Repository navigation
CI EFFICIENCY MANAGER (DEDICATED, operator-directive 2026-05-29): end-to-end audit of .github/workflows/ci.yml + every script it invokes, drive CI toward seconds-scale execution on unchanged subgraphs. Operator quote: 'we need to get CI down to a very minimal/efficient fast state (i'm thinking like, - #3883
briansrls wants to merge 1 commit into
Conversation
|
Closing as abandoned per operator course correction (2026-05-29, relayed via PM): stop micro-tuning legacy YAML/shell paths; CI efficiency work pivots to proper This draft only patched Superseded by:
Branch — sent from clever-cat-115 |
briansrls
left a comment
There was a problem hiding this comment.
Review metadata
- Provider / model:
codex/unknown - Commit:
933089d2· Trigger:schedule - Thinking:
260s wall
BLOCKING (1)
Root Cause
dsl/gunbc/ci_github_actions_workflow.daggenerated mirror stayed stale at the M1 probe step and lacks the new timeout fail-closed step → regenerate it from .github/workflows/ci.yml.
| - name: M1 v4 full-tree rust emit probe (v2 emit + cargo check) | ||
| if: needs.affected.outputs.v4 == 'true' || needs.affected.outputs.workflow_policy == 'true' | ||
| id: m1_rust_emit_probe | ||
| if: needs.affected.outputs.v4 == 'true' |
There was a problem hiding this comment.
BLOCKING: INVARIANTS.md P2 says the committed generated CI model must track ci.yml, but this v4-only M1 gate change leaves dsl/gunbc/ci_github_actions_workflow.dag still declaring the old v4-or-workflow_policy condition.
…(2026-05-29) (#3885) * docs(audit): CI anatomy, profiling, redundancy, and test corpus rationalization Canonical anchor for the CI-efficiency lane (operator 2026-05-29): workflow anatomy, run 26615772294 profiling, redundancy/overlap tables, M1 fail-open bug, dependency-modeled fixes, and test consolidation candidates (§8). Co-authored-by: Cursor <cursoragent@cursor.com> * docs(audit): reframe §8 — affected-set skips overlap, not delete duplicates Operator clarification: deduplication via declared input subgraph + IRT frontier; T-IG gap table; overlap → Table B; delete only scaffold/DEAD. Co-authored-by: Cursor <cursoragent@cursor.com> * docs(audit): fix §8.6 IRT-1 vs IRT-4 cache-key semantics (TASKS.md) Verdict reuse keys content_hash(whole TestClaim node), not input_subgraph alone. Split frontier selection from IRT-4 cache in north star, Table B, §8.3, and §8.6 formula block. Co-authored-by: Cursor <cursoragent@cursor.com> * docs(audit): align §7 with no-bridging design (#3886) Replace micro-P0 YAML plan with atoms A0–A14; drop bucket-tuning and superseded PR refs; point to interpreter-direct S2′. Co-authored-by: Cursor <cursoragent@cursor.com> * docs(audit): reconcile B01/B03 fix status with #3886 atoms Drop #3883 as implementer target (closed/superseded); point M1 and bucket-gating fixes at #3886 A1/A9 per §7. Addresses review #20367. Co-authored-by: Cursor <cursoragent@cursor.com> * docs(audit): fix T-IG-04 owner — #3886 not #3853 for frontier wiring #3853 stays witness/detect-bin only per §7; ci_select frontier work belongs to #3886 A1/A2 + T-38. Addresses openai-pro review #20386. Co-authored-by: Cursor <cursoragent@cursor.com> * docs(audit): §5.1 srv2 jobserver infra compounding for M1 timeouts Separate operator infra lane; modeled fixes (A9/A1) remain required. Co-authored-by: Cursor <cursoragent@cursor.com> * docs(audit): reword Table B R10 away from GHA if: bridge Replace "frontier → if:" with interpreter-frontier language so the audit cannot be read as endorsing bucket→YAML scheduling. #20401. Co-authored-by: Cursor <cursoragent@cursor.com> * docs(audit): split Table B IRT-1 schedule vs IRT-4 cache keys Per §8.6 and codex/inline review: selection (frontier) must not sit in the verdict-identity column; cache keys are content_hash(whole node). Co-authored-by: Cursor <cursoragent@cursor.com> * docs(audit): R07 M1 cache key includes toolchain/resources IRT-4/P2: verdict key must fold cargo-check toolchain facts, not only src/v4 merkle + v2 binary. Addresses inline review on #3885. Co-authored-by: Cursor <cursoragent@cursor.com> * docs(audit): Table B whole-node IRT-4 keys for all CiCommand arms Verdict column uses content_hash(whole node) for non-TestClaim commands; schedule column holds frontier only. §8.6 extends IRT-4 to CiCommand. Addresses codex review on cb21f37. Co-authored-by: Cursor <cursoragent@cursor.com> --------- Co-authored-by: Cursor <cursoragent@cursor.com>
Auto-opened by session-dashboard for session
clever-cat-115.Pushing to
session/clever-cat-115advances this PR.Worker attestation
Before flipping this PR to ready for review, confirm each item:
npm test,cargo test) and the result.Closes #Ndirective.Summary
TODO: replace this paragraph with one or two sentences naming the change and its motivation. Reviewers read this first.
Test plan