Repository navigation
Pre-push testing hook: affected-set test gate + non-invasive gates - #6706
Conversation
…non-invasive gates) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…ntract Asserts: fmt checked but no auto-commit, zero 'git commit' in the hook, NEED_WITNESS/claim_executor/gunbc_ci_witness_corpus_only_batches wired, --no-verify escape, dag//src/v2 witness trigger globs. Discriminating: reds if auto-commit is reintroduced. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
… catch-22, not this PR's) The stage0_crate_layout emit produces the .rs without trailing commas (drift-clean form) but rustfmt requires them, so post-#6677 main cannot have that file both drift-clean and fmt-clean. Leave it matching main; the fix belongs to the crate-layout lane, not this hook PR.
…invocations The doc-reachability witness (and others) import v2.std.* modules; the hook's emit_claim_batch passed only --source-root dag, so an enabled hook fail-closed with 'unresolved import v2.std.live_tree' on any docs-changing push. Caught by firing the enabled hook end-to-end. Witness now asserts both source roots. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
gunbai-bot
Bot
changed the title
|
Addressed review 38662 (cursor/composer-2.5, REQUEST_CHANGES): emit_freshness_witness now derives its trigger globs from gunbc.ci_layer_roots.witness_layer_roots (via map(r => r + "/")), and the witness test projects from layer_root_glob_prefixes() — matching witness_trigger_globs_derived — so the values can no longer fork from the single authority. Emitted hook is byte-identical (no drift). Fixed in the latest commit. — sent from royal-carp-451 |
…e roots from authority - Review 38666 (cursor): the note claimed a diff-observation gap 'widens to the full hermetic corpus' — that is the §5 absorbing-fallback the floor was repaired to NOT do. Floor selection REFUSES (typed, located) on a provenance gap, never widens. Note corrected to match ci_spec_discovery_flip_note / compile_door_ledger / cli_run AFFECTED-SET REFUSAL. - Root of the doc-gate resolve bug: emit_claim_batch now derives its --source-root flags from gunbc.ci_layer_roots.witness_layer_roots (single authority) instead of hardcoding, so the doc/artifact gates can't miss src/v2. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
|
Addressed review 38666 (cursor/composer-2.5, REQUEST_CHANGES): you're right — the note misstated fail-closed selection. It claimed a diff-observation gap 'widens to the full hermetic corpus', which is exactly the §5 absorbing-fallback the floor was repaired to NOT do. Corrected: a provenance/diff-observation gap REFUSES with a typed, located diagnostic (never widens, never silently skips), aligned with ci_spec_discovery_flip_note, compile_door_ledger, and the cli_run AFFECTED-SET REFUSAL. Also fixed the root of a separate doc-gate resolve bug found by firing the hook: emit_claim_batch now derives its --source-root flags from ci_layer_roots.witness_layer_roots instead of hardcoding, so the doc/artifact gates can't miss src/v2. Both in the latest commit. — sent from royal-carp-451 |
|
Addressed review 38672 (6212367). emit_run_witness no longer hardcodes the source roots — it now derives them from gunbc.ci_layer_roots.witness_layer_source_flags_rooted(roots: witness_layer_roots), the same authority ci_spec.dag:364/439 uses for the floor's claim_executor invocation. And I closed the deeper fork the finding pointed at: emit_claim_batch was hand-rolling the same fold (a second spelling of the unrooted flags), so it now routes through witness_layer_source_flags(roots: witness_layer_roots). One authority for the --source-root flags across claim_executor (rooted) and claim_batch (unrooted); a witness_layer_roots change can no longer diverge them. Emitted .githooks/pre-push bytes are byte-identical (regen: zero drift, same sha 18d8bf05...); the hook witness githooks_pre_push_emit_witnesses still PASSes green-by-execution. — sent from royal-carp-451 |
Revert .githooks/pre-push and githooks_pre_push_emit.dag to main — the --source-root src/v2 fix is redundant with #6706 ci_layer_roots authority. Revert src/v2/std/algebra.dag to main; §3 dissolve ships in a separate PR. Host-converge Slice 1 does not import v2.std.algebra. Co-authored-by: Cursor <cursoragent@cursor.com>
|
CI investigation: the failing check is NOT introduced by this PR — it is an inherited main-red. Failing witness: doc_graph_has_no_orphan_docs (dag/test/claim/doc_reachability_witness_test.dag) returned Bool(false) in batch 2 of the floor. Root: main commit 8322580 (#6708) added docs/plans/emitted-crate-partition-design.md with no inbound link, so it is an orphan in the doc-reachability graph. That witness fails identically on main itself (run 29471295796, head eba7679) and on the prior main run — it is a standing main-red, blocking every open PR via the refs/pull/N/merge ref (head union main). Evidence this PR is clean: this branch changes NO docs (git diff origin/main...HEAD -- docs/ is empty) and doc_graph_has_no_orphan_docs PASSes when run against this branch in isolation. The affected-set widened to pull the doc witness in only because the shared ci_floor_plan.dag edit touches a widely-imported module — the witness itself is green on our contents. I am fixing the orphan separately (a narrow main-targeting PR that links the doc), NOT bundling it here — bundling would swap an inherited red for a doc change that reads as this PR's. This PR stays byte-identical hook wiring. Merge readiness otherwise: 2 distinct approving providers (claude review 38679, cursor review 38680, both on this head 6212367), no open REQUEST_CHANGES, mergeable=MERGEABLE. Its CI will re-run green once main's orphan is fixed. — sent from royal-carp-451 |
Wire WitnessCorpusRun into githooks_pre_push_plan (layer-root trigger globs + gunbc_ci_witness_corpus_only_batches authority) and stage0 transport; keeps the emitted hook a thin exec shim. Co-authored-by: Cursor <cursoragent@cursor.com>
* WIP: Host-converge Slice 1: wire first real knob (srv1 memory-cap)
* WIP: Host-converge Slice 1: wire first real knob (srv1 memory-cap)
* WIP: Host-converge Slice 1: wire first real knob (srv1 memory-cap)
* WIP: Host-converge Slice 1: wire first real knob (srv1 memory-cap)
* WIP: Host-converge Slice 1: wire first real knob (srv1 memory-cap)
* Dissolve v2.std.algebra fork onto flat std.algebra (P0 field wall).
The nested Monoid/Semigroup/etc. definitions in v2.std.algebra conflicted with
std.algebra after the field-presence wall landed: lookup saw flat fields while
module imports still expected nested literals. Re-export std.algebra types so
all consumers share one authority.
Co-authored-by: Cursor <cursoragent@cursor.com>
* WIP: Host-converge Slice 1: wire first real knob (srv1 memory-cap)
* Mark host_converge_slice1 shell transport with dissolution scaffold.
Mirror fleet_show_effective_read: concat systemctl show/set-property
scripts executed via shell.Exec.Run carry host_converge_slice1_transport_script_scaffold
bound to live_deploy_emit_shell_dissolution_trigger; witness asserts Scaffold row.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Route srv1 converge CLI dispatch via operator_host_srv1 identity.
Replace stringly `host == "srv1"` with HostIdentity comparison against
the fleet_intent_network authority so slice-2 dispatch stays discoverable.
Co-authored-by: Cursor <cursoragent@cursor.com>
* WIP: Host-converge Slice 1: wire first real knob (srv1 memory-cap)
* WIP: Host-converge Slice 1: wire first real knob (srv1 memory-cap)
* Fix fleet_converge_cli parse error in srv1 dispatch guard.
Parenthesize `(operator_host_srv1 as String)` so `as` does not bind the
following `if` block — unparenthesized form caused expected LBrace parse
failure and compile-clean gate red on batch 1.
Co-authored-by: Cursor <cursoragent@cursor.com>
* WIP: Host-converge Slice 1: wire first real knob (srv1 memory-cap)
* Fail-closed on probed_at: refuse instead of "unknown" sentinel.
host_converge_slice1_probed_at returns Absent when date -Iseconds fails
or is empty; all converge/reapply/read-back paths refuse with a typed
reason rather than fabricating a timestamp (§5 absorbing fallback).
Co-authored-by: Cursor <cursoragent@cursor.com>
* Address review 38621: drop grep witness, note ByteSize dissolve-on.
Remove host_converge_slice1_apply_path_wired substring assertion; knob
coverage stays on execution witnesses. Document inherited String seam
debt on scope disposition for live_deploy.emit.
Co-authored-by: Cursor <cursoragent@cursor.com>
* WIP: Host-converge Slice 1: wire first real knob (srv1 memory-cap)
* WIP: Host-converge Slice 1: wire first real knob (srv1 memory-cap)
* WIP: Host-converge Slice 1: wire first real knob (srv1 memory-cap)
* WIP: Host-converge Slice 1: wire first real knob (srv1 memory-cap)
* WIP: Host-converge Slice 1: wire first real knob (srv1 memory-cap)
* Fix Optional null matches in slice1; document MemoryLimitAbsent routing.
Use wildcard Optional arms in host_converge_slice1 converge_loop (runtime
null ≠ ConvergeVerdict::Absent). Route MemoryLimitAbsent through
converge_verdict("", …) with a one-line intent note — semantics unchanged
because effective=="" always yields Absent. Pre-push doc witness passes
--source-root src/v2 for v2.std.live_tree resolution.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Fix Optional null matches in slice1; document MemoryLimitAbsent routing.
Use wildcard Optional arms in host_converge_slice1 converge_loop (runtime
null ≠ ConvergeVerdict::Absent). Route MemoryLimitAbsent through
converge_verdict("", …) with a one-line intent note — semantics unchanged
because effective=="" always yields Absent. Pre-push claim_batch calls pass
--source-root src/v2 for v2 module resolution.
Co-authored-by: Cursor <cursoragent@cursor.com>
* WIP: Host-converge Slice 1: wire first real knob (srv1 memory-cap)
* WIP: Host-converge Slice 1: wire first real knob (srv1 memory-cap)
* Fix fleet_converge_cli witness live-tree disposition; pin Absent routing.
Replace srv1 converge_cli_run with hermetic converge_cli_routes_to_slice1 +
host_converge_slice1_fixture_receipt so SubstrateInputsOnly stays honest.
Add executable witness that MemoryLimitAbsent routes to Absent via converge_verdict.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Unbundle #6695: revert hook edits and std/algebra per ruling.
Revert .githooks/pre-push and githooks_pre_push_emit.dag to main — the
--source-root src/v2 fix is redundant with #6706 ci_layer_roots authority.
Revert src/v2/std/algebra.dag to main; §3 dissolve ships in a separate PR.
Host-converge Slice 1 does not import v2.std.algebra.
Co-authored-by: Cursor <cursoragent@cursor.com>
* WIP: Host-converge Slice 1: wire first real knob (srv1 memory-cap)
---------
Co-authored-by: Brian Searls <briansrls@gunb.ai>
Co-authored-by: Cursor <cursoragent@cursor.com>
1 participant
Pre-push testing hook: affected-set test gate + non-invasive gates
Rebuilds the pre-push hook (
dag/gunbc/githooks_pre_push_emit.dag) so it actually runs the affected tests before push, closing the "local-green ≠ CI-red" gap, and makes every gate non-invasive (report + block, never auto-commit).What changed
ci_floor_plan.dag: gunbc_ci_witness_corpus_only_batches): the same witness corpus CI runs, scoped to the push diff via the floor's own selection authority (no forked affected-set logic). Excludes only the 72-min whole-tree compile gate and the Wet/bin lanes — declared, not silent (gunbc_ci_witness_corpus_only_batches_note). Blocks the push and reports on any failing affected test;git push --no-verifyis the rare-case escape (CI stays the real wall).exit 1(was: autocargo fmt/main_wet+git commiton your behalf + abort). Zerogit commitin the emitted hook.--source-root dag --source-root src/v2(the doc witness importsv2.std.*; the single-root invocation fail-closed onunresolved import). Caught by firing the enabled hook end-to-end.githooks_pre_push_emit_test.dagupdated to the new contract (asserts no auto-commit, the test gate +--no-verifywired, both source roots) — discriminating: reds if auto-commit is reintroduced.Proof (green-by-execution)
Fired the enabled hook exactly as
git pushdoes:HOOK_EXIT=0, 0 FAIL / 231 PASS across doc → artifact → witness → fmt gates. The affected-set gate’s first run caught a real regression (the witness above), which was fixed and re-verified.To activate (per-clone, local config — not a committed artifact)
git config core.hooksPath .githooks(already set in the primary dev container). Cost: ~6–7 min for a small.dagdiff, dominated by the whole-tree resolve; witness caching is the follow-on.