:33

Kyzcreig mentioned this pull request Sep 22, 2026

fix(kanban): close the 4 P1 FleetReview recorded on #848's head 0757be86 #886

Merged

daedalus-opus and others added 4 commits September 23, 2026 03:40

Kyzcreig force-pushed the daedalus/t_e41d3dd4-mixed-carry-pin branch from f14afd1 to 7dbd9ed Compare September 23, 2026 10:42

Kyzcreig added this pull request to the merge queue Sep 23, 2026

Merged via the queue into main with commit d4ee4fb Sep 23, 2026
54 checks passed

Kyzcreig deleted the daedalus/t_e41d3dd4-mixed-carry-pin branch September 23, 2026 11:13

Kyzcreig added the fleetreview:post-merge Ask FleetReview to review this MERGED pull (merge commit vs first parent) label Sep 24, 2026

Navigation Menu

Repository navigation

test(kanban): pin that carrying a ref does not cost the bundle beside it - #866

test(kanban): pin that carrying a ref does not cost the bundle beside it#866
Kyzcreig merged 4 commits into
mainfrom
daedalus/t_e41d3dd4-mixed-carry-pin

Kyzcreig commented Sep 22, 2026 •

edited by blacksmith-sh Bot

Loading

Uh oh!

Kyzcreig commented Sep 22, 2026

Uh oh!

Kyzcreig commented Sep 23, 2026

Uh oh!

Kyzcreig commented Sep 23, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

Kyzcreig commented Sep 22, 2026 • edited by blacksmith-sh Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

The carded defect is already fixed — this closes the GATE, not the bug

What IS missing: the MIXED shape has no gate

Measured (my run, isolated --basetemp, four survivor files)

Notes

Uh oh!

Kyzcreig commented Sep 22, 2026

FleetReview

Findings

Uh oh!

Kyzcreig commented Sep 23, 2026

Uh oh!

Kyzcreig commented Sep 23, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Kyzcreig commented Sep 22, 2026 •

edited by blacksmith-sh Bot

Loading

Measured (my run, isolated `--basetemp`, four survivor files)