fix(kanban-home-cards): prewarm probe, cached lineage, mtime board filter, stage-attributed timeout (t_11f2cf60) - #988
Conversation
e059f3e to
e683cdb
Compare
|
🤖 merged-by: apollo · lane: discord · gate: BYPASS: FR paused by Ace ruling 09-22; Ace directed landing without Argus 09-24 12:45; Argus r1 findings addressed per the r2 summary and covered by the 89 tests · why: t_11f2cf60 r2 (Ace: skip Argus reviews, 09-24 12:45): kanban-home-cards first turn reads one write-path home index instead of fanning out 78 board DBs (home_ids cold 4-15 s vs 750 ms budget). Apollo gate: CI 16/16 python slices green; local at e683cdb: PR test files 89/89 + kanban-home selection 97/97 |
|
🤖 merged-by: apollo · lane: discord · gate: BYPASS: FR paused by Ace ruling 09-22; Ace directed landing without Argus 09-24 12:45 · why: t_11f2cf60 r3 @ 3ece02d (Ace: no Argus, 09-24 12:45/14:07): kanban-home-cards first turn reads one write-path home index, lineage+dedupe on background threads (no cold state.db read gates the turn), delete_task ghost-row fixed. Gate: CI 39/0 green; local 103/103 (home_index, lineage, plugin, #951 AST contract) |
|
🤖 merged-by: apollo · lane: claude-bridge-session · gate: BYPASS: FR PAUSED by Ace ruling 2026-09-22 · why: re-arm after merge-queue ejection by the test_kanban_progress_stall load flake on the local pool (fix #1040); PR untouched |
|
🤖 merged-by: apollo · lane: claude-bridge-session · gate: BYPASS: FR PAUSED by Ace ruling 2026-09-22 · why: re-arm (auto-merge disable+enable) after merge-queue ejection by the test_kanban_progress_stall load flake on the local pool; fix is #1040; PR untouched |
|
🤖 merged-by: apollo · lane: claude-bridge-session · gate: BYPASS: FR PAUSED by Ace ruling 2026-09-22 · why: re-arm after fleet/attribution required-check app_id drift (15368 -> -1 restored per docs/ci/attribution-gate.md); PR untouched |
0756e19 to
318c731
Compare
|
🤖 merged-by: apollo · lane: discord · gate: BYPASS: FR paused by Ace ruling 09-22; slice card under review_policy=milestone_only · why: t_11f2cf60: kanban-home-cards r4 rebased onto main after MQ conflict ejection (add/add in test_kanban_home_lineage.py, both kept). Local 104/104 PR tests. Slice card: CI is the gate |
|
🤖 merged-by: apollo · lane: claude-bridge-session · gate: BYPASS: FR PAUSED by Ace ruling 2026-09-22 (state/fleetreview-pause-20260922.md); Apollo-reviewed lands via bypass · why: re-arm for the overnight-drain closeout: PR CI green on current head, 0 conflicts; #988 lost its per-SHA fleet/attribution status on the 02:38Z rebase push, #1012 was ejected by a group whose other member was red; Apollo self-review, FR paused |
|
🤖 merged-by: apollo · lane: boil-ocean · gate: BYPASS: FR PAUSED by Ace ruling 2026-09-22 (state/fleetreview-pause-20260922.md); Apollo-reviewed lands via bypass · why: fix(kanban-home-cards): prewarm probe, cached lineage, mtime board filter (t_11f2cf60); CI 39 green, review_policy=none => land on CI green — card done, approved-but-unlanded; landed in the 09-25 boil-ocean pass |
|
DEQUEUED by Apollo 07:2x PT: this PR breaks main's tests/plugins/test_kanban_home_facet.py::test_viewer_home_uses_session_lineage_not_exact_id (from #1014, home-session facet) — it failed in every merge group containing #988 (r-1012 group, r-1043 group), ejecting siblings and costing ~45 min per cycle. Reconcile kanban_home_index.py lineage lookup with the facet's session-lineage expectation, push, then re-enqueue via fleet-merge.sh. Card t_11f2cf60. |
…lter, stage-attributed timeout (t_11f2cf60) - kanban_db.home_ids: per-process TTL/LRU cache (keyed sid+path+inode); fail-open answers not cached. New home_lineage() also returns the home's earliest started_at. - plugin: home_ids now delegates to the shared lineage helper. - plugin: probe (lineage -> dedupe -> boards) runs on a daemon thread, started at pre_gateway_dispatch from the in-process session store (peek_session_id, no DB) so it overlaps auth/agent build; pre_llm_call waits at most BUDGET_S. - boards whose kanban.db predates the home start (minus 300s slack) with an empty WAL are not opened; root always scanned. - BUDGET_S 0.75 -> 1.5; timeout log names stage=lineage|dedupe|boards, ms, prewarmed; success log carries per-stage ms and mtime_skipped. Tests: tests/plugins/kanban_home_cards + tests/hermes_cli/test_kanban_home_lineage.py 84 passed (nice -n 19, load ~88).
… boards on the turn (t_11f2cf60 r2) Apollo ruling on Argus r1 (B1/B2/B3): the first turn no longer opens ~78 board DBs. hermes_cli/kanban_home_index.py keeps a single cross-board index (<kanban_home>/kanban/home_index.db: board, task_id, session_id, status, title, last_comment, last_activity, ev) maintained from kanban_db.write_txn after every outer COMMIT -- the one boundary guarded mutators and execution-lane writers share. Event-driven (per-board task_events watermark, per-row ev so older snapshots never win), best effort, never fails a board write. `hermes kanban home-index` backfills; `--check` reports drift (exit 1) for the daily cron. Plugin: lineage (cached home_ids) -> state.db dedupe -> ONE indexed read. Missing/unbackfilled/unreadable index renders nothing with one log line; boards are never scanned. Removed the mtime filter, pre_gateway_dispatch prewarm and _BudgetExceeded. Budget back to 750 ms. Every first turn logs one line incl. dedupe (history|state.db) and per-stage ms. Verified: tests/plugins/kanban_home_cards + test_kanban_home_lineage + test_kanban_home_index: 89 passed. Live read-only (scratch index built from the 77 live boards, load 79-80, no head start): 8/8 sessions match an independent raw-sqlite oracle, hook 18-130 ms.
…onverge event-less writers (t_11f2cf60 r3) B1: lineage + dedupe run on background threads with STATE_WAIT_S=0.35; past that the home is id + in-process parent_session_id + the lineage persisted in the home index (homes table, written off the turn path after home_ids resolves). Subset of truth: may under-count, never foreign. Dedupe not answered in time -> inject. Log names lineage=state.db|index| in-process and dedupe=clean|pending. B2: sync_after_commit fingerprints the board's stamped (id, session_id, status, title) set on every commit and diffs it against the index, so delete_task / delete_archived_task (events deleted), transfer scrub, swarm, and raw dashboard UPDATEs converge at their commit or the next write_txn commit on that board. Test per path. Verified: focused suite 99 passed; the 5 new sweep tests fail on r2's index code; live read-only probe 4/4 vs oracle, cold-lineage session rendered in 360 ms via in-process fallback.
…o staged state.db gate (t_15d21849) (#1036) - BUDGET_S 0.75 -> 0.22; STATE_WAIT_S removed. Lineage, dedupe and a fallback-home index read start together; the turn wakes on each completion and at the ceiling takes exact > fallback cards, else skips. - Wait sliced into 2 ms: the loaded Studio oversleeps a single 0.25 s timed wait by ~140 ms (measured 383-397 ms); sliced 251-265 ms. - Tests: 77-board fleet fixture (warm, cold state.db, locked index, 27 concurrent first turns), all under the 250 ms ceiling. Bench on a copy of the live 77 boards under load: main p50 726/p95 1293 ms (27/27 over 250); this head p50 21/p95 214, cold state.db p95 248 ms, 0 over. Co-authored-by: Kyzcreig <9063726+Kyzcreig@users.noreply.github.com>
|
🤖 merged-by: apollo · lane: kanban-merge-pass · gate: BYPASS: FleetReview paused by Ace 2026-09-22 (state/fleetreview-pause marker present) · why: t_9bd6852d: hermes-agent CI load-sensitive tests, part 2 (after t_6ee81e82 / PR #1062): boot; Argus off card review (Ace 13:08), CI green |
…at (t_11f2cf60) r1 added started_at to every lineage SELECT; a schema lacking it (the #1014 dashboard facet fixture, older state.db) raised and home_ids failed open to the exact id, breaking test_viewer_home_uses_session_lineage_not_exact_id. Select started_at only when the column exists; start is None otherwise. Verified: test_kanban_home_facet + home_lineage + home_index 47 passed, kanban_home_cards 66 passed; new regression test fails without the fix.
318c731 to
674fb8c
Compare
|
🤖 merged-by: apollo · lane: kanban-merge-pass · gate: BYPASS: FleetReview paused by Ace 2026-09-22 (state/fleetreview-pause marker present) · why: t_11f2cf60: kanban-home-cards: fires but times out on the live Studio (home_ids cold 4–15 s ; worker completed in place, CI green |
|
🤖 merged-by: apollo · lane: discord · gate: BYPASS: FR paused by Ace ruling 09-22; slice card under review_policy=milestone_only; gate = CI green on the current head · why: t_11f2cf60: kanban-home-cards r4 (674fb8c); CI green (58 checks) + local 113/113. Ace 09-25 10:32: finish the session; MQ 27 deep, 0 merges to main in 4h, GitHub ETA 2.5h+ with cascade ejections |
Kyzcreig
added
the
fleetreview:post-merge
FleetReviewReview: post-merge · head Post-merge review ( Reviewed with 1 of 3 model families — anthropic, xai unavailable. Confidence: 2/5 Findings
FleetReview provenance · models: C=claude-code-opus-5, D=grok-4.6 · cost: $23.60 · duration: 1h 23m 07s · rounds: 3 · files examined: 8 |
1 participant
Card t_11f2cf60. Live: plugin fired but logged
unavailable=timeout partial_cards=0on every resumed session (750 ms budget).Changes
kanban_db.home_ids: per-process TTL(300s)/LRU(256) cache keyed (sid, path, inode); fail-open answers not cached.home_lineage()also returns the home's earlieststarted_at.home_idsdelegates to the shared lineage helper (was exact-id).pre_gateway_dispatchfrom the in-process store (peek_session_id, no DB);pre_llm_callwaits <= BUDGET_S.stage=lineage|dedupe|boards partial_cards ms prewarmed; success log: per-stage ms + mtime_skipped.Verified
Known limits
Need help on this PR? Tag
@codesmith-botwith what you need. Autofix is disabled.