f7df920681f233ae0fe064ee659550bdf41ff708
4 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
Tom Boucher
|
738f42f4fd |
feat(#2398): consensus gate for CYCLE_SUMMARY on multi-reviewer runs (#3755)
* test(#2398): failing-first suite for the CYCLE_SUMMARY consensus gate Binds the gate before it exists, so the suite is RED against next. The load-bearing rows are the two the closed PR #2417 did not have. The B2 regression row asserts a judgment-class lone HIGH counts WITHOUT corroboration when its raiser is unmarked — if anyone re-couples that class to corroboration, more reviewers again produce a weaker gate than one, which is what closed #2417. The parity row asserts every marker literal the gate names is one review-lane-runner actually emits, so the gate cannot key on a signal nothing produces; a mutation row and a seeded fast-check property prove that guard runs its failure branch rather than only reading a correct tree. Also pinned: gate position before Counting rules, the untouched CYCLE_SUMMARY line shape the orchestrator greps, fence balance, the single-reviewer no-op, classification by what a claim asserts rather than by citation presence, the all-marked fail-open, current_actionable staying out of scope, and the leading-marker requirement that stops a review which merely quotes a marker from suppressing its own findings. * feat(#2398): consensus gate for CYCLE_SUMMARY on multi-reviewer runs With review.reviewer_instances running several reviewer identities off one adapter, any single instance's fabricated HIGH could force a full replan cycle on its own. Across ~9 real cycles on two projects each of four instances fabricated at least once, and each was also the most accurate reviewer in some other cycle, so dropping to fewer reviewers trades away real signal. The gate engages only when 2+ reviewers actually ran, and weighs a lone HIGH by what the claim asserts rather than by whether anyone agreed with it. An existence claim -- a symbol, file, flag, commit or ID exists, is absent, or says something specific -- counts only if source-grounding confirms it or another reviewer raised the same concern. A judgment claim -- a design or correctness property -- counts unless that reviewer's own section opens with an evidence-quality discount marker the review lane already stamps ([reviewed-without-source-citations] #3194, [reviewed-without-repo-access] #2176, or a diff-only lane). That split is what resolves B2, the finding that closed PR #2417. B2 showed the approved wording made more reviewers produce a WEAKER gate than one: condition (a) pointed at the source-grounding pass, which verifies every symbol THE PLAN cites and never takes reviewer claims as input, so a genuine architectural HIGH that one reviewer caught and another missed was neither groundable nor corroborated and stopped gating. Judgment-class findings are therefore exempt from corroboration entirely -- reviewers catch materially different classes of issue, and demanding two of them independently raise the same architectural concern suppresses exactly what a multi-reviewer setup exists to surface. Guards on the gate itself: an all-marked cycle disengages it, so a cycle in which nothing was verified can never be counted as converged; the marker must OPEN a reviewer's section, so a review that merely quotes a marker does not suppress its own findings; a suppressed HIGH stays listed and tagged rather than dropped; current_actionable is untouched; and a single-reviewer run is unchanged. No new command, config key, or dependency -- the gate reads signals that already exist. The CYCLE_SUMMARY line shape the orchestrator greps is unchanged; only the integer it computes moves, and only for 2+ reviewers. Known limit, inherited rather than introduced: SOURCE_CITATION_RE checks citation presence, not resolution, which src/review-lane-runner.cts records as a deliberate #3194 scope boundary. A fabricated but plausible file:line still gates. Scope revised and re-approved on the issue before any code was written. * test(#2398): make marker parity behavioral, and stop overclaiming the gate Review found the parity tests were vacuous: they asserted a marker STRING appeared in review-lane-runner.cjs's source text, never requiring the module or calling the stampers, so they would pass even if stampUngroundedReview were broken or never invoked. They now invoke the real exported functions and assert what those functions PRODUCE — that an uncited review gains a leading marker blockquote, that a review carrying a file:line does not, that a self-reported blind review is stamped, and that stamping is idempotent. Removing the source read also removes an incidental no-source-grep evasion via a parameterized path. Review also found the changeset headline false for the class it matters most in. The discount markers detect 'cited nothing' and 'had no repo access'; they cannot detect 'drew a wrong conclusion from a real citation', so a judgment-class finding invented by an evidence-bearing reviewer still counts alone. That is the deliberate side of the tradeoff jags-faith named when closing #2417 — the alternative is requiring corroboration for design findings, which is B2 — but the changeset claimed lone hallucinations no longer force a cycle, full stop. Corrected there, and stated plainly in docs/COMMANDS.md and the design record. Also dropped the reviewer-instances.md entry from the emitted-drift ack: the growth ratchet's currentSizes() scans only gsd-core/workflows/ and agents/ (tests/helpers/emitted-runtime.cjs:916-929), so references/ is outside it and that entry acknowledged a delta the gate cannot see. * chore(#2398): backfill changeset pr number to 3755 --------- Co-authored-by: sim <sim@local> |
||
|
Tom Boucher
|
3f6b063fbb |
chore(#2799): invoke_reviewers and write_reviews iterate declared lanes (#2861)
* chore(#2799): resolve reviewer lanes into executable invocation plans Phase 5b of ADR-2782. Adds the resolver and runner that let invoke_reviewers iterate declared lanes instead of hand-authored per-CLI bash. Five additive descriptor amendments, each forced by a lane that ships today: - LaneHandler gains 'opencode' — the lane rebuilds its review from assistant text parts of a --format json stream; a plain stdout copy re-breaks #1936. - modelConfigKey — antigravity's key is review.models.agy, not .antigravity, so resolving by slug silently dropped a configured model. - defaultHost/fallbackModel — Phase 4 federated every *_host with a default of empty string; the real fallback only existed in the bash. - args becomes an argv template with a closed four-placeholder vocabulary. Positional splicing produced 'codex --model M -o F exec --ephemeral', which is not a valid invocation: codex injects in the middle, twice. - kimi-code lane, with the bounded command-capability probe (needle --output-format) that tells Kimi Code from the legacy python kimi-cli. Parity gate re-pointed: the workflow-text families it scanned are the text this phase deletes, so they are replaced by descriptor-to-registry parity plus an anti-parity check that no bespoke leg returns. jq, curl and external timeout/gtimeout all drop out of the review path. Refs #2782 * chore(#2799): add review-lane query surface and widen the manifest vocabulary Adds the gsd-tools 'review-lane' route (plan/invoke/sections) the workflow loops over, projects all twelve lanes into their capability manifests, and widens capability-validator for the amendments. opencode admitted to VALID_LANE_HANDLERS under the second arm of the enum's own admission rule: one lane, justified by a documented upstream defect data cannot express (#1936 — the agent can end its turn with zero output tokens and --format default then drops the assistant text entirely). Two bugs caught by an end-to-end stub run and fixed here: - loadConfigResolved returns a provenance wrapper, not the config; using it directly resolved every key to undefined, which reads as 'nothing configured' and silently dropped every model override. - hasBinary used shell:true with an args array (Node 26 DEP0190). Replaced with a PATH scan that spawns nothing at all. Refs #2782 * chore(#2799): iterate declared lanes in invoke_reviewers and write_reviews Replaces the eleven hand-authored per-CLI bash legs with a loop over resolved lanes, and renders REVIEWS.md sections from each lane's declared reviewsSection instead of thirteen hardcoded headings. review.md drops from 1104 lines to 507 (61KB to 28.7KB). Parity gate re-pointed, as agreed: the leg-marker and section-heading families scanned exactly the text this phase deletes, so they are replaced by descriptor-to-registry parity in both directions, plus an anti-parity check that fires if a bespoke leg is ever re-added. Enum, emitting sites and the Object.keys lock moved together. The budget-trim helper is hoisted out of the Ollama leg: it was always lane-agnostic, and any lane may now declare a promptBudgetKey. Refs #2782 * feat(#2799): bind the consented egress host and re-verify it at invocation Completes ADR-2782 D5. Rule 1 was recorded in the ADR as delivered by Phase 3 but was not implemented: ConsentRecord had no host field and nothing in the tree bound one, so this phase's rule-4 comparison had no baseline. ConsentRecord gains an OPTIONAL reviewerHost. Optional is the whole design: isValidConsentRecord does not require it, so every record already on disk stays valid and no re-consent storm fires (D4 rule 5). It is deliberately excluded from disclosureSignature — the loader has no config resolver, so folding a config-derived value in would make loader and lifecycle compute different signatures for the same manifest and re-prompt forever. Install resolves hostConfigKey (falling back to the lane's declared defaultHost, which is what the invocation path uses) and records it. Invocation re-resolves and blocks on mismatch rather than silently redirecting. Absence allows: no record, or a record predating the field, means nothing to compare — denying there would break every existing local-model user on upgrade. Refs #2782 * test(#2799): cover the resolver, runner and handlers; retarget the parity suites Adds the golden invocation-plan table (one row per shipped lane, derived from the bash legs rather than the descriptor types) plus runner coverage for the probe, empty-output policy, the three handlers and the egress check. Retargets the existing suites onto the new contract: descriptor-to-registry parity, the anti-parity check, the opencode handler, and the twelfth lane. Two corrections found by running them: - modelConfigKey was required; that breaks D4 rule 2, since a reviewer manifest authored before this phase would fail validation on upgrade. It is optional, read as null when absent. - the antigravity non-zero-exit test pre-seeded the transcript, which asserted that a STALE entry leaks through — the exact bug the watermark prevents. The spawn now appends, as the real tool does. Refs #2782 * fix(#2799): restore agy --add-dir and the self-report prompt in the handler Retargeting the three legacy reviewer suites off the deleted bash surfaced two real regressions in the port, both #2176: - --add-dir was dropped. Without it agy's permission context never receives the cwd repo, so the agent anchors on its own scratch dir and reviews the plan text in isolation — the exact failure the Review Instructions forbid. It is capability-probed, because an older agy rejects the unknown flag outright and a lane that fails to start is worse than one running on the prompt anchor. - the prompt lost the clause mandating a REVIEWED-WITHOUT-REPO-ACCESS self-report, which is what makes a blind review distinguishable from a grounded one. antigravity now builds its own prompt variant. Also ports the #2073 mode-2 cli.log diagnostic, which was dropped: a pinned model that 404s exits 0 with empty stdout AND an empty transcript, so agy's own log is the only evidence that anything failed. The three suites now assert against the plan and the handler instead of matching fence text, so they no longer need allow-test-rule exemptions. Refs #2782 * docs(#2799): document the declared lanes, the new flag, and dropped prerequisites COMMANDS.md gains --kimi-code and replaces the jq-prerequisite paragraph, which is now false: no lane requires jq, curl or an external timeout. Adds the changed-egress-destination behavior, since a blocked lane is something a user can hit. CONFIGURATION.md records that the model config key is declared per lane rather than derived from the flag — antigravity's is review.models.agy — and adds review.models.kimi-code. reviewer-instances.md now routes an instance through its lane's single invocation seam instead of a copied per-adapter bash block, which is what lets a cross-cutting fix reach instances for free. That required implementing the --model/--agent/--as flags it documents; --model re-resolves through the lane's argv template rather than splicing, so the flag lands where the lane declares it rather than ahead of a subcommand. CONTEXT.md glossary gains both new modules. Refs #2782 * chore(#2799): drop the stale emitted-drift acknowledgment The only entry was #2797's, acknowledging COMMENT-ONLY GROWTH in review.md. That file now shrinks by ~32KB and every emitted hash that moved is attributable to this diff, so the ack no longer explains anything. Removing the last entry means removing the file: its presence is the alarm, and an empty one signals nothing. Verified by deleting it and re-running the attribution and provenance gates plus lint:ci — all green without it. Refs #2782 * docs(#2799): record the Phase 5b vocabulary widenings in ADR-2782 Five additive amendments, each forced by a lane that ships today, plus two corrections the phase had to make rather than work around: - D5 rule 1 was recorded as delivered by Phase 3 and was not implemented, so this phase's rule-4 comparison had no baseline. Recorded because an ADR asserting a rule was delivered is exactly what stops a later phase checking. - The DEFECT.GENERATIVE-FIX gate is re-pointed: its workflow-text families scanned the text this phase deletes. Also records that D7's 'skip the probe where no bounding mechanism exists' carve-out is obsolete — in practice it meant the Antigravity lane ran unbounded on every stock macOS host, which ships neither timeout nor gtimeout. Refs #2782 * fix(#2799): close four defects found by adversarial review Two confirmed bugs, both reproduced before fixing: - resolveLanePlan was not total. An openai-http lane with a missing or non-object invoke dereferenced inv.hostConfigKey and threw, contradicting the module's own documented contract; the spawn branch guarded correctly and the http branch did not. The CLI seam resolves every selected lane in one map, so one malformed overlay manifest would have aborted the whole review rather than dropping its own lane. Guarded, plus a per-lane try/catch at the seam so a throw can never take down siblings. - A reviewer-instance model was silently dropped for any lane declaring modelConfigKey null (cursor, qwen, coderabbit). reviewer_instances validates that cli is a known slug but never that the slug accepts a model, so a user could configure one, get a clean run, and never learn a different model reviewed their plan. Now warns explicitly. Two hardening fixes: - The slug is concatenated into artifact paths, so LANE_SLUG_RE is enforced in the resolver rather than inherited from a validator that does not run on this path — the module documents itself as the overlay-manifest trust boundary, so it should not depend on someone else having checked. - normalizeHost mangled a scheme-less value: new URL('localhost:11434') parses with an empty hostname, so it became 'localhost://11434' and was compared and requested as if real. An empty hostname now means not-a-URL. Also documents the one gap that cannot be closed here: the antigravity watermark is keyed by workspace, so two concurrent reviews of the same repo share a transcript. agy exposes no per-invocation id to filter on, so the handler now states which half of its never-stale guarantee actually holds. Refs #2782 * test(#2799): retarget the remaining eight review.md-asserting suites The remote runner found 37 failures the local sweep missed (it hit the shell's two-minute cap before reaching these). All eight extract per-CLI bash from review.md that this phase deletes; each protects a real invariant, so each is retargeted onto the plan, the runner or the handler rather than removed. Three real defects surfaced by doing so: - effort args never reached ANY lane. model-resolver.cjs exports no resolveExecution, so effortFor silently returned [] every time. Restored by calling the same bounded resolve-execution query the bash legs used — and NOT with --raw, which prints the resolved effort rather than the picked field, so claude got 'low' instead of '--effort low'. - the timeout guidance lost 'a silent empty output is a timeout kill, not a crash' — the operator note that exists because of the Codex 0xc0000142 misdiagnosis. Restored. - the opencode handler dropped EMPTY assistant text parts. The shipped jq was , and only substitutes for false/null — an empty string is truthy in jq and contributed a blank line. Found by a property test shrinking to ['', '']. The opencode property suite no longer spawns jq at all, which deletes the #2099 hang mechanism it was architected around rather than mitigating it. Refs #2782 * fix(#2799): register the two new generated modules, and untrack them The remote runner caught build output committed to git. Both new modules compile from src/*.cts into gsd-core/bin/lib/*.cjs, and every sibling generated that way is gitignored and eslint-ignored (ADR-457) - including Phase 1's own review-lane-descriptor.cjs. Mine were neither, so repo-invariants' "each bin/lib/*.cjs is linted xor ignored according to migration state" failed. Registered both in .gitignore and eslint.config.mjs alongside the Phase 1 module, and dropped them from the index. Nothing about the shipped behaviour changes; the artifacts are rebuilt by build:lib. This is the new-.cts-module registration ripple, and it is the one part of it I had not completed - the CONTEXT.md glossary and the inventory manifest were already done. Refs #2782 * chore(#2799): backfill changeset pr number to 2861 * chore(#2799): backfill changeset pr number to 2861 --------- Co-authored-by: Test <test@example.com> |
||
|
Tom Boucher
|
40ce95f882 |
fix(#2358): scope review.md and ship.md temp files to a per-run mktemp directory (#2433)
* fix(#2358): scope review workflow temp files to a per-run mktemp dir /gsd-review wrote every prompt/section/output temp file to a hardcoded /tmp path keyed only on the phase number, so two GSD projects sharing a small phase number collide on the exact same path and a crashed run's leftover file becomes bait a later, unrelated run can silently read. ship.md's external peer-review stderr capture was strictly worse — one shared, unqualified path across every project/phase/run. Thread a single mktemp -d "${TMPDIR:-/tmp}/gsd-review.XXXXXX" run directory through every review.md temp path (67 sites) via a new {run_dir}/$RUN_DIR placeholder, mirroring the existing {phase} substitution mechanism, and clean it up at the end of the run. Route ship.md's stderr capture through a per-run mktemp file the same way. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * chore(#2358): regenerate fixtures + lint gate-prep * fix(#2358): repair failing tests after gate verification * fix(#2358): thread RUN_DIR scoping into reviewer-instances.md (#1517) review.md's own invoke_reviewers step lazily loads gsd-core/references/reviewer-instances.md for the review.reviewer_instances codepath, but that doc was missed when review.md and ship.md were moved to the run-scoped {run_dir} temp directory. It still read the combined prompt from the old /tmp/gsd-review-prompt-{phase}.md (which build_prompt no longer writes, breaking reviewer-instances functionality outright) and wrote each instance's output to the old unscoped /tmp/gsd-review-${INSTANCE_NAME}-{phase}.md, leaving the exact cross-project temp-file collision bug open for that code path. Both paths now thread through {run_dir}, matching every other reviewer block in review.md. Extends the existing #2358 regression test with assertions pinning reviewer-instances.md's prompt read and output write to {run_dir}, and adds the Fixed changeset fragment. Regenerated the golden-install-parity content hashes for reviewer-instances.md via `npm run gen:golden` (paths unchanged; only the modified file's hash moved). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * chore(#2358): backfill changeset pr (#2433) --------- Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com> |
||
|
Tom Boucher
|
b0d5ca3379 |
feat(#1517): support custom reviewer instances for /gsd:review (#1766)
* feat(#1517): support custom reviewer instances for /gsd:review Add a bounded review.reviewer_instances config surface so one model-capable adapter (e.g. opencode) can run as several independent reviewer identities in a single /gsd:review pass. Instances participate only via review.default_reviewers, expand before built-in slugs, are available iff their cli is detected, and a non-matching entry is a hard error (typo must be loud). >=2 same-cli instances emit a shared-adapter caveat in REVIEWS.md. Default path with no instances is byte-for-byte unchanged. Single-source instance->cli resolution lives in resolveReviewerSelection / normalizeReviewerInstances (parity-locked in tests/review-reviewer-instances.test.cjs). cli validated against KNOWN_REVIEWER_SLUGS only (never arbitrary shell); model/agent opaque, never shell-interpolated. Closes #1517 * chore(#1517): backfill changeset pr:1766 --------- Co-authored-by: review-bot <review-bot@gsd> |