0.14.65 — war-memory CLI correctness & lint completeness - #1173
Merged
Conversation
…rchive/ The tighten lesson was archived on 2026-07-27 (PR #1167, operator-directed RESOLVED sweep), so Task 1.2's declared path no longer resolves. Archiving is a move, not a deletion — the lesson stays FTS-queryable, so its inaccurate zsh-fuse mechanism prose still misleads and still needs correcting. Re-anchored in place rather than restoring the lesson to the hot set: restoring would add its row back and push the projection over the 17,000 B advisory that the same sweep just cleared (16,797 B, verdict ok). The sibling cmdquery lesson is unaffected — still hot at its declared path. Lead adjudication at the campaign decompose gate. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
- End state 1 + Purpose narrowed to space-separated flag values; the =-attached residual (D10, frozen parseArgv) lifted into the Intent as a named still-silent out-of-scope path instead of a Notes-only mention. - Task 1.2 sweep + End state 9 greps made case-insensitive (grep -in), token-anchored — a re-cased straggler could previously false-negate. - ADR 0028 quotes the hub-WARN verbatim: added to Task 1.1 Files with a dated append-only amendment; per-medium gate (NEW-present on the ADR, OLD-absent on live skills/ surfaces) added to End state 8. - The existing spawn-the-CLI hub-WARN test pins the message bytes and reds on the noun flip: named as a required in-task edit; End state 8's stay-green clause narrowed to the inboundCiters unit tests. - cmdQuery guard triple gets coupling comments both ways (cmdTightenPlan's --target block named as reference implementation). - End state 3's explicit-defaults invocation built from imported constants, not hand-typed 10/4096 literals. - End state 6 re-filed HARD (task gate runs the recursive live-tree lint pre-merge); backstop 1 narrowed to the post-gate window. - Backstop 2 rewritten: in-run runner = Phase 1 servitor (recorded resolved-lesson-stamp convention); Lead Gate-2 duty to exclude the archived tighten slug from promotion (flat-copy would duplicate it hot). - Contention note corrected: three siblings, not one; plan 4's README + version-slots.test.mjs Phase-1 overlap recorded with mitigation. - Doc-sweep note widened to the WARN message string. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…ings patched, sandbox escape contained Round 0: 14 probes on-target (6 executed / 8 analyzed), 20 agents, opus/high. 8 gate blockers + 4 needsDecision patched into the plan at 4453863; round-1 re-verify confirms all 12 RESOLVED. 5 Minors absorbed by the same patches. Incident: the ff-topology probe built its synthetic fixture in the shared sandbox worktree (5 rogue commits + working-tree edits). Contained to the disposable sandbox — real repo clean, remote tips unchanged; taint audit shows every load-bearing finding anchored on unmutated surfaces; sandbox restored and escape guard re-run exit 0 before the verdict. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
) Three CLI correctness fixes, each bound at its owning verb's boundary: - cmdQuery (#1145): --top-k and --budget get the three-way argv-boundary resolution #1059 ratified for tighten-plan's --target — absent ⇒ default (byte-identical flagless path), a positive finite string binds, anything else refuses loud (exit 1, stderr naming the flag and the received token). The guards sit above walkCorpus, so a refusal writes nothing: no query-log line, no prompt block. No integer check (D3 — --top-k 3.5 binds, mirroring --target's predicate exactly). parseArgv is frozen, so the =-attached form (--top-k=abc) remains the named, deliberately silent residual (D10). Three comment-coupled sites, no shared helper. - cmdLint (#1135): the directory branch walks recursively, so the committed archive/ subtree is linted on every surface sharing the one invocation — CI (memory-audit.yml) and the gate wrapper (war-memory-lint.test.sh) both stay byte-unchanged. The catch { continue } fail-open is preserved. - cmdArchive (#1154): the advisory concept-hub WARN counts slug-deduped citers for both the >=2 predicate and the printed number, converging on the inbound verb and tightenPlan's Floor 2; the count noun becomes "inbound citers" (the old noun would be false on the deduped count). ADR 0028 quotes the WARN verbatim: a dated append-only amendment records the new counting and noun; the historical quote stands (superseded, not rewritten). Regressions (all red pre-fix, verified by reverting the source alone): eight spawn-the-CLI refusal cases with per-case token fidelity plus no-query-log-on- refusal, unchanged-path cases built from the imported DEFAULT_TOP_K/ DEFAULT_BUDGET constants, archive/-nested lint fixtures at both the CLI and wrapper surfaces, and three hub-WARN dedupe cases (twin ⇒ no WARN, two distinct ⇒ 2, twin + distinct ⇒ 2 not 3). The existing hub-WARN CLI test is updated in place — it pins the message bytes, so the noun flip reds it by design. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…amp (#1171) Two committed-lesson corrections, sourced from the red-team report (docs/red-team/2026-07-24-memory-tooling-hardening.md, Minors, "Purpose, zsh threading") and spec §4 of docs/specs/2026-07-26-war-memory-cli-correctness-design.md — not re-derived: - tighten-target-flag-...-degradation-paths.md (archive/, moved there by the 2026-07-27 RESOLVED sweep): failure-mode-3 now states the true fuse mechanism — zsh, lacking SH_WORD_SPLIT, does not word-split the unquoted ${TIGHTEN_TARGET:+...} expansion, so the whole replacement fuses into one argv word (`--target 2000`), which parseArgv keys as `{"target 2000": true}`; argv.target itself stays undefined. The "same silent outcome as never setting $TIGHTEN_TARGET" clause is kept verbatim (it was the accurate part). A case-insensitive grep for dropped/collapses now finds zero hits. The closing "Not yet extended to siblings" paragraph is rewritten past-tense as "Extended to siblings (war-memory-cli-correctness/1.1, #1145)", citing Task 1.1's landed cmdQuery guard by construct name. - cmdquery-topk-budget-share-tighten-targets-pre-fix-truthy-ternary-shape.md (D9): description RESOLVED-prefixed; appended `## RESOLVED` section names Task 1.1's guarded cmdQuery resolution and its `refusal (#1145): --<flag> ...` test banner in war-memory.test.mjs. Per the ratified stamp convention the "deliberately left open" H1 and pre-fix code snippet stay frozen — the appended section is what corrects the record. Hand-scan survey (both files, full body): zero additional paraphrase stragglers beyond failure-mode-3 — matches the spec's own spec-time survey result (the sibling lesson carries no zsh claim). Redaction lint clean on both edited files. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…1171) Absorb-disposition Minor findings from the task 1.2 audit: 1. [Minor] New RESOLVED prose misattributes the NaN --budget failure mode to selectForBudget's slice (it uncapped, it never emptied). Rationale: selectForBudget slices only on topK; budget is used in 'bytes + size > budget' — with NaN that comparison is always false, so the loop never breaks and the block runs UNCAPPED, the opposite of empty. Fix: split the clause by flag — a NaN --top-k empties via slice(0, topK); a NaN --budget silently uncaps (the break never fires). 2. [Minor] 'bare/NaN --top-k had silently emptied the prefetch' — the bare flag gave a ONE-lesson block, not an empty one. Rationale: pre-fix a bare --top-k mapped to boolean true and Number(true) === 1 selected one lesson; only NaN/zero/negative values emptied the block via slice. Fix: split bare (one lesson) from NaN/zero/negative (emptied) in the archive lesson's 'Extended to siblings' paragraph, and mirror the same split in the sibling lesson's RESOLVED-prefixed description (the stamp convention's correction channel). Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
#1170) Lead manual completion of a held:escalation phase. The post-merge gate-audit (execution-evidence seat) found End state 8's OLD-absent clause false at the confirmed tip: the retired WARN literal survived on three live `skills/` surfaces, none of them exempt. Reproduced by the Lead at 36e04d0 before acting — three hits, all real: skills/lessons-learned/SKILL.md:256, :257 (the Phase-3 hub-check thresholds) skills/lessons-learned/lessons-learned-doc-contract.test.mjs:88 (comment) Not cosmetic. Those SKILL.md lines document `war-memory inbound`'s thresholds, and cmdInbound dedupes by slug — `[...new Set(inboundCiters(...).map(r => r.slug))]` — so on a deduped count `refs` is false: a cross-root twin is 2 refs but 1 citer and is NOT a hub. That is precisely the falsehood this plan retired, so the surviving prose contradicted the tool it documents (ADR 0025 cascade class). Verified the doc-contract test asserts on the `inbound <slug>` invocation and the absence of the retired prose grep, never on the retired literal — so the flip reds nothing. Scope widening adjudicated by the Lead rather than deferred: End state 8 is a plan End state, and a validation in neither the gate, a floor, nor the backstops section may not be waived in prose (ADR 0017). Fixing it makes the End state true; a follow-up issue would have left it false. End state 8 now holds both halves at this tip: OLD-absent = 0 hits across skills/, hooks/, agents/, CONTEXT.md, README.md; NEW-present verified in war-memory.mjs (4) and ADR 0028's dated amendment. Gate green: 990 JS tests, all shell tests, war-memory lint docs/learnings/ clean. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
All four release slots bumped in lock-step to 0.14.65, the next free patch: plugin.json#version, marketplace.json#metadata.version and #plugins[0].version, and the README ## Status token. Resolved from the live slots at the Phase-1 landed tip 086db74 and origin/master (both 0.14.64), not from any plan literal; version-slots.test.mjs green on all four checks, including the monotonic floor (window max 0.14.64). The Status blurb names the one consumer-visible behavior change first-class: the redaction-lint walk is RECURSIVE, so a plugin user's previously green learnings tree holding a violation under archive/ exited 0 before this release and exits 1 now. Phrased as the walk being recursive and as a real new failure with a remedy — never as "lints all archives clean" (the recorded overstating-guard-semantics trap). Also enumerated: cmdQuery's --top-k/--budget refusal guards (a supplied-but-invalid space-separated value exits 1 instead of binding a degraded one; the =-attached form named as the frozen D10 residual) and cmdArchive's slug-deduped hub WARN with its `inbound citers` noun, plus the ADR 0028 amendment, the /lessons-learned threshold noun retirement, and the two lesson truth-ups. Two traps checked against the drafted prose, not assumed: - Count word: "Three correctness fixes" is followed by exactly three numbered items; the one other count ("three comment-coupled sites") names its own members inline (--target, --top-k, --budget). - The retired WARN literal is NOT reintroduced — README.md is inside End state 8's OLD-absent grep scope (skills/, hooks/, agents/, CONTEXT.md, README.md), so describing the noun flip in the blurb would have re-fired the rename's own absence guard. Re-verified zero hits over that scope after the edit. Code facts in the blurb were traced at the tip rather than copied from the Phase-1 commit message: selectForBudget's records.slice(0, topK) empties on a non-numeric or zero --top-k while a NaN --budget leaves the block UNCAPPED (the `bytes + size > budget` break never fires), and parseArgv keys `--top-k=abc` whole, leaving argv['top-k'] undefined. "Negative empties" was deliberately not claimed — slice(0, -n) truncates from the end, it does not empty. Gate: 990 JS tests pass, 27 shell suites pass (exit 0), including war-memory-lint.test.sh linting the live docs/learnings/ tree clean with the recursive walk active. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…e-2 into dev/2026-07-26-war-memory-cli-correctness
Gate-2 promotion, deferred until phase 2 landed so the commit could not strand phase 2's land CAS as land_stale (its integration branch was cut from the pre-commit working tip). - redteam-sandbox-probe-branch-residue-blocks-next-provision: a git-mutating red-team probe that escapes into a shared sandbox worktree leaves REFS behind that a working-tree reset --hard/clean -fd does not undo; a residual branch then collided with the next run's Provision barrier (ADR 0003). - old-absent-gate-half-relies-on-unrecorded-hand-grep-fails-silently: in a per-medium NEW-present/OLD-absent retirement gate, the OLD-absent half had no committed assertion and rested on an unrecorded hand grep — which is the half that silently returned the wrong answer (caught at post-merge gate-audit, 3 live stragglers). - release-blurb-quoted-code-literal-can-diverge-from-actual-identifier: a blurb quoting a call with the CLI argument name instead of the source identifier defeats the grep-the-rendered-literal verification this repo relies on. EXCLUDED from promotion (red-team adjudication 3, fail-closed): tighten-target-flag-has-three-independent-silent-degradation-paths — its repo copy lives under docs/learnings/archive/, and promotion copies FLAT into docs/learnings/, which would land it hot and duplicate the archived copy. NOT promoted: cmdquery-topk-budget-share-… — Task 1.2 already corrected the repo copy in place; the servitor only mirrored that correction locally. Release slots re-verified at 0.14.65 before staging (#1083 hazard). Redaction lint clean over the full recursive docs/learnings/ tree. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This was referenced Jul 28, 2026
This was referenced Jul 28, 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.
Campaign 2026-07-26-doc-truth-and-engine-hardening, plan 2 of 4. Both phases landed via
/war --afk --ace. Plan 1 (#1166) is already on master, so this cuts from and lands tomasterwith no stacking constraint.Plan:
docs/plans/2026-07-26-war-memory-cli-correctness.md· Red-team:docs/red-team/2026-07-26-war-memory-cli-correctness.md(CLEARED, 12 blocking findings patched) · Epics #1168, #1169 · Source issues #1145, #1135, #1154, #1147.What lands
cmdQuery --top-k/--budgetguards (war-memory: cmdQuery --top-k/--budget keep the truthy-ternary silent-degradation shape --target just retired #1145) — a supplied-but-invalid space-separated value now exits 1 with empty stdout, stderr naming the flag and the received token, and no query-log line (the guard sits beforewalkCorpus, so refusal writes nothing). 8/8 refusal regressions, red pre-fix. The=-attached form (--top-k=abc) remains a known, still-silent, deliberately frozen residual —parseArgvis frozen by ratified spec decision D10, and the Intent names it rather than waiving it.cmdLintrecursive walk (war-followup: redaction lint (CI + new gate wrapper) never walks docs/learnings/archive/ — ~104 committed lessons unlinted #1135) —lint <dir>now descendsarchive/. CI (memory-audit.yml) and the gate wrapper (war-memory-lint.test.sh) inherit it with zero edits (both byte-unchanged, checked from the task diff). Consumer-visible: a previously-green learnings tree with a nested violation will start failing — that is the lint doing its job.cmdArchivehub WARN slug-deduped (war-memory: archive's concept-hub warning counts a cross-root-dupe citer twice, disagreeing with inbound's slug-deduped count #1154) — the>= 2predicate and the printed count now both come fromnew Set(hotInbound.map(r => r.slug)).size, and the noun isinbound citers. On a deduped countrefswas false: a cross-root twin is 2 refs but 1 citer and is not a hub. Independently observed live by the Lead during a memory sweep before this plan ran.Unexecuted backstops
3885e9e:lint: cleanover the full recursivedocs/learnings/tree (145+ archived files). Red-team narrowed this at round 0 — the task gate already runs it pre-merge and blocking, so End state 6 was re-filed HARD; only the gate→merge window was ever deferred, and CI re-checks it on this PR.--argslive transport: PARTIALLY DISCHARGED. Both phases dispatched through the newly landedstage-workflow.mjs --args <file>CLI path (50,330 B and 32,414 B) with no Workflowargs; the phases running at all is the transport observation. The ~104.5 KB class remains unobserved, so this is partial, not closed. Its second half — carrying both lesson corrections into the local memory root — was discharged in-run by the Phase 1 servitor, per red-team adjudication 3.Notable during execution
escalationon a post-merge gate-audit Major: End state 8's OLD-absent clause was false at the tip — the retiredinbound refsliteral survived on three liveskills/surfaces, including the operator-facing hub thresholds inskills/lessons-learned/SKILL.md, which thereby contradicted the tool this task converged. Reproduced by the Lead, then fixed at086db74rather than deferred: a plan End state may not be waived in prose (ADR 0017), and a follow-up would have left it false.workflow-errorat the Provision barrier — a red-team probe escape had left a repo-global integration branch that the barrier correctly refused (ADR 0003). Cleaned and relaunched fresh. Both incidents are now recorded lessons in this PR'sdocs(learnings)commit.Residual risk (accepted, all named in the plan)
The
=-attached flag form stays silent (D10); recursivereaddirSyncdoes not descend directory symlinks; an unreadable subdirectory skips its whole target fail-open. Two Nits landed unfixed and recorded: the release blurb renders the recursive call with the CLI argument name rather than the source identifier, and the plan Notes' three one-line residuals are unclaimed by any task slice.🤖 Generated with Claude Code