Repository navigation
claw: autofix recipe table -- every verified gate fix as data (AF1, closes #9282) - #9290
Merged
Merged
Conversation
specs/claw/recipes.t27: 19 rows, one per verified recipe, parallel arrays pinned by tests (same discipline as the gates inventory). GATES are workflow stems joined to specs/claw/gates.t27 NAMES, never display names. Classes: C_AUTO 13 (loop fixes, PR, auto-merge), C_REVIEW 3 (prepared patch, person merges), C_OWNER 3 (never executed; report names the keystroke). corpus-ratchet carries 5 rows, and the types pair (bless + verdicts) is pinned to land in ONE PR -- the #9145 lesson. Unknown kinds fail closed to fallback-manual. Folds the claw-af-seed rows (specs/claw/autofix.t27, never PR'd) into this one table so a single SSOT remains; display names re-keyed to real stems. Ledger: str_eq 2 recorded in the same commit -- a deliberate local helper, same shape as gates.t27's (specs cannot import). The two remaining dupe errors (slices_equal, starts_with) are inherited master reds owned by the night loop's next train item; NOT blessed here. AF5 #9287 (owner 2026-10-12): t27o transpiles foreign code into .t27 with no AI in the loop; spec-first-port moves REVIEW -> AUTO on the first verified transpile. Direction recorded in the header, table stays truthful for today. Lab-verified: typecheck OK, 11/11 tests, 0 FAIL, 0 vacuous; dupe scan scoped. Closes #9282. Part of #9271 and #9281. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
gHashTag
enabled auto-merge (squash)
October 12, 2026 17:09
Contributor
gHashTag
added a commit
that referenced
this pull request
Oct 12, 2026
…l apply (AF2, Refs #9283) specs/claw/engine.t27: pure decision logic over recipes' resolved (tool, class, attempt) triple -- kind-slug lookups stay in specs/claw/recipes.t27 (#9290), the table is never re-stated here. Fail-closed on both sides (unmatched signature, unknown instrument), one-class escalation per failed round, two-round budget, and the AF5 invariant as a population test: no apply kind consults a model; every apply is a tool call or a t27o port delegation, which reads deterministic today so recipes' planned spec-first-port flip needs no engine change. Lab-verified: 16/16 pass, 0 vacuous, dupe-clean. Refs #9283, #9281, #9288, #9290, #9271. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
gHashTag
added a commit
that referenced
this pull request
Oct 12, 2026
…l apply (AF2, Refs #9283) (#9292) specs/claw/engine.t27: pure decision logic over recipes' resolved (tool, class, attempt) triple -- kind-slug lookups stay in specs/claw/recipes.t27 (#9290), the table is never re-stated here. Fail-closed on both sides (unmatched signature, unknown instrument), one-class escalation per failed round, two-round budget, and the AF5 invariant as a population test: no apply kind consults a model; every apply is a tool call or a t27o port delegation, which reads deterministic today so recipes' planned spec-first-port flip needs no engine change. Lab-verified: 16/16 pass, 0 vacuous, dupe-clean. Refs #9283, #9281, #9288, #9290, #9271. Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This was referenced Oct 12, 2026
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.
AF1: the autofix knowledge as data. One row per recipe the loops have actually run and verified; the runner (AF2 #9283) and CLI (AF3 #9284) read this table, they do not invent fixes.
Table
GATES / KINDS / TOOLS / CMDS / CONDS / CLASSES), alignment pinned by index tests -- same discipline asspecs/claw/gates.t27.GATESare workflow file stems, the same keys as the inventory'sNAMES(display names never appear, so the join cannot rot).anyis the fail-closed fallback.C_AUTO13 (the loop runs the fix, opens the PR, arms auto-merge),C_REVIEW3 (patch prepared; a person merges -- foreign net-zero, policy-PR-first),C_OWNER3 (never executed; the report names the exact one keystroke).corpus-ratchetcarries 5 rows (its five instruments); the types pair is pinned to land in one PR -- a bless alone leaves the classified gate red (types classification: write the verdicts for Entry, Inputs, SkillMeta (Refs #2774) #9145).CONDSentry is an incident: quiet-window bars for blesses, lab-only + zig-on-PATH for compiler instruments, base64+sha256 transfer protocol, never-force-push, policy-PR-first, label-is-the-owner's-click.Fold + supersession
Folds the claw-af-seed rows (
specs/claw/autofix.t27, never PR'd) into this one table; its display-name keys are re-keyed to real stems (duplicate-bodies -> dupe-ratchet, loop-tools-tracked -> loop-tools-gate, coverage -> seal-coverage, "Documented t27c subcommands exist" -> documented-commands, "Conflict markers" -> conflict-markers). One SSOT remains.Ledger move (same commit)
str_eq 2recorded: a deliberate local helper (specs cannot import; same shape as gates.t27's). The two remaining dupe errors (slices_equal,starts_with) are inherited master reds owned by the night loop's next train item -- deliberately NOT blessed here (census-pin trap).AF5 direction (recorded, not yet truth)
#9287 (owner 2026-10-12): t27o transpiles foreign code into
.t27from any language, no AI in the loop. When the first transpile is verified,spec-first-portmoves REVIEW -> AUTO with toolt27o. The header records the direction; the table records the day.Verification
tools/dupe_scan.py: my group recorded in-ledger; no new groups beyond it.Closes #9282. Part of #9271, #9281; feeds #9283 (engine), #9284 (CLI), #9285 (harvest continues), #9287 (transpiler).
🤖 Generated with Claude Code