fix(agent-actions): close hard-blocked contributor PRs in auto mode - #2856
Merged
Conversation
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## main #2856 +/- ##
=======================================
Coverage 96.01% 96.02%
=======================================
Files 259 259
Lines 28304 28334 +30
Branches 10291 10298 +7
=======================================
+ Hits 27177 27207 +30
Misses 491 491
Partials 636 636
🚀 New features to boost your workflow:
|
loopover-orb
Bot
added
the
gittensor:bug
Contributor
|
Warning 🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨 ⏸️ Gittensory review result - manual review recommendedReview updated: 2026-07-04 02:33:52 UTC
⏸️ Suggested Action - Manual Review
Review summary Nits — 6 non-blocking
Review context
Contributor next steps
Signal definitions
🟩 Safe / merged · 🟦 Advisory · 🟨 Held for review · 🟥 Blocked / closed 💰 Earn for open-source contributions like this. Gittensor lets GitHub contributors earn for the work they already do — register to start earning →. Checked by Gittensory, a quiet PR intelligence layer for OSS maintainers.
|
JSONbored
force-pushed
the
claude/close-hard-blocked-contributor-prs
branch
4 times, most recently
from
July 4, 2026 01:38
ea61c4b to
87a286b
Compare
VPS evidence (system_flags) showed closehold:<repo> engaged for all three review-active repos since 2026-06-28, silently downgrading every heuristic close to a human hold regardless of autonomy.close=auto. Three compounding, generic self-host engine bugs: - The close-precision circuit breaker treated every heuristic close identically, including ones backed by concrete, non-judgment evidence (a committed secret, red CI, a base conflict, a deterministic linked-issue duplicate, a dual-model AI consensus defect). Added closeConcreteEvidence, round-tripped through the approval-queue staging path, and scoped downgradeCloseToHold to exempt it - a hard blocker now closes even while the breaker is engaged, matching every other deterministic close kind. - The self-tune breaker's auto-clear tick only reconsidered projects present in the current gate-eval report, but a project whose closes are 100% suppressed stops producing new decided samples for that action class and can drop out of the report entirely - stranding its flag engaged forever regardless of cooldown. Widened the auto-clear candidate set to the union of the eval report and every currently-engaged per-project flag. - The eval read across every review_audit source, including the frozen, no-longer-written 'reviewbot' source from before the gittensory-native convergence. A dead system's historical close predictions could permanently anchor a live repo's measured close precision. Scoped the tick to source='gittensory-native'. Also adds a bounded-cardinality gittensory_precision_breaker_downgrades_total counter so an engaged breaker is visible without querying review_audit directly.
…tric miscount Gate review (round 2) found two reachable defects in the prior commit: - ai_consensus_defect was wrongly classified as closeConcreteEvidence, letting a green PR with only an AI-derived blocker bypass the close-precision breaker it exists to check. Two independent models agreeing is still a judgment call, not deterministic evidence - removed it from CONCRETE_EVIDENCE_BLOCKER_CODES, and hardened hasConcreteCloseEvidence to also exclude anything in advisory.ts's own AI_JUDGMENT_BLOCKER_CODES so this can't silently regress again. - precisionBreakerDowngradeDirections detected a close downgrade by checking whether any close action remained in the post-breaker plan, so a plan with a kept deterministic close plus a dropped heuristic close recorded no downgrade. Switched to reference-identity survival checks (does this specific planned action still exist in the breaker's output), which correctly handles multiple close actions. Also reuses advisory.ts's/pre-merge-checks.ts's existing exported code constants for two of the nine concrete-evidence codes instead of retyping them, and adds a source-text parity test guarding the remaining seven against silent producer-side drift.
… fix npm run test:coverage (and CI's validate/validate-code) failed with "DUPLICATE_ONLY_BLOCKER_CODES is not iterable": agent-actions.ts sits inside a genuine module-load cycle (scoring/model.ts -> db/repositories.ts -> agent-actions.ts -> rules/advisory.ts -> scoring/preview.ts -> scoring/model.ts), and eagerly spreading another module's export into a top-level array literal reads it before that module has necessarily finished initializing on the cycle's first pass. Reverted the two CONCRETE_EVIDENCE_BLOCKER_CODES entries back to plain string literals and extended the parity test to cover all nine codes instead of seven. AI_JUDGMENT_BLOCKER_CODES stays imported - it is only read inside a function body, not at module-eval time, so it isn't exposed to the same hazard.
codecov/patch flagged src/queue/processors.ts's incr() call inside runAgentMaintenancePlanAndExecute as uncovered: precisionBreakerDowngradeDirections and applyPrecisionBreakers were already exercised as pure functions in precision-breakers-chain.test.ts, but nothing drove the LIVE integration path (a real webhook, a real engaged breaker, a real withheld mutation) through this exact call site. Added an end-to-end regression alongside the existing convergence-chain test, seeding a holdonly breaker flag before the CI-completion re-review step and asserting both the withheld merge and the resulting counter.
…dProjectScopes codecov/patch flagged an else-if arm as a partial branch: the query's own WHERE clause (key LIKE 'holdonly:%' OR 'closehold:%') guarantees prefix can only ever be "holdonly" or "closehold", so the second === check could never see any other value. Collapsed to a plain else -- same behavior, no unreachable branch left to cover.
…r, address gate nits The gate's deterministic secret scanner flagged reason: "leaked secret" / "secret leaked" test fixtures as a possible generic_secret_assignment hit - no real credential, just wording that matched the heuristic's shape. Reworded to "hard blocker", which the tests don't depend on semantically. Also addresses the round-3 review nits: - Added a dedicated regression test that forces agent-actions.ts and scoring/model.ts to load together in the same module graph, so a reintroduced eager cross-module read on the documented load cycle fails a test directly instead of only being caught incidentally by unrelated suites. - Tightened the CONCRETE_EVIDENCE_BLOCKER_CODES parity test to require the actual producer assignment shape (code: "..." or a SOME_CONST = "..." export) rather than the bare literal appearing anywhere in the file, which a stale comment could have satisfied. - Documented the maintenance path in listEngagedProjectScopes's plain else: if the WHERE clause ever grows a third prefix, it must go back to an explicit branch with its own test, not stay silently bucketed as closehold.
JSONbored
force-pushed
the
claude/close-hard-blocked-contributor-prs
branch
from
July 4, 2026 01:53
87a286b to
400e03d
Compare
…gression The deterministic secret scanner's generic_secret_assignment pattern matches any `token: "..."` (or api_key/secret/password/etc.) assignment whose value is 16+ characters and doesn't look like a known placeholder string. The new holdonly-breaker test's hand-rolled fetch stub used "installation-token" (18 chars, no placeholder markers) in an ADDED line, which the scanner flags regardless of the hundreds of pre-existing, unchanged occurrences of that same literal elsewhere in this file. Shortened to a value under the length threshold; the token's actual content was never asserted on.
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.
Summary
system_flagson edge-us-01) showedclosehold:<repo>engaged for all three review-active repos since 2026-06-28 07:00:17, silently downgrading every heuristic close to a human hold — regardless ofautonomy.close=auto,agent_dry_run=0, and correct repo config. Investigated the full gate-to-action decision path (planAgentMaintenanceActions→applyPrecisionBreakers→executeAgentMaintenanceActions) and found three compounding, generic self-host engine bugs (no repo-specific logic anywhere in this diff):downgradeCloseToHolddowngraded everycloseKind: "heuristic"close identically, including ones backed by concrete, non-judgment evidence — a committed secret, red CI, a base conflict, or a deterministic linked-issue-overlap duplicate. AddedcloseConcreteEvidencetoPlannedAgentAction, round-tripped it through the approval-queue staging path (AgentPendingActionParams,actionParams), and scoped the breaker to exempt it — a hard blocker now closes even while the breaker is engaged, matching the existing exemption already given toblacklist/contributor_cap/review_nag/linked-issue-hard-rulecloses. Bothai_consensus_defect(a dual-model AI verdict) andai_review_splitstay fully subject to the breaker — the gate review (round 2) correctly flagged an earlier version of this diff for treating a dual-model AI consensus as "concrete": two models agreeing is still a judgment call, not deterministic evidence, and a systematically-wrong AI-driven close is exactly the failure mode this breaker exists to catch.hasConcreteCloseEvidencenow also defensively excludes anything in advisory.ts's ownAI_JUDGMENT_BLOCKER_CODES, so this can't silently regress again.runSelfTuneBreakeronly reconsidered projects present in the currentcomputeGateEvalreport, but a project whose closes are 100% suppressed stops producing new decided samples for that action class and can drop out of the report entirely — stranding its flag engaged forever, no matter how long the 24h cooldown has elapsed. Widened the auto-clear candidate set to the union of the eval report and every currently-engaged per-project flag (still never touches a human-set:globalflag).computeGateEvalread across everyreview_auditsource, including the frozen, no-longer-writtenreviewbotsource from before the gittensory-native convergence (verified on the VPS: 256 oldclosepredictions vs. zero fromgittensory-native, which only ever recordsmerge/holdfor its own gate check). A dead system's historical close predictions could permanently anchor a live repo's measured close precision, so it could never legitimately "recover." Scoped the tick tosource: 'gittensory-native'.gittensory_precision_breaker_downgrades_total{direction}counter (extracted as a pure, directly-testedprecisionBreakerDowngradeDirectionshelper) so an engaged breaker is visible without queryingreview_auditdirectly — no raw SHAs/PR numbers/free text.precisionBreakerDowngradeDirectionscompares each planned merge/close action's survival by reference identity (not "is any action of that class still present anywhere"), so a plan carrying both a kept deterministic close and a dropped heuristic close is still correctly counted as a close downgrade — another defect the round-2 gate review caught in the first version of this metric.last_regated_at/no current-head gate outcome) is already adequately covered by the existing hourly per-repo sweep plus the foreground-liveness/reconciliation-repair work already merged in fix(selfhost): keep review queue live and verify configured CI #2843/fix(queue): prioritize live PR work under GitHub rate pressure #2847 — no additional code needed there.CONCRETE_EVIDENCE_BLOCKER_CODESnow reuse advisory.ts's/pre-merge-checks.ts's existing exported constants (DUPLICATE_ONLY_BLOCKER_CODES,PRE_MERGE_CHECK_BLOCKING_CODE) instead of retyping the literal; the remaining seven have no single canonical export to import (each is either module-private or a literal duplicated across several unrelated producer files), so a new source-text parity test guards them against silent producer-side drift instead.Scope
src/settings/agent-actions.ts,src/types.ts,src/services/agent-action-executor.ts,src/review/outcomes-wire.ts,src/queue/processors.ts,src/selfhost/metrics.ts+ their test files onlyValidation
git diff --checknpm run typecheckagent-actions.test.ts,agent-action-executor.test.ts,outcomes-wire.test.ts,precision-breakers-chain.test.ts,agent-approval-queue.test.ts,auto-tune.test.ts— 415/415 passingnpm audit --audit-level=moderate— 0 vulnerabilitiesnpm run test:coverage(full unsharded) — not re-run locally beyond the directly-affected suite; CI runs the full gate unshardedIf any required check was skipped, explain why:
test:coverage/test:workers/UI checks were not re-run locally beyond the directly-affected suite, per this repo's own guidance that CI already runs the full gate unsharded.Safety
Notes
SELECTqueries againstsystem_flags/review_auditover the existing Tailscale SSH access) is not reproduced here — only the code fix and its reasoning.