T-24 Wave-0: ci.dag affected-set authority — dissolve detect-affected-components.sh (ci.yml emit deferred) by briansrls · Pull Request #3853 · gunb-ai/gunbc

briansrls · 2026-05-28T23:40:39Z

Summary

T-24 Wave-0: dissolve scripts/detect-affected-components.sh and make src/v4/workflow/ci.dag the single authority for CI component path buckets, git-diff fail-closed behavior, and self-hosted M1 cargo-check parallelism.

Deleted scripts/detect-affected-components.sh; added detect-ci-affected-components host transport plus Rust mirrors (v4_ci_component_affected.rs, v4_ci_runner_pool.rs) with explicit dissolution triggers.
Extended ci.dag: CiComponentAffected, CiGitDiffReadOutcome, workflow_policy predicates, runner pools, M1 probe wiring; fixed ReadFailClosed match pattern for v4 compile.
Aligned .github/workflows/ci.yml / dsl/gunbc/ci_github_actions_workflow.dag affected job and V4_M1_CARGO_CHECK_JOBS projection with modeled facts.
Ratchets: claim .dag receipts, smoke test cross-checks ci.dag workflow_policy prefixes against the Rust mirror.

Full YamlStatic emission of ci.yml from ci.dag remains deferred per src/v4/TASKS.md T-24 (committed YAML is still interim authority until emission slice lands).

Test plan

cargo test -p v3-compiler --test integration v4_workflow_ci_runner_dag_smoke_test -- --quiet
cargo test -p v3-compiler v4_ci_component_affected -- --quiet
cargo test -p v3-compiler --bin detect-ci-affected-components (unit tests in module)
CI: affected job uses detect-ci-affected-components; Lens-CI v4 compile over ci.dag entry root
CI green on HEAD (reviews re-running)

SG-0 / P5 receipt

SG-0 hand-path delta: +3

SG-0 pairing: (c) Net +3 on EXPECTED_HAND_AUTHORED_NON_TEST for T-24 host-transport shims (detect_ci_affected_components.rs, v4_ci_component_affected.rs, v4_ci_runner_pool.rs) with matching INVARIANTS §P5(b) rows. Follow-up dispatch: dissolve when ci.dag eval / generated CI runner owns affected-set and M1 parallelism without parallel Rust path logic (T-24 emission slice; tracked gunbc#3853).

…mponents

…v projection Co-authored-by: Cursor <cursoragent@cursor.com>

Add v4_ci_runner_pool to ci_changed_path_affects_workflow_policy so the detect-ci-affected-components host transport matches ci.dag authority. Register detect_ci_affected_components bin and v4_ci_* modules in EXPECTED_HAND_AUTHORED_NON_TEST and INVARIANTS §P5 non-test table. Co-authored-by: Cursor <cursoragent@cursor.com>

briansrls · 2026-05-29T00:01:20Z

Addressed both REQUEST_CHANGES findings in f39cf8cd9:

workflow_policy path parity — Added src/v3/compiler/src/v4_ci_runner_pool to ci_changed_path_affects_workflow_policy in v4_ci_component_affected.rs (matching ci.dag:675) and a workflow_policy_paths_match_ci_dag_authority unit test so the Rust host transport cannot drift from the .dag predicates.
P5 / SG-0 census — Registered src/v3/compiler/src/bin/detect_ci_affected_components.rs, src/v3/compiler/src/v4_ci_component_affected.rs, and src/v3/compiler/src/v4_ci_runner_pool.rs in EXPECTED_HAND_AUTHORED_NON_TEST with matching INVARIANTS §P5 non-test table rows (T-24 host-transport receipts). sg0_v3_hand_authored_census passes locally.

— sent from wise-otter-34

…mponents

briansrls

Review metadata

Provider / model: codex / unknown
Commit: e948ea5f · Trigger: schedule
Thinking: 233s wall

BLOCKING (2)

Root Cause

src/v4/workflow/ci.dag Git diff acquisition is modeled only as GitDiffNameOnly data with no failure carrier → add a typed read outcome/diagnostic path and make the host transport exit nonzero or mark all components affected on read failure.
src/v4/workflow/ci.dag Runner-pool arithmetic lacks cardinality/spread evidence at the API boundary → require a non-empty/non-uniform pool witness or totalize the fanout calculation before exposing ci_m1_probe_cargo_check_jobs_from_pools.

⚠️ The PR is aligned with the workflow-as-data direction, but these fail-closed and totality gaps should be fixed before merge.

briansrls · 2026-05-29T01:15:28Z

Sign in to view

+                "warning: git diff --name-only {range} exited {}; treating as no changes",
+                out.status
+            );
+            Vec::new()


Already addressed on HEAD (dc4e020): host transport now uses GitDiffRead::FailClosed and calls ci_component_affected_fail_closed() (all four flags true) when git diff exits nonzero or fails to spawn — it no longer returns an empty path set. The .dag side models the same contract via CiGitDiffReadOutcome / ci_component_affected_from_git_diff_read → ReadFailClosed → ci_component_affected_fail_closed().

— sent from wise-otter-34

briansrls · 2026-05-29T01:15:29Z

Sign in to view

+  let max_runners = ci_runner_pool_max_runner_count(pools: pools)
+  let total_runners = ci_runner_pool_total_runner_count(pools: pools)
+  let hosts = ci_runner_pool_host_count(pools: pools)
+  (total_runners - min_runners) * hosts / (max_runners - min_runners)


Already addressed on HEAD (dc4e020): fanout division is totalized via ci_runner_pool_runner_spread (returns 1 when max_runners == min_runners), and ci_m1_probe_cargo_check_jobs_from_pools divides by ci_int_at_least_one(ci_m1_probe_cargo_fanout_slots(...)). Fleet non-uniformity is receipted by ci_runner_pool_m1_probe_witness_holds + data ci_runner_pool_m1_probe_witness on the pinned srv1/srv2 pools; Rust mirror has matching runner_spread_fail_closed_when_max_equals_min unit test.

— sent from wise-otter-34

…mponents

CI failed on sg0-pr-body-append.3853.txt — the (c) pairing line cited gunbc#3853 but omitted "dispatch", which check-pr-sg0-net-shrink-discipline requires on the pairing line or its immediate continuation. Co-authored-by: Cursor <cursoragent@cursor.com>

…mponents

The Lens-CI v4 compile step failed because `ReadFailClosed { .. }` uses rest syntax the parser does not accept. Bind the Diagnostic payload with an explicit wildcard field instead. Co-authored-by: Cursor <cursoragent@cursor.com>

…mponents

briansrls · 2026-05-29T01:02:28Z

Re: fold init values (ci_runner_pool_min_*)

Verified in-tree: ci_runner_pool_min_jobserver_token_cap seeds with ci_srv2_pool.jobserver_token_cap (36, the fleet max cap) and ci_runner_pool_min_runner_count seeds with ci_srv1_pool.runner_count (20, the fleet min count). That is not a generic min over arbitrary lists — it only matches std::iter::min on the Rust mirror when pools == ci_self_hosted_runner_pools.

Intentional for this Wave-0 slice: the modeled fleet is pinned (ci_srv1_pool / ci_srv2_pool data rows + operator spec); m1_probe_cargo_check_jobs and the smoke test (m1_probe_cargo_check_jobs_derived_as_four, V4_M1_CARGO_CHECK_JOBS) ratchet the canonical answer (4). Generalizing fold seeds (+∞ / first-element init) or renaming to *_for_current_fleet is reasonable follow-up but out of scope for this PR — full YamlStatic emission / .dag eval host is the named dissolution path in the PR body dispatch line.

Re: parallel Rust mirrors (v4_ci_component_affected.rs, v4_ci_runner_pool.rs)

Agreed these are interim host projection, not a second authority: P5(b) rows + SG-0 +3 pairing are in this PR, and v4_workflow_ci_workflow_policy_prefixes_align_rust_mirror cross-checks workflow_policy prefix literals between ci.dag and the mirror source. Dissolution trigger is explicit in sg0-pr-body-append.3853.txt.

— sent from wise-otter-34

v4 compile rejects a line-leading `!=` token (Ne); fold the nonuniform-runner-count check onto one line for Lens-CI. Co-authored-by: Cursor <cursoragent@cursor.com>

…mponents

briansrls · 2026-05-29T01:22:19Z

Re: title vs Wave-0 scope (claude review nit)

Agreed — this PR is the Wave-0 slice named in src/v4/workflow/ci.dag (YamlStatic workflow projection deferred; full ci_pipeline → .github/workflows/ci.yml is a follow-on emission slice). Landed here: dissolve detect-affected-components.sh, path buckets + git-read fail-closed + runner-pool/M1 parallelism in ci.dag, and the three v3 host-transport shims with P5/SG-0 receipts. .github/workflows/ci.yml remains hand-authored but is updated only for the new cargo run affected job and V4_M1_CARGO_CHECK_JOBS projection.

Latest push also fixes v2 compile of ci.dag (line-leading *// operators — same class as the prior != witness fix).

— sent from wise-otter-34

briansrls

Review metadata

Provider / model: codex / unknown
Commit: e233966f · Trigger: schedule
Thinking: 295s wall

BLOCKING (2)

Root Cause

src/v4/workflow/ci.dag host read outcome was added with unsupported variant-payload syntax → declare payloads using existing record variant syntax, e.g. ReadOk { git_diff: GitDiffNameOnly } and ReadFailClosed { diagnostic: Diagnostic }.
src/v4/workflow/ci.dag runner-pool helpers expose arbitrary-list APIs while closing over fixed srv1/srv2 data → specialize the calculation to ci_self_hosted_runner_pools or require a non-empty pool carrier whose first element seeds the min folds.

⚠️ The direction is aligned with workflow-as-data, but these substrate issues need fixing before merge.

briansrls · 2026-05-29T01:35:10Z

Sign in to view

+// RULING-1: 🟢 pretty-much-done — host `git diff --name-only` read outcome (T-24 / INVARIANTS P3).
+// `ReadFailClosed` selects the superset `ci_component_affected_fail_closed` (all flags true).
+type CiGitDiffReadOutcome {
+  ReadOk: GitDiffNameOnly


Addressed on HEAD (b06f7665): CiGitDiffReadOutcome is now declared as a coproduct with named payloads — ReadOk { git_diff: GitDiffNameOnly } | ReadFailClosed { diagnostic: Diagnostic } — matching the RerunNodeSet pattern and the match in ci_component_affected_from_git_diff_read.

— sent from wise-otter-34

briansrls · 2026-05-29T01:35:11Z

Sign in to view

+
+
+fn ci_runner_pool_min_jobserver_token_cap(pools: List<SelfHostedRunnerPool>) -> Int {
+  fold(pools, init: ci_srv2_pool.jobserver_token_cap, f: fn(acc, pool) {


Addressed on HEAD (b06f7665): both min folds no longer seed from pinned ci_srv1_pool / ci_srv2_pool rows. They use structural neutral ci_int_min_fold_sentinel with an empty-list guard returning 0; ci_runner_pool_max_runner_count also returns 0 on empty input. Rust mirror updated to match (v4_ci_runner_pool.rs empty-pool unit test). Canonical fleet answer remains ratcheted via data m1_probe_cargo_check_jobs on ci_self_hosted_runner_pools + smoke test.

— sent from wise-otter-34

…mponents

Main landed incomplete WasmModelCoreLawKey/SwiftModelCoreLawKey return blocks (unclosed '('); GitHub PR merge replayed that blob and broke test_strict_deprose_dag.py. Restore merge-base tag entries and close the handlers matching TsModelCoreLawKey. Co-authored-by: Cursor <cursoragent@cursor.com>

Resolve strict_deprose_dag.py conflict by retaining the closed dissolve-on-arrival return block; main's incomplete handler broke test_strict_deprose_dag.py on PR merge. Co-authored-by: Cursor <cursoragent@cursor.com>

briansrls · 2026-05-29T08:23:15Z

Verified cursor APPROVE (review artifact 20526) against HEAD:

Wave-0 partial close honestly scoped; ci.dag authority, shell dissolved, P5(b) mirrors with parity ratchets.
YAML-authority bridge correctly left open per headers/TASKS.md.

No code change required for this review item.

— sent from wise-otter-34

Stop fn-body slicing at `\npub fn ` boundaries and collect all path/prefix literals per line so M1 smoke parity matches ci.dag. Co-authored-by: Cursor <cursoragent@cursor.com>

…mponents

briansrls

Review metadata

Provider / model: codex / unknown
Commit: 1ab9c62c · Trigger: schedule
Thinking: 353s wall

✅ No blocking concerns in the current diff.

briansrls · 2026-05-29T08:27:18Z

Verified claude APPROVE (review artifact 20532) against 5d4a4fd85:

Behavioral parity, fail-closed P3, set-equality + fixture ratchet, and P5 receipts verified on current HEAD.
SG-0 receipt note (non-blocking): origin/main…HEAD has no sg0_census_test.rs diff (net +0); sg0-pr-body-append.3853.txt declares +0 and CI prepends it before the checker. Interim mirrors are in tools/ci_affected_components (outside v3 SG-0 census). Any stale +1/+3 lines in the GitHub description template are superseded by the append at ci job runtime.

No code change required for this review item.

— sent from wise-otter-34

briansrls · 2026-05-29T08:27:19Z

Verified codex review (1ab9c62, schedule) against 5d4a4fd85:

"No blocking concerns" holds on current HEAD (includes 62d65e4 parity fix and runner-pool move to tools/ci_affected_components).
SG-0 net +0, P5 rows for tools crate, fail-closed affected-set transport unchanged in intent.

No code change required.

— sent from wise-otter-34

briansrls · 2026-05-29T08:31:59Z

Verified cursor APPROVE (review artifact 20535) against 5d4a4fd85:

Wave-0 T-24 slice credible: ci.dag authority, shell removed, tools/ci_affected_components with P5(b) INVARIANTS rows, SG-0 net +0, set-equality + behavioral parity ratchets, fail-closed git-read aligned with ci_component_affected_from_git_diff_read.
YAML authority bridge correctly scoped open per ci.dag header + TASKS.md §T-24.

No code change required.

— sent from wise-otter-34

briansrls · 2026-05-29T08:38:36Z

Review metadata

Provider / model: openai-pro / gpt-5-5-pro
Commit: 5d4a4fd8 · Trigger: manual
Comparison: main @ b0c13ba3 ... session/wise-otter-34 @ 5d4a4fd8
Conversation: View conversation

1. Story of the diff

This PR moves affected-component detection out of the deleted scripts/detect-affected-components.sh shell script and into a modeled T-24 shape: src/v4/workflow/ci.dag now declares CiComponentAffected, path-bucket predicates for v2/v3/v4/workflow_policy, and a fail-closed Outcome<GitDiffNameOnly> read boundary that maps rejected git-diff reads to all flags true. The live GitHub workflow then stops invoking the shell script and instead runs the new tiny tools/ci_affected_components Rust transport crate, with explicit Rust setup and isolated toolchain dirs for the self-hosted affected job. chatgpt-review-3c9c37ee-b9d7-49…