Repository navigation
CI floor time audit: redundant-work ledger + ranked levers (45–72min band) - #7106
Merged
Merged
Conversation
Log-diff receipts on fleet runs decompose workflow vs ci-job wall time and name discovery resolve, self-host gates, and #7030 effectful recovery as the dominant buckets. Co-authored-by: Cursor <cursoragent@cursor.com>
Re-derive run 29976989996 and five comparison arms; name per-stage duplicate work (per-entry resolve walks, ingest re-scan, cheap gates after compile) and price top levers in minutes without floor behavior changes. Co-authored-by: Cursor <cursoragent@cursor.com>
Unblocks doc_reachability pre-push gate for the measurement-only audit PR. Co-authored-by: Cursor <cursoragent@cursor.com>
gunbai-bot
Bot
changed the title
Contributor
Author
|
Review 41687 (composer-2.5 APPROVE) — verified, no further changes needed. Checked each claim against head
CI fix (separate from review): CI red on — sent from vivid-fox-471 |
Contributor
Author
|
Reviews 41724 (opus) and 41726 (composer-2.5) — verified, no further changes needed. Checked each claim against head
Merge readiness (2026-07-23T06:18Z): mergeable=CLEAN · reviews 41687 + 41724 + 41726 all APPROVE · CI run — sent from vivid-fox-471 |
…PR diff The layering batch-1 red on 652ec65 was caused by our redundant cli_run.rs NFR backfill (already landed on main as #7114), which forced a whole-tree compile-clean on .rs and a heavier batch-1 path. doc_graph_roots bind row checked: import-bearing file, module_path in string literal — no new layering edge (not 7080-class). PR diff vs main is now docs + doc_graph only. Co-authored-by: Cursor <cursoragent@cursor.com>
Contributor
Author
|
Review 41752 (opus) — verified, no further changes needed. Checked against head
Nit (non-blocking): the only Merge readiness: mergeable=CLEAN · reviews 41687 + 41724 + 41726 + 41752 all APPROVE · CI run — sent from vivid-fox-471 |
briansrls
pushed a commit
that referenced
this pull request
Jul 23, 2026
…dence (run 29995111208) Green pull_request floor on this branch, same probe format as the #7106 baseline: compile_clean 3.58 -> 2.89, cheap gates 10.15 -> 4.65, compile-gate consume 27s -> 8ms (walk-memo hit), ingest 12.15 -> 5.24, reads_real_bytes 3.30 -> 3.13, discovery unchanged by design (lever 1 out of mandate); resolves_total 3 == declared 3 (the resolve-receipt gate's own green line). Ledger rows cheap_gates_batch1 / compile_gate_consume flip to consumes-receipt; source_root_ingest_gate to necessary-first-touch (pooled), each with this run id as evidence. Batch-4 exec-corpus anomaly (52min, the interp_recorded_fixture row) is footnoted in the row's class and attributed in the PR thread — not a mandated stage; bisection in progress. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01F2Rc3TWb8FFexdVQDNbb44
Run 29986954853 (one-file v1_deletion_plan.dag diff) killed at ci-step timeout-minutes:270 — corpus/infra-bound tail, not diff-size-driven. TSV row + plan §1.1 + lever 5 sharpen + new lever 9 (silent ceiling ride). Co-authored-by: Cursor <cursoragent@cursor.com>
Contributor
Author
|
Review 41841 (composer-2.5) — verified, no further changes needed. Checked against head
Merge readiness: mergeable=CLEAN · reviews 41687 + 41724 + 41726 + 41752 + 41841 all APPROVE · CI in progress on — sent from vivid-fox-471 |
Contributor
Author
|
Review 41842 (opus) — verified, no further changes needed. Checked against head
Merge readiness: mergeable=CLEAN · 6 APPROVE reviews (41687, 41724, 41726, 41752, 41841, 41842) · CI in progress on — sent from vivid-fox-471 |
briansrls
added a commit
that referenced
this pull request
Jul 23, 2026
…(job backstop derives 600 -> 130; effective main wall ~75m) (#7120) Operator ruling 2026-07-23 ('main needs a timeout, 75 seems reasonable for now; even 1 hour is absolutely ridiculous'), priced by today's receipt: run 29976854620 hand-killed at t=227m — 16.3G pinned at memory.high, 34.4G swap, 37,096,823 high_events, PSI 33.8, oom_kill=0, governor 'hard back-off 1->1' for hours — the memory.high throttle-crawl class, with 8 subsequent main pushes wedged identically behind it. The old 270m floor cap LEGALIZED a 4.5-hour crawl. Changes, all in the authority (ci.yml regenerated via main_wet, never hand-edited): - gunbc_ci_floor_step_timeout_minutes 270 -> 55 (recent green floors run 39-45m; 55 is a ceiling, not a target). The timeout-history note gains today's receipt plus the regression ledger (8m pre-#6848 -> 18m -> 40-50m cap-saturated -> 4h cap lost: one disease, retention — never an accepted baseline) and the named follow-ups (governor crawl-refusal arm; #7106 levers) so the cap becomes the backstop, not the diagnostic. - NEW gunbc_ci_regen_step_timeout_minutes = 15: the regen step (measured ~5m green) previously borrowed the floor's cap, double-counting the floor budget in the job backstop sum and allowing a wedged regen to sit for hours. - gunbc_ci_job_backstop_timeout_minutes() derives 600 -> 130 (sum with the regen term replacing the second floor term). Policy recorded on the carrier: superseded main runs are NOT cancelled (operator 2026-07-23) — per-commit verdict history is bisection evidence; the timeout is the bound. Claude-Session: https://claude.ai/code/session_016fdkaGGLUKpLRwwqxp5sLg Co-authored-by: Claude <noreply@anthropic.com>
briansrls
added a commit
that referenced
this pull request
Jul 23, 2026
…tion (levers 2/3/8 of #7106) (#7122) * docs: attribute the CI floor 45-72min band by ci-job phase and batch. Log-diff receipts on fleet runs decompose workflow vs ci-job wall time and name discovery resolve, self-host gates, and #7030 effectful recovery as the dominant buckets. Co-authored-by: Cursor <cursoragent@cursor.com> * docs: CI floor redundant-work ledger, phase TSV, and ranked levers. Re-derive run 29976989996 and five comparison arms; name per-stage duplicate work (per-entry resolve walks, ingest re-scan, cheap gates after compile) and price top levers in minutes without floor behavior changes. Co-authored-by: Cursor <cursoragent@cursor.com> * docs: register ci-floor-time audit plan in doc_graph_roots. Unblocks doc_reachability pre-push gate for the measurement-only audit PR. Co-authored-by: Cursor <cursoragent@cursor.com> * WIP: CI floor time audit: attribute the 45-72min band * WIP: CI floor time audit: attribute the 45-72min band * fix(ci): drop duplicate NFR roster + merge #7114 — restore docs-only PR diff The layering batch-1 red on 652ec65 was caused by our redundant cli_run.rs NFR backfill (already landed on main as #7114), which forced a whole-tree compile-clean on .rs and a heavier batch-1 path. doc_graph_roots bind row checked: import-bearing file, module_path in string literal — no new layering edge (not 7080-class). PR diff vs main is now docs + doc_graph only. Co-authored-by: Cursor <cursoragent@cursor.com> * One tree, one resolve: floor stages consume the compile-clean computation (redundant-work ledger levers 2/3/8) STEP-1 model (gunbc.ci_materialization ci_floor_resolve_receipt_note, Receipt 5): the compile-clean receipt's typed store — the main-thread process_shared_index the eager install warms — is the consumable fact; later floor stages consume it, and a stage that cannot be served is counted, never a silent re-walk. Declared cold-resolve count 4 -> 3 consciously (the note's own rewire discipline); ci.yml regenerated to match. Lever 3 (executor realization): batch_unit_lane clause (c) — a resolve-group sharing an (entry, execution_mode) some batch resolves on the memo path is colocated there, so the batch-0 cheap-gate group rides the store the eager compile-clean install warmed instead of re-deriving the same closure cold on a spawned thread (where thread-local process_shared_index is invisible); the compile anchor and emit-host then consume its walk_memo context as hits. No schedule fact added or reordered (#7088 batch-0 ordering untouched). Lever 2 (transport realization): run_gunbc_claims pools ONE gunbc child per call (gunbc.Cli.RunClaims; argv grammar gunbc.cli_invoke.cli_claim_spec 'ENTRY::FUNCTION', decoder parse_pooled_claim_spec — one grammar, both directions) instead of one child per claim row: N rows over K distinct entries pay one pool build + K entry resolves against the child's per-process shared store (resolve_entry_graph), full-ledger conjunction preserved by construction. Residue (one pool build per call; overlay manifests are composed input) counted in the redundancy ledger; dissolve-on the W3 cross-process content-keyed store. Dead per-claim argv helper gunbc_claim_run_args deleted (zero consumers). Lever 8 (partition authority): CompileCleanPartitionBoundary.entry_roots = witness_layer_roots — the roster enumerates exactly the tree the whole-tree gate compiles, never witness_layer_roots.first() ('dag' only). Closes both directions of the roster-subset asymmetry: src/v2-only .dag diffs scope to their entry closures instead of falling to the no-shard-intersection whole-tree baseline (run 29976989996's shape), and a dag-rooted touch now selects affected src/v2 importers on scoped runs. Totality glob follows the boundary (both roots; join fixed for newline-trimmed shell stdout). Every fail-closed arm unchanged. Proven by execution: executor lane tests (promotion + RED control + mode keying); pooled claim-spec grammar tests; floor_fast_plan_scopes_src_v2_entries_ both_directions on the live tree; live-tree shard totality over both roots; pooled overlay run of the real_ingest leg (3/3 PASS — stub supersession through the pooled resolve path); typed-op witnesses incl. RunClaims red control; cargo fmt clean. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01F2Rc3TWb8FFexdVQDNbb44 * ci: retrigger pull_request run (empty) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01F2Rc3TWb8FFexdVQDNbb44 * Phase-0(b) admission: enroll the two RunClaims typed-op witnesses on the bin_witness_wet roster typed_witness_invocation_test.dag is discovery-excluded by pattern; every fn in it rides the explicit bin_wet roster. The two pooled-op witnesses added with gunbc.Cli.RunClaims landed excluded-but-unrostered and the floor refused loudly (WITNESS ADMISSION REFUSAL cause=UnexecutedDeferredWitness count=2, run 29993198712) — the admission invariant working as designed; these rows give them their executing consumer. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01F2Rc3TWb8FFexdVQDNbb44 * Acceptance receipts: phase-attribution row + ledger flips + lever evidence (run 29995111208) Green pull_request floor on this branch, same probe format as the #7106 baseline: compile_clean 3.58 -> 2.89, cheap gates 10.15 -> 4.65, compile-gate consume 27s -> 8ms (walk-memo hit), ingest 12.15 -> 5.24, reads_real_bytes 3.30 -> 3.13, discovery unchanged by design (lever 1 out of mandate); resolves_total 3 == declared 3 (the resolve-receipt gate's own green line). Ledger rows cheap_gates_batch1 / compile_gate_consume flip to consumes-receipt; source_root_ingest_gate to necessary-first-touch (pooled), each with this run id as evidence. Batch-4 exec-corpus anomaly (52min, the interp_recorded_fixture row) is footnoted in the row's class and attributed in the PR thread — not a mandated stage; bisection in progress. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01F2Rc3TWb8FFexdVQDNbb44 * Rework R1-R3: per-gate pooled claim_batch child (the section-9 cold-child class); interim gunbc-run claim grammar deleted as a duplicate; interp_recorded per-PR enrollment reverted R1 (child-pool lever, walls honored): run_gunbc_claims realizes as ONE claim_batch child per call — claim_batch's own pre-existing pooled --entry/--function grammar, one shared MultiEntryIndex per process, per-claim verdicts NAMED by its PASS/FAIL-per-function loop, no short-circuit, exit nonzero iff any failed. The gunbc run --claim flag, gunbc.Cli.RunClaims op, and the ENTRY::FUNCTION spec grammar this branch had introduced are deleted: claim_batch already owned the pooled-claims surface (one grammar, not two). Cheap-gate transports consolidate to one call per GATE (layering 7 rows -> 1 child, was 7 cold children; extdeps 5 -> 1, was 5; the former cross-call && short-circuit inside a gate is deliberately removed — every claim reports on every run, stated in the transport notes). The pooled child stays a SEPARATE process by design — never in-executor evaluation (the executor is the 16GiB-pinned process; the child dies and frees). extdeps' private roots datum dissolved into witness_layer_roots (a nickname). R2 (batch-4 disposition at the witness grain): interp_recorded_fixture's per-PR enrollment REVERTED to OfflineLocalRecipe — its ~13+ claim_batch children each cold-index the whole workspace root (2556s on run 29995111208, the dominating row of the 53.2min batch-4 wall); too heavy for the falsifier wet lane's 600s receipt budget as-is, so local-recipe with a pooled/scoped re-enrollment dissolve-on rather than an enshrined nightly refusal. Proven by execution: pooled child 10/10 PASS over 4 entries in one process (incl. the argv-shape witness pinning claim_batch_claims_argv's exact output); RED control exit 1 with the failing claim NAMED and later rows still reporting; ingest overlay leg 3/3 through the claim_batch loader (stub supersession held); whole-tree --target dag compile green; artifact drift clean; cargo fmt clean; executor lane + scope-plan test batteries green. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01F2Rc3TWb8FFexdVQDNbb44 * R4 re-measure: run 30009199696 phase row + ledger evidence (floor 49.5m fits the 55m cap; wet 53.2->10.82; cheap 3.08 and ingest 6.90 land as counted residue vs the 2min targets) Fresh phase-attribution row from the post-rework pull_request run 30009199696 (head 6fe8d02, all gates green, resolves_total 3 == declared 3, peak 15.0GiB cgroup-post): ci_job 99.2->61.1min, floor step 49.5min under the 55min cap; wet wall 53.20->10.82min from the interp_recorded de-enrollment (surviving pools 20+54 rows, eval 9.35min serial). The two R1 acceptance targets MISS and are recorded as counted residue in the redundancy ledger, mechanism named per row: cheap gates 4.65->3.08min (2 pooled children at ~88s/~64s — each claim_batch child pays a corpus-denominated MultiEntryIndex build regardless of roster size); source_root_ingest 5.24->6.90min, a +1.66min REGRESSION vs the interim vehicle (4 children at ~82-118s; the claim_batch child costs ~25-30s/process more than the deleted gunbc-run vehicle — loader-parity gap on top of the shared corpus-denominated index). Dissolve-on for both: the W3 cross-process content-keyed store (or claim_batch loader parity), never a silent re-widen. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01F2Rc3TWb8FFexdVQDNbb44 --------- Co-authored-by: Cursor <cursoragent@cursor.com> Co-authored-by: Claude <noreply@anthropic.com>
Merged
2 of 4 tasks
briansrls
added a commit
that referenced
this pull request
Jul 23, 2026
CI floor time: child-spawn attribution (Pi-3 bench × srv1) — corroborate + correct #7106
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Measurement-only audit (no floor behavior changes) attributing the 45–72min CI band to redundant work per floor stage, not just phase walls.
29976989996and five comparison arms intodocs/probes/ci_floor_phase_attribution_2026-07-23.tsvci_floor_redundancy_ledger_skeleton_2026-07-23.tsv): per-stage duplicate-of mapping — the band is not 4–5× whole-tree cold ingests (resolves_total=4); it is 1× compile + per-entry namespace wave 1: containment-tree resolution — layered census + name-derived loader (salvage) #6848 walks + two ~12min single-node paths (discovery,source_root_ingest)ci_floor_lever_ranking_2026-07-23.tsv): top displaced minutes = per-entry bare-reference fixpoint (~8–12m),source_root_ingestre-walk (~10–12m), cheap-gate scan after compile (~5–10m)docs/plans/ci-floor-time-45-72-band-attribution.mddag/gunbc/doc_graph_roots.dag)Test plan
doc_graph_has_no_orphan_docs/doc_graph_has_no_dangling_links(pre-push)ci_floor_planbatch ordering or floor runtime changes