akm-cli 0.9.0-beta.6 → 0.9.0-rc.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +663 -0
- package/README.md +12 -4
- package/dist/akm +38 -0
- package/dist/akm-migrate-storage +38 -0
- package/dist/assets/help/help-improve.md +9 -6
- package/dist/assets/hints/cli-hints-full.md +6 -5
- package/dist/assets/profiles/default.json +9 -4
- package/dist/assets/profiles/frequent.json +1 -1
- package/dist/assets/profiles/memory-focus.json +1 -1
- package/dist/assets/profiles/proactive-maintenance.json +25 -0
- package/dist/assets/profiles/quick.json +1 -1
- package/dist/assets/profiles/recombine-only.json +21 -0
- package/dist/assets/profiles/reflect-distill.json +30 -0
- package/dist/assets/profiles/synthesize.json +15 -0
- package/dist/assets/profiles/thorough.json +1 -1
- package/dist/assets/prompts/consolidate-system.md +23 -0
- package/dist/assets/prompts/contradiction-judge.md +33 -0
- package/dist/assets/prompts/distill-knowledge-system.md +22 -0
- package/dist/assets/prompts/distill-lesson-system.md +36 -0
- package/dist/assets/prompts/extract-session.md +11 -3
- package/dist/assets/prompts/graph-extract-system.md +1 -0
- package/dist/assets/prompts/graph-extract-user-prompt.md +1 -1
- package/dist/assets/prompts/memory-infer-system.md +1 -0
- package/dist/assets/prompts/memory-infer-user.md +5 -0
- package/dist/assets/prompts/metadata-enhance-system.md +1 -0
- package/dist/assets/prompts/procedural-system.md +44 -0
- package/dist/assets/prompts/recombine-system.md +40 -0
- package/dist/assets/prompts/staleness-detect-system.md +6 -0
- package/dist/assets/prompts/validate-summary-judge.md +1 -0
- package/dist/assets/stash-skeleton/facts/conventions/assets/agent.md +38 -0
- package/dist/assets/stash-skeleton/facts/conventions/assets/command.md +38 -0
- package/dist/assets/stash-skeleton/facts/conventions/assets/fact.md +39 -0
- package/dist/assets/stash-skeleton/facts/conventions/assets/knowledge.md +40 -0
- package/dist/assets/stash-skeleton/facts/conventions/assets/lesson.md +43 -0
- package/dist/assets/stash-skeleton/facts/conventions/assets/memory.md +38 -0
- package/dist/assets/stash-skeleton/facts/conventions/assets/script.md +43 -0
- package/dist/assets/stash-skeleton/facts/conventions/assets/skill.md +40 -0
- package/dist/assets/stash-skeleton/facts/conventions/assets/workflow.md +43 -0
- package/dist/assets/templates/html/health.html +281 -111
- package/dist/assets/wiki/ingest-workflow-template.md +45 -16
- package/dist/assets/wiki/schema-template.md +4 -4
- package/dist/cli/clack.js +56 -0
- package/dist/cli/config-migrate.js +7 -1
- package/dist/cli/confirm.js +1 -1
- package/dist/cli/parse-args.js +46 -1
- package/dist/cli/shared.js +28 -0
- package/dist/cli.js +25 -14
- package/dist/commands/agent/agent-dispatch.js +3 -2
- package/dist/commands/agent/agent-support.js +0 -7
- package/dist/commands/agent/contribute-cli.js +26 -7
- package/dist/commands/config-cli.js +26 -13
- package/dist/commands/env/child-env.js +47 -0
- package/dist/commands/env/env-cli.js +220 -227
- package/dist/commands/env/env.js +14 -67
- package/dist/commands/env/secret-cli.js +140 -138
- package/dist/commands/feedback-cli.js +153 -147
- package/dist/commands/graph/graph-cli.js +5 -13
- package/dist/commands/graph/graph.js +76 -72
- package/dist/commands/health/advisories.js +151 -0
- package/dist/commands/health/checks.js +103 -16
- package/dist/commands/health/html-report.js +447 -81
- package/dist/commands/health/improve-metrics.js +771 -0
- package/dist/commands/health/llm-usage.js +65 -0
- package/dist/commands/health/md-report.js +103 -0
- package/dist/commands/health/metrics.js +278 -0
- package/dist/commands/health/stash-exposure.js +46 -0
- package/dist/commands/health/surfaces.js +216 -0
- package/dist/commands/health/task-runs.js +135 -0
- package/dist/commands/health/types.js +26 -0
- package/dist/commands/health/windows.js +195 -0
- package/dist/commands/health.js +91 -1083
- package/dist/commands/improve/anti-collapse.js +170 -0
- package/dist/commands/improve/calibration.js +161 -0
- package/dist/commands/improve/collapse-detector.js +421 -0
- package/dist/commands/improve/consolidate/chunking.js +141 -0
- package/dist/commands/improve/consolidate/eligibility.js +64 -0
- package/dist/commands/improve/consolidate/merge.js +145 -0
- package/dist/commands/improve/consolidate/sanitize.js +231 -0
- package/dist/commands/{lint.js → improve/consolidate/types.js} +1 -1
- package/dist/commands/improve/consolidate.js +1313 -1278
- package/dist/commands/improve/dedup.js +482 -0
- package/dist/commands/improve/distill/content-repair.js +202 -0
- package/dist/commands/improve/distill/promote-memory.js +229 -0
- package/dist/commands/improve/distill/quality-gate.js +236 -0
- package/dist/commands/improve/distill-guards.js +127 -0
- package/dist/commands/improve/distill-promotion-policy.js +826 -167
- package/dist/commands/improve/distill.js +243 -599
- package/dist/commands/improve/eligibility.js +434 -0
- package/dist/commands/improve/encoding-salience.js +205 -0
- package/dist/commands/improve/extract-cli.js +179 -59
- package/dist/commands/improve/extract-prompt.js +55 -4
- package/dist/commands/improve/extract-watch.js +140 -0
- package/dist/commands/improve/extract.js +409 -43
- package/dist/commands/improve/feedback-valence.js +54 -0
- package/dist/commands/improve/hot-probation.js +45 -0
- package/dist/commands/improve/improve-auto-accept.js +160 -7
- package/dist/commands/improve/improve-cli.js +115 -73
- package/dist/commands/improve/improve-profiles.js +32 -8
- package/dist/commands/improve/improve-result-file.js +15 -25
- package/dist/commands/improve/improve-session.js +58 -0
- package/dist/commands/improve/improve.js +510 -2537
- package/dist/commands/improve/locks.js +154 -0
- package/dist/commands/improve/loop-stages.js +1100 -0
- package/dist/commands/improve/memory/memory-belief.js +14 -15
- package/dist/commands/improve/memory/memory-contradiction-detect.js +83 -60
- package/dist/commands/improve/memory/memory-improve.js +27 -27
- package/dist/commands/improve/outcome-loop.js +270 -0
- package/dist/commands/improve/preparation.js +2002 -0
- package/dist/commands/improve/proactive-maintenance.js +115 -0
- package/dist/commands/improve/procedural.js +398 -0
- package/dist/commands/improve/recombine.js +818 -0
- package/dist/commands/improve/reflect-noise.js +0 -0
- package/dist/commands/improve/reflect.js +212 -45
- package/dist/commands/improve/salience.js +455 -0
- package/dist/commands/improve/schema-similarity-gate.js +168 -0
- package/dist/commands/improve/shared.js +51 -0
- package/dist/commands/improve/triage.js +93 -0
- package/dist/commands/lint/agent-linter.js +19 -24
- package/dist/commands/lint/base-linter.js +173 -60
- package/dist/commands/lint/command-linter.js +19 -24
- package/dist/commands/lint/env-key-rules.js +38 -1
- package/dist/commands/lint/fact-linter.js +39 -0
- package/dist/commands/lint/index.js +31 -13
- package/dist/commands/lint/memory-linter.js +1 -1
- package/dist/commands/lint/registry.js +7 -2
- package/dist/commands/lint/task-linter.js +3 -3
- package/dist/commands/lint/workflow-linter.js +26 -1
- package/dist/commands/observability-cli.js +4 -4
- package/dist/commands/proposal/drain-policies.js +13 -4
- package/dist/commands/proposal/drain.js +45 -51
- package/dist/commands/proposal/legacy-import.js +115 -0
- package/dist/commands/proposal/proposal-cli.js +24 -34
- package/dist/commands/proposal/proposal.js +7 -1
- package/dist/commands/proposal/propose.js +8 -3
- package/dist/commands/proposal/repository.js +829 -0
- package/dist/commands/proposal/validators/proposal-quality-validators.js +9 -8
- package/dist/commands/proposal/validators/proposals.js +93 -882
- package/dist/commands/read/curate.js +419 -103
- package/dist/commands/read/knowledge.js +10 -3
- package/dist/commands/read/remember-cli.js +133 -138
- package/dist/commands/read/search-cli.js +15 -8
- package/dist/commands/read/search.js +22 -11
- package/dist/commands/read/show.js +106 -14
- package/dist/commands/registry-cli.js +76 -87
- package/dist/commands/remember.js +11 -12
- package/dist/commands/sources/add-cli.js +91 -95
- package/dist/commands/sources/history.js +1 -1
- package/dist/commands/sources/init.js +66 -18
- package/dist/commands/sources/installed-stashes.js +11 -3
- package/dist/commands/sources/schema-repair.js +44 -46
- package/dist/commands/sources/self-update.js +2 -2
- package/dist/commands/sources/source-add.js +7 -3
- package/dist/commands/sources/sources-cli.js +3 -3
- package/dist/commands/sources/stash-cli.js +29 -41
- package/dist/commands/sources/stash-skeleton.js +57 -8
- package/dist/commands/tasks/default-tasks.js +15 -2
- package/dist/commands/tasks/tasks-cli.js +20 -29
- package/dist/commands/tasks/tasks.js +39 -11
- package/dist/commands/wiki-cli.js +23 -38
- package/dist/commands/workflow-cli.js +15 -1
- package/dist/core/asset/asset-registry.js +3 -1
- package/dist/core/asset/asset-spec.js +21 -4
- package/dist/core/asset/frontmatter.js +188 -167
- package/dist/core/asset/markdown.js +8 -0
- package/dist/core/authoring-rules.js +92 -0
- package/dist/core/common.js +4 -23
- package/dist/core/concurrent.js +10 -1
- package/dist/core/config/config-io.js +10 -1
- package/dist/core/config/config-migration.js +18 -40
- package/dist/core/config/config-schema.js +389 -58
- package/dist/core/config/config-types.js +3 -3
- package/dist/core/config/config.js +67 -22
- package/dist/core/deep-merge.js +38 -0
- package/dist/core/errors.js +1 -0
- package/dist/core/eval/rank-metrics.js +113 -0
- package/dist/core/events.js +4 -7
- package/dist/core/improve-types.js +47 -8
- package/dist/core/logs-db.js +14 -75
- package/dist/core/parse.js +36 -16
- package/dist/core/paths.js +21 -18
- package/dist/core/standards/resolve-standards-context.js +87 -0
- package/dist/core/standards/resolve-stash-standards.js +99 -0
- package/dist/core/standards/resolve-type-conventions.js +66 -0
- package/dist/core/state/migrations.js +770 -0
- package/dist/core/state-db.js +142 -1091
- package/dist/core/structured.js +69 -0
- package/dist/core/time.js +53 -0
- package/dist/core/warn.js +21 -0
- package/dist/core/write-source.js +37 -0
- package/dist/indexer/db/db.js +356 -780
- package/dist/indexer/db/entry-mapper.js +41 -0
- package/dist/indexer/db/graph-db.js +129 -86
- package/dist/indexer/db/llm-cache.js +2 -2
- package/dist/indexer/db/schema.js +516 -0
- package/dist/indexer/ensure-index.js +103 -24
- package/dist/indexer/feedback/utility-policy.js +75 -0
- package/dist/indexer/graph/graph-boost.js +51 -41
- package/dist/indexer/graph/graph-extraction.js +207 -4
- package/dist/indexer/index-writer-lock.js +106 -0
- package/dist/indexer/index-written-assets.js +105 -0
- package/dist/indexer/indexer.js +291 -310
- package/dist/indexer/passes/dir-staleness.js +114 -0
- package/dist/indexer/passes/memory-inference.js +13 -5
- package/dist/indexer/passes/metadata.js +20 -0
- package/dist/indexer/read-preflight.js +23 -0
- package/dist/indexer/search/db-search.js +89 -13
- package/dist/indexer/search/fts-query.js +51 -0
- package/dist/indexer/search/ranking-contributors.js +95 -9
- package/dist/indexer/search/ranking.js +79 -3
- package/dist/indexer/search/search-fields.js +6 -0
- package/dist/indexer/search/search-source.js +32 -21
- package/dist/indexer/search/semantic-status.js +4 -0
- package/dist/indexer/walk/matchers.js +9 -0
- package/dist/indexer/walk/walker.js +21 -13
- package/dist/integrations/agent/builders.js +39 -13
- package/dist/integrations/agent/config.js +20 -59
- package/dist/integrations/agent/detect.js +9 -0
- package/dist/integrations/agent/index.js +3 -19
- package/dist/integrations/agent/model-aliases.js +7 -2
- package/dist/integrations/agent/profiles.js +7 -1
- package/dist/integrations/agent/prompts.js +75 -9
- package/dist/integrations/agent/runner-dispatch.js +59 -0
- package/dist/integrations/agent/runner.js +13 -9
- package/dist/integrations/agent/spawn.js +69 -67
- package/dist/integrations/harnesses/claude/agent-builder.js +1 -1
- package/dist/integrations/harnesses/claude/index.js +2 -0
- package/dist/integrations/harnesses/claude/session-log.js +11 -1
- package/dist/integrations/harnesses/index.js +2 -3
- package/dist/integrations/harnesses/opencode/agent-builder.js +1 -1
- package/dist/integrations/harnesses/opencode/index.js +2 -0
- package/dist/integrations/harnesses/opencode/session-log.js +173 -3
- package/dist/integrations/harnesses/opencode-sdk/index.js +2 -2
- package/dist/integrations/harnesses/opencode-sdk/sdk-runner.js +98 -17
- package/dist/integrations/harnesses/types.js +1 -0
- package/dist/integrations/session-logs/index.js +16 -0
- package/dist/llm/call-ai.js +2 -2
- package/dist/llm/client.js +57 -15
- package/dist/llm/embedder.js +67 -4
- package/dist/llm/embedders/cache.js +3 -1
- package/dist/llm/embedders/deterministic.js +66 -0
- package/dist/llm/embedders/local.js +73 -3
- package/dist/llm/feature-gate.js +16 -15
- package/dist/llm/graph-extract.js +67 -44
- package/dist/llm/memory-infer-impl.js +138 -0
- package/dist/llm/memory-infer.js +1 -127
- package/dist/llm/metadata-enhance.js +44 -31
- package/dist/llm/structured-call.js +49 -0
- package/dist/migrate-storage-node.mjs +8 -0
- package/dist/output/context.js +5 -5
- package/dist/output/renderers.js +85 -14
- package/dist/output/shapes/curate.js +14 -2
- package/dist/output/shapes/helpers.js +0 -3
- package/dist/output/shapes/passthrough.js +2 -1
- package/dist/output/text/helpers.js +29 -1
- package/dist/output/text/workflow.js +1 -0
- package/dist/registry/providers/skills-sh.js +21 -147
- package/dist/registry/providers/static-index.js +15 -157
- package/dist/registry/resolve.js +27 -9
- package/dist/runtime.js +25 -1
- package/dist/scripts/migrate-storage.js +2718 -2354
- package/dist/scripts/migrations/import-fs-improve-runs-to-db.js +891 -597
- package/dist/setup/detect.js +9 -0
- package/dist/setup/legacy-config.js +106 -0
- package/dist/setup/prompt.js +57 -0
- package/dist/setup/providers.js +14 -0
- package/dist/setup/registry-stash-loader.js +12 -0
- package/dist/setup/semantic-assets.js +124 -0
- package/dist/setup/setup.js +52 -1614
- package/dist/setup/steps/connection.js +734 -0
- package/dist/setup/steps/output.js +31 -0
- package/dist/setup/steps/platforms.js +124 -0
- package/dist/setup/steps/semantic.js +27 -0
- package/dist/setup/steps/sources.js +222 -0
- package/dist/setup/steps/stashdir.js +42 -0
- package/dist/setup/steps/tasks.js +152 -0
- package/dist/sources/include.js +6 -2
- package/dist/sources/providers/filesystem.js +0 -1
- package/dist/sources/providers/git-install.js +210 -0
- package/dist/sources/providers/git-provider.js +234 -0
- package/dist/sources/providers/git-stash.js +248 -0
- package/dist/sources/providers/git.js +10 -661
- package/dist/sources/providers/npm.js +2 -6
- package/dist/sources/providers/provider-utils.js +13 -7
- package/dist/sources/providers/sync-from-ref.js +9 -1
- package/dist/sources/providers/tar-utils.js +16 -8
- package/dist/sources/providers/website.js +9 -5
- package/dist/sources/website-ingest.js +187 -29
- package/dist/sources/wiki-fetchers/registry.js +53 -0
- package/dist/sources/wiki-fetchers/youtube.js +239 -0
- package/dist/storage/database.js +45 -10
- package/dist/storage/managed-db.js +82 -0
- package/dist/storage/repositories/canaries-repository.js +107 -0
- package/dist/storage/repositories/consolidation-repository.js +38 -0
- package/dist/storage/repositories/embeddings-repository.js +72 -0
- package/dist/storage/repositories/events-repository.js +187 -0
- package/dist/storage/repositories/extract-sessions-repository.js +96 -0
- package/dist/storage/repositories/improve-runs-repository.js +146 -0
- package/dist/storage/repositories/index-db.js +14 -8
- package/dist/storage/repositories/proposals-repository.js +220 -0
- package/dist/storage/repositories/recombine-repository.js +213 -0
- package/dist/storage/repositories/registry-cache.js +93 -0
- package/dist/storage/repositories/registry-index-cache-repository.js +46 -0
- package/dist/storage/repositories/task-history-repository.js +93 -0
- package/dist/storage/sqlite-pragmas.js +146 -0
- package/dist/tasks/backends/cron.js +1 -1
- package/dist/tasks/backends/index.js +9 -0
- package/dist/tasks/backends/launchd.js +1 -1
- package/dist/tasks/backends/schtasks.js +1 -1
- package/dist/tasks/{resolveAkmBin.js → resolve-akm-bin.js} +2 -2
- package/dist/tasks/runner.js +15 -13
- package/dist/text-import-hook.mjs +0 -0
- package/dist/wiki/wiki.js +52 -11
- package/dist/workflows/cli.js +1 -0
- package/dist/workflows/db.js +3 -4
- package/dist/workflows/runtime/runs.js +43 -118
- package/dist/workflows/runtime/workflow-asset-loader.js +125 -0
- package/dist/workflows/validate-summary.js +2 -7
- package/docs/README.md +69 -18
- package/docs/data-and-telemetry.md +5 -4
- package/docs/migration/release-notes/0.7.0.md +1 -1
- package/docs/migration/release-notes/0.9.0.md +39 -0
- package/package.json +10 -10
- package/dist/assets/tasks/core/update-stashes.yml +0 -4
- package/dist/commands/db-cli.js +0 -23
- package/dist/indexer/db/db-backup.js +0 -376
- package/dist/indexer/passes/staleness-detect.js +0 -488
|
@@ -4,89 +4,46 @@
|
|
|
4
4
|
import fs from "node:fs";
|
|
5
5
|
import path from "node:path";
|
|
6
6
|
import { assertNever } from "../../core/assert.js";
|
|
7
|
-
import {
|
|
8
|
-
import {
|
|
9
|
-
import {
|
|
10
|
-
import { getDefaultLlmConfig, loadConfig } from "../../core/config/config.js";
|
|
11
|
-
import { ConfigError, NotFoundError, rethrowIfTestIsolationError, UsageError } from "../../core/errors.js";
|
|
7
|
+
import { daysToMs } from "../../core/common.js";
|
|
8
|
+
import { loadConfig } from "../../core/config/config.js";
|
|
9
|
+
import { rethrowIfTestIsolationError } from "../../core/errors.js";
|
|
12
10
|
import { appendEvent, readEvents } from "../../core/events.js";
|
|
13
|
-
import {
|
|
14
|
-
import { classifyImproveAction } from "../../core/improve-types.js";
|
|
15
|
-
import { openLogsDatabase, purgeOldTaskLogs } from "../../core/logs-db.js";
|
|
11
|
+
import { classifyImproveAction, foldDistillSkipped } from "../../core/improve-types.js";
|
|
16
12
|
import { getDbPath, getStateDbPathInDataDir } from "../../core/paths.js";
|
|
17
|
-
import { openStateDatabase
|
|
13
|
+
import { openStateDatabase } from "../../core/state-db.js";
|
|
18
14
|
import { info, warn } from "../../core/warn.js";
|
|
19
|
-
import { closeDatabase,
|
|
15
|
+
import { closeDatabase, getEntryCount, openExistingDatabase } from "../../indexer/db/db.js";
|
|
20
16
|
import { ensureIndex } from "../../indexer/ensure-index.js";
|
|
21
|
-
import { runGraphExtractionPass } from "../../indexer/graph/graph-extraction.js";
|
|
22
17
|
import { akmIndex } from "../../indexer/indexer.js";
|
|
23
|
-
import {
|
|
24
|
-
import {
|
|
25
|
-
import { getWritableStashDirs, resolveSourceEntries } from "../../indexer/search/search-source.js";
|
|
26
|
-
import { countUsageEventsByType } from "../../indexer/usage/usage-events.js";
|
|
27
|
-
import { resolveAssetPath } from "../../indexer/walk/path-resolver.js";
|
|
28
|
-
import { resolveImproveProcessRunnerFromProfile, resolveTriageJudgmentRunner } from "../../integrations/agent/runner.js";
|
|
29
|
-
import { getAvailableHarnesses } from "../../integrations/session-logs/index.js";
|
|
30
|
-
import { isLlmFeatureEnabled, isProcessEnabled } from "../../llm/feature-gate.js";
|
|
18
|
+
import { resolveSourceEntries } from "../../indexer/search/search-source.js";
|
|
19
|
+
import { resolveTriageJudgmentRunner } from "../../integrations/agent/runner.js";
|
|
31
20
|
import { installLlmUsagePersistence } from "../../llm/usage-persist.js";
|
|
32
21
|
import { withLlmStage } from "../../llm/usage-telemetry.js";
|
|
33
22
|
import { isGitBackedStash, resolveWritableOverride, saveGitStash } from "../../sources/providers/git.js";
|
|
34
|
-
import { akmLint } from "../lint/index.js";
|
|
35
23
|
import { drainProposals } from "../proposal/drain.js";
|
|
36
24
|
import { resolveDrainPolicy } from "../proposal/drain-policies.js";
|
|
37
|
-
import {
|
|
38
|
-
|
|
39
|
-
import {
|
|
40
|
-
import {
|
|
41
|
-
import {
|
|
42
|
-
|
|
43
|
-
|
|
44
|
-
import {
|
|
45
|
-
|
|
46
|
-
import {
|
|
25
|
+
import { akmDistill } from "./distill.js";
|
|
26
|
+
// Eligibility / candidate-selection predicates live in ./eligibility.
|
|
27
|
+
import { buildLatestProposalTsMap, collectEligibleRefs, memoryCleanupParentRef, resolveImproveScope, shouldAnalyzeMemoryCleanup, } from "./eligibility.js";
|
|
28
|
+
import { countEvalCases } from "./eval-cases.js";
|
|
29
|
+
import { resolveImproveProfile, resolveProcessEnabled } from "./improve-profiles.js";
|
|
30
|
+
// #607 per-process lock primitives live in ./locks. Imported for internal use;
|
|
31
|
+
// resetHeldProcessLocks is re-exported (the test seam imports it from here).
|
|
32
|
+
import { PROCESS_LOCK_DEFS, processLockPath, releaseAllProcessLocks, releaseHeldLocksIfOwned, releaseProcessLock, tryAcquireProcessLock, withOptionalProcessLock, } from "./locks.js";
|
|
33
|
+
// The cycle loop / post-loop / maintenance stages live in ./loop-stages.
|
|
34
|
+
import { runImproveLoopStage, runImprovePostLoopStage } from "./loop-stages.js";
|
|
47
35
|
import { detectAndWriteContradictions } from "./memory/memory-contradiction-detect.js";
|
|
48
|
-
import { analyzeMemoryCleanup
|
|
36
|
+
import { analyzeMemoryCleanup } from "./memory/memory-improve.js";
|
|
37
|
+
// The pre-loop preparation pipeline lives in ./preparation.
|
|
38
|
+
import { runImprovePreparationStage } from "./preparation.js";
|
|
39
|
+
import { DEFAULT_DUE_DAYS, filterProactiveDue } from "./proactive-maintenance.js";
|
|
49
40
|
import { akmReflect } from "./reflect.js";
|
|
50
|
-
|
|
51
|
-
|
|
52
|
-
|
|
53
|
-
|
|
54
|
-
|
|
55
|
-
|
|
56
|
-
return { mode: "ref", value: trimmed };
|
|
57
|
-
}
|
|
58
|
-
catch {
|
|
59
|
-
if (!isAssetType(trimmed)) {
|
|
60
|
-
throw new UsageError(`Unknown asset type: "${trimmed}". Valid types: memory, knowledge, skill, lesson, workflow, agent, command, script, wiki, env, secret, task.\n` +
|
|
61
|
-
`If you passed --format to akm improve, that flag is not supported — use it with akm search or akm show instead.`, "INVALID_FLAG_VALUE");
|
|
62
|
-
}
|
|
63
|
-
return { mode: "type", value: trimmed };
|
|
64
|
-
}
|
|
65
|
-
}
|
|
66
|
-
/**
|
|
67
|
-
* Render the end-of-run stash-sync commit message, expanding `{token}`
|
|
68
|
-
* placeholders against this run's results. Unknown tokens are passed through
|
|
69
|
-
* verbatim so adding new tokens later never breaks an existing template, and so
|
|
70
|
-
* a literal brace in a message is harmless.
|
|
71
|
-
*
|
|
72
|
-
* Supported tokens (the "free" set — derived from data already on the result):
|
|
73
|
-
* {timestamp} `YYYY-MM-DD HH:MM:SS` (UTC)
|
|
74
|
-
* {date} `YYYY-MM-DD` (UTC)
|
|
75
|
-
* {time} `HH:MM:SS` (UTC)
|
|
76
|
-
* {scope} scope value (e.g. a ref/type) or the scope mode (`all`)
|
|
77
|
-
* {refs} number of planned refs this run processed
|
|
78
|
-
* {accepted} number of proposals auto-accepted by the confidence gate
|
|
79
|
-
* {triage_promoted} proposals promoted by the triage pre-pass (0 if triage did not run)
|
|
80
|
-
* {triage_rejected} proposals rejected by the triage pre-pass (0 if triage did not run)
|
|
81
|
-
* {runId} this run's id (empty string when absent)
|
|
82
|
-
*
|
|
83
|
-
* The result is still passed through `sanitizeCommitMessage` downstream in
|
|
84
|
-
* `saveGitStash`, so token values never widen the commit-message attack surface
|
|
85
|
-
* (newlines/control chars are collapsed there).
|
|
86
|
-
*
|
|
87
|
-
* `nowMs` is injected (not read from `Date.now()`) so the function is pure and
|
|
88
|
-
* deterministically testable.
|
|
89
|
-
*/
|
|
41
|
+
import { errMessage } from "./shared.js";
|
|
42
|
+
export { resetHeldProcessLocks } from "./locks.js";
|
|
43
|
+
// Re-exported from ./loop-stages for test importers (improve-db-locking).
|
|
44
|
+
export { runImproveMaintenancePasses } from "./loop-stages.js";
|
|
45
|
+
// Re-exported from ./preparation so existing importers (tests, callers) resolve.
|
|
46
|
+
export { maybeAutoTuneThreshold } from "./preparation.js";
|
|
90
47
|
export function renderSyncCommitMessage(template, result, nowMs) {
|
|
91
48
|
const iso = new Date(nowMs).toISOString();
|
|
92
49
|
const tokens = {
|
|
@@ -102,334 +59,6 @@ export function renderSyncCommitMessage(template, result, nowMs) {
|
|
|
102
59
|
};
|
|
103
60
|
return template.replace(/\{(\w+)\}/g, (match, key) => (Object.hasOwn(tokens, key) ? tokens[key] : match));
|
|
104
61
|
}
|
|
105
|
-
async function collectEligibleRefs(scope, stashDir, improveProfile) {
|
|
106
|
-
if (scope.mode === "ref" && scope.value) {
|
|
107
|
-
const parsed = parseAssetRef(scope.value);
|
|
108
|
-
const writableDirs = new Set(getWritableStashDirs(stashDir).map((dir) => path.resolve(dir)));
|
|
109
|
-
const filePath = await findAssetFilePath(scope.value, stashDir, writableDirs);
|
|
110
|
-
if (!filePath) {
|
|
111
|
-
return {
|
|
112
|
-
plannedRefs: [],
|
|
113
|
-
memorySummary: { eligible: 0, derived: 0 },
|
|
114
|
-
profileFilteredRefs: [],
|
|
115
|
-
};
|
|
116
|
-
}
|
|
117
|
-
return {
|
|
118
|
-
plannedRefs: [{ ref: scope.value, reason: "scope-ref", filePath }],
|
|
119
|
-
memorySummary: {
|
|
120
|
-
eligible: parsed.type === "memory" ? 1 : 0,
|
|
121
|
-
derived: parsed.type === "memory" && parsed.name.endsWith(".derived") ? 1 : 0,
|
|
122
|
-
},
|
|
123
|
-
profileFilteredRefs: [],
|
|
124
|
-
};
|
|
125
|
-
}
|
|
126
|
-
let sources;
|
|
127
|
-
try {
|
|
128
|
-
sources = resolveSourceEntries(stashDir);
|
|
129
|
-
}
|
|
130
|
-
catch {
|
|
131
|
-
return { plannedRefs: [], memorySummary: { eligible: 0, derived: 0 }, profileFilteredRefs: [] };
|
|
132
|
-
}
|
|
133
|
-
if (sources.length === 0) {
|
|
134
|
-
return { plannedRefs: [], memorySummary: { eligible: 0, derived: 0 }, profileFilteredRefs: [] };
|
|
135
|
-
}
|
|
136
|
-
// Only operate on writable sources — never mutate read-only registry caches
|
|
137
|
-
// or remote stashes that the user did not mark writable.
|
|
138
|
-
let writableDirs;
|
|
139
|
-
try {
|
|
140
|
-
writableDirs = getWritableStashDirs(stashDir);
|
|
141
|
-
}
|
|
142
|
-
catch {
|
|
143
|
-
writableDirs = sources.slice(0, 1).map((s) => s.path); // fallback: primary only
|
|
144
|
-
}
|
|
145
|
-
const writableDirSet = new Set(writableDirs.map((d) => path.resolve(d)));
|
|
146
|
-
let db;
|
|
147
|
-
try {
|
|
148
|
-
db = openExistingDatabase();
|
|
149
|
-
const entries = getAllEntries(db, scope.mode === "type" ? scope.value : undefined).filter((indexed) => {
|
|
150
|
-
// First apply the existing stashDir-scope filter (no-op when stashDir is unset).
|
|
151
|
-
if (!isEntryInScope(indexed.stashDir, indexed.filePath, stashDir))
|
|
152
|
-
return false;
|
|
153
|
-
// Then restrict to writable sources only.
|
|
154
|
-
return isEntryInWritableSource(indexed.stashDir, indexed.filePath, writableDirSet);
|
|
155
|
-
});
|
|
156
|
-
const planned = new Map();
|
|
157
|
-
const profileFiltered = new Map();
|
|
158
|
-
let memoryEligible = 0;
|
|
159
|
-
let memoryDerived = 0;
|
|
160
|
-
for (const indexed of entries) {
|
|
161
|
-
const ref = makeAssetRef(indexed.entry.type, indexed.entry.name);
|
|
162
|
-
const isDerived = indexed.entry.name.endsWith(".derived");
|
|
163
|
-
// `.derived` memories are LLM-inferred and intentionally skip reflect
|
|
164
|
-
// (see the synthetic `derived-memory-reflect-skipped` branch in the
|
|
165
|
-
// improve loop). Enqueueing them here just produced one synthetic skip
|
|
166
|
-
// per derived memory per hour with no real work — pure churn observed
|
|
167
|
-
// 2026-05-21: 11 derived refs re-planned every hour during idle periods.
|
|
168
|
-
// The cleanup phase (analyzeMemoryCleanup) inspects derived memories
|
|
169
|
-
// independently of `plannedRefs`, so dropping them here loses nothing.
|
|
170
|
-
if (!isDerived && !planned.has(ref) && !profileFiltered.has(ref)) {
|
|
171
|
-
// 2026-05-27: extend the .derived precedent to profile-incompatible
|
|
172
|
-
// refs. If every per-ref pass (reflect + distill) on the active
|
|
173
|
-
// profile would refuse this ref, drop it from `plannedRefs`. The
|
|
174
|
-
// caller emits `improve_skipped { reason: profile_filtered_all_passes }`
|
|
175
|
-
// once `eventsCtx` is available so the audit trail is preserved in a
|
|
176
|
-
// single event per ref instead of 2× synthetic actions per run.
|
|
177
|
-
// Background: see /tmp/akm-health-investigations/planner-profile-metrics-deep-analysis.md
|
|
178
|
-
if (improveProfile && isProfileFilteredForAllPasses(ref, improveProfile)) {
|
|
179
|
-
profileFiltered.set(ref, {
|
|
180
|
-
ref,
|
|
181
|
-
reason: "profile_filtered_all_passes",
|
|
182
|
-
filePath: indexed.filePath,
|
|
183
|
-
});
|
|
184
|
-
}
|
|
185
|
-
else {
|
|
186
|
-
planned.set(ref, {
|
|
187
|
-
ref,
|
|
188
|
-
reason: scope.mode === "type" ? "scope-type" : indexed.entry.type === "memory" ? "memory-cleanup" : "scope-type",
|
|
189
|
-
filePath: indexed.filePath,
|
|
190
|
-
});
|
|
191
|
-
}
|
|
192
|
-
}
|
|
193
|
-
if (indexed.entry.type === "memory") {
|
|
194
|
-
memoryEligible += 1;
|
|
195
|
-
if (isDerived)
|
|
196
|
-
memoryDerived += 1;
|
|
197
|
-
}
|
|
198
|
-
}
|
|
199
|
-
return {
|
|
200
|
-
plannedRefs: [...planned.values()],
|
|
201
|
-
memorySummary: { eligible: memoryEligible, derived: memoryDerived },
|
|
202
|
-
profileFilteredRefs: [...profileFiltered.values()],
|
|
203
|
-
};
|
|
204
|
-
}
|
|
205
|
-
catch (error) {
|
|
206
|
-
// The bun-test isolation guard must never be downgraded to "empty plan".
|
|
207
|
-
rethrowIfTestIsolationError(error);
|
|
208
|
-
if (error instanceof NotFoundError || error instanceof Error) {
|
|
209
|
-
return { plannedRefs: [], memorySummary: { eligible: 0, derived: 0 }, profileFilteredRefs: [] };
|
|
210
|
-
}
|
|
211
|
-
throw error;
|
|
212
|
-
}
|
|
213
|
-
finally {
|
|
214
|
-
if (db)
|
|
215
|
-
closeDatabase(db);
|
|
216
|
-
}
|
|
217
|
-
}
|
|
218
|
-
function isEntryInScope(entryStashDir, filePath, stashDir) {
|
|
219
|
-
if (!stashDir)
|
|
220
|
-
return true;
|
|
221
|
-
const resolvedEntryStashDir = path.resolve(entryStashDir);
|
|
222
|
-
const resolvedFilePath = path.resolve(filePath);
|
|
223
|
-
const resolvedScopeStashDir = path.resolve(stashDir);
|
|
224
|
-
return (resolvedEntryStashDir === resolvedScopeStashDir ||
|
|
225
|
-
resolvedEntryStashDir.startsWith(`${resolvedScopeStashDir}${path.sep}`) ||
|
|
226
|
-
resolvedFilePath.startsWith(`${resolvedScopeStashDir}${path.sep}`));
|
|
227
|
-
}
|
|
228
|
-
/**
|
|
229
|
-
* Return true when the indexed entry belongs to one of the writable source
|
|
230
|
-
* directories. Entries from read-only registry caches or remote stashes that
|
|
231
|
-
* the user has not marked writable must never enter the improve/distill loop.
|
|
232
|
-
*/
|
|
233
|
-
function isEntryInWritableSource(entryStashDir, filePath, writableDirSet) {
|
|
234
|
-
const resolvedEntryStashDir = path.resolve(entryStashDir);
|
|
235
|
-
const resolvedFilePath = path.resolve(filePath);
|
|
236
|
-
for (const writableDir of writableDirSet) {
|
|
237
|
-
if (resolvedEntryStashDir === writableDir ||
|
|
238
|
-
resolvedEntryStashDir.startsWith(`${writableDir}${path.sep}`) ||
|
|
239
|
-
resolvedFilePath.startsWith(`${writableDir}${path.sep}`)) {
|
|
240
|
-
return true;
|
|
241
|
-
}
|
|
242
|
-
}
|
|
243
|
-
return false;
|
|
244
|
-
}
|
|
245
|
-
function memoryCleanupParentRef(scope, stashDir) {
|
|
246
|
-
if (scope.mode !== "ref" || !scope.value)
|
|
247
|
-
return undefined;
|
|
248
|
-
const parsed = parseAssetRef(scope.value);
|
|
249
|
-
if (parsed.type !== "memory")
|
|
250
|
-
return undefined;
|
|
251
|
-
if (!parsed.name.endsWith(".derived"))
|
|
252
|
-
return scope.value;
|
|
253
|
-
const sources = resolveSourceEntries(stashDir);
|
|
254
|
-
for (const source of sources) {
|
|
255
|
-
const candidate = path.join(source.path, "memories", `${parsed.name}.md`);
|
|
256
|
-
if (!fs.existsSync(candidate))
|
|
257
|
-
continue;
|
|
258
|
-
const raw = fs.readFileSync(candidate, "utf8");
|
|
259
|
-
const fm = parseFrontmatter(raw).data;
|
|
260
|
-
const sourceRef = typeof fm.source === "string" ? fm.source : undefined;
|
|
261
|
-
if (sourceRef) {
|
|
262
|
-
try {
|
|
263
|
-
const parent = parseAssetRef(sourceRef.trim());
|
|
264
|
-
if (parent.type === "memory")
|
|
265
|
-
return makeAssetRef(parent.type, parent.name);
|
|
266
|
-
}
|
|
267
|
-
catch { }
|
|
268
|
-
}
|
|
269
|
-
}
|
|
270
|
-
return makeAssetRef("memory", parsed.name.slice(0, -".derived".length));
|
|
271
|
-
}
|
|
272
|
-
function isLessonCandidate(ref) {
|
|
273
|
-
// Only lesson assets need lesson-schema validation (description + when_to_use).
|
|
274
|
-
// Memories have their own distill path via shouldDistillMemoryRef.
|
|
275
|
-
// All other types go through reflect, not distill.
|
|
276
|
-
return parseAssetRef(ref).type === "lesson";
|
|
277
|
-
}
|
|
278
|
-
/**
|
|
279
|
-
* Planner-side check: should this ref enter the distill queue?
|
|
280
|
-
*
|
|
281
|
-
* Distill produces lessons from non-lesson sources. Two cases are eligible:
|
|
282
|
-
*
|
|
283
|
-
* 1. Memory refs that pass {@link shouldDistillMemoryRef} (the existing
|
|
284
|
-
* memory→lesson/knowledge promotion path).
|
|
285
|
-
*
|
|
286
|
-
* Refs whose `type` is in {@link DISTILL_REFUSED_INPUT_TYPES} (currently
|
|
287
|
-
* `lesson:*`) are explicitly excluded — distill refuses them at runtime and
|
|
288
|
-
* queuing them just produces a no-op `skipped` outcome per ref per hour. That
|
|
289
|
-
* planner waste was the bug fixed in commit
|
|
290
|
-
* fix(improve): drop distill-refused types from planner.
|
|
291
|
-
*
|
|
292
|
-
* Note: prior to this fix the gate used `isLessonCandidate(ref)` directly,
|
|
293
|
-
* which was true *only* for `lesson:*` refs — exactly the set distill refuses.
|
|
294
|
-
* The result: every hourly run re-queued the same lesson refs, the same skip
|
|
295
|
-
* message returned, and no work was ever done. See
|
|
296
|
-
* `tests/commands/improve-distill-planner-skip-lessons.test.ts`.
|
|
297
|
-
*/
|
|
298
|
-
function isDistillCandidateRef(ref, stashDir) {
|
|
299
|
-
const parsed = parseAssetRef(ref);
|
|
300
|
-
if (isDistillRefusedInputType(parsed.type))
|
|
301
|
-
return false;
|
|
302
|
-
return shouldDistillMemoryRef(ref, stashDir);
|
|
303
|
-
}
|
|
304
|
-
function shouldDistillMemoryRef(ref, stashDir) {
|
|
305
|
-
const parsed = parseAssetRef(ref);
|
|
306
|
-
if (parsed.type !== "memory")
|
|
307
|
-
return false;
|
|
308
|
-
const sources = resolveSourceEntries(stashDir);
|
|
309
|
-
for (const source of sources) {
|
|
310
|
-
const candidate = `${source.path}/memories/${parsed.name}.md`;
|
|
311
|
-
if (!fs.existsSync(candidate))
|
|
312
|
-
continue;
|
|
313
|
-
const raw = fs.readFileSync(candidate, "utf8");
|
|
314
|
-
const fm = parseFrontmatter(raw).data;
|
|
315
|
-
const quality = typeof fm.quality === "string" ? fm.quality : undefined;
|
|
316
|
-
if (quality === "proposed")
|
|
317
|
-
return false;
|
|
318
|
-
return !parsed.name.endsWith(".derived");
|
|
319
|
-
}
|
|
320
|
-
return !parsed.name.endsWith(".derived");
|
|
321
|
-
}
|
|
322
|
-
// ── Signal-delta eligibility helpers (0.8.0) ────────────────────────────────
|
|
323
|
-
//
|
|
324
|
-
// The 0.8.0 redesign replaced flat time-based cooldowns for reflect/distill
|
|
325
|
-
// with a *signal-delta* gate: a ref is re-eligible iff new feedback has
|
|
326
|
-
// landed since the last proposal was generated for it. These helpers build
|
|
327
|
-
// the two timestamp maps the gate needs in bulk, so the planner avoids
|
|
328
|
-
// N+1 queries across the full postCleanupRefs set.
|
|
329
|
-
/**
|
|
330
|
-
* Latest feedback event timestamp per ref in the active window. Reads all
|
|
331
|
-
* `feedback` events newer than `sinceIso` in one query and indexes by ref,
|
|
332
|
-
* keeping the maximum `ts` per ref.
|
|
333
|
-
*
|
|
334
|
-
* Only events with a meaningful payload count as "signal" — `metadata.signal`
|
|
335
|
-
* (positive/negative) OR `metadata.note` (a free-form annotation). Empty
|
|
336
|
-
* metadata events are ignored so a stray `akm feedback <ref>` invocation
|
|
337
|
-
* without a flag doesn't trigger downstream re-processing.
|
|
338
|
-
*/
|
|
339
|
-
function buildLatestFeedbackTsMap(refs, sinceIso) {
|
|
340
|
-
const out = new Map();
|
|
341
|
-
if (refs.length === 0)
|
|
342
|
-
return out;
|
|
343
|
-
const refSet = new Set(refs);
|
|
344
|
-
const { events } = readEvents({ type: "feedback", since: sinceIso });
|
|
345
|
-
for (const e of events) {
|
|
346
|
-
const ref = e.ref;
|
|
347
|
-
if (!ref || !refSet.has(ref))
|
|
348
|
-
continue;
|
|
349
|
-
const meta = e.metadata;
|
|
350
|
-
const hasSignal = meta !== undefined && (typeof meta.signal === "string" || typeof meta.note === "string");
|
|
351
|
-
if (!hasSignal)
|
|
352
|
-
continue;
|
|
353
|
-
const ts = e.ts ?? "";
|
|
354
|
-
if (ts > (out.get(ref) ?? ""))
|
|
355
|
-
out.set(ref, ts);
|
|
356
|
-
}
|
|
357
|
-
return out;
|
|
358
|
-
}
|
|
359
|
-
/**
|
|
360
|
-
* Latest proposal timestamp per input-ref, filtered by source ('reflect' or
|
|
361
|
-
* 'distill'). Reads the corresponding `*_invoked` events from state.db —
|
|
362
|
-
* these events are emitted at proposal creation time and carry the *input*
|
|
363
|
-
* asset ref (memory:foo, skill:bar, etc.) directly. We use them rather than
|
|
364
|
-
* `listProposals` because distill proposals are keyed by the derived
|
|
365
|
-
* lesson/knowledge ref, not the source memory — joining back through the
|
|
366
|
-
* payload would be fragile.
|
|
367
|
-
*/
|
|
368
|
-
function buildLatestProposalTsMap(refs, source) {
|
|
369
|
-
const out = new Map();
|
|
370
|
-
if (refs.length === 0)
|
|
371
|
-
return out;
|
|
372
|
-
const refSet = new Set(refs);
|
|
373
|
-
const eventType = source === "reflect" ? "reflect_invoked" : "distill_invoked";
|
|
374
|
-
const { events } = readEvents({ type: eventType });
|
|
375
|
-
for (const e of events) {
|
|
376
|
-
const ref = e.ref;
|
|
377
|
-
if (!ref || !refSet.has(ref))
|
|
378
|
-
continue;
|
|
379
|
-
// For distill_invoked we only count attempts that produced (or attempted
|
|
380
|
-
// to produce) a real proposal — config_disabled / parse-error outcomes
|
|
381
|
-
// should not move the signal-delta cursor forward.
|
|
382
|
-
if (eventType === "distill_invoked") {
|
|
383
|
-
const outcome = e.metadata?.outcome;
|
|
384
|
-
if (outcome !== "queued" && outcome !== "skipped" && outcome !== "validation_failed")
|
|
385
|
-
continue;
|
|
386
|
-
}
|
|
387
|
-
const ts = e.ts ?? "";
|
|
388
|
-
if (ts > (out.get(ref) ?? ""))
|
|
389
|
-
out.set(ref, ts);
|
|
390
|
-
}
|
|
391
|
-
return out;
|
|
392
|
-
}
|
|
393
|
-
/**
|
|
394
|
-
* Signal-delta eligibility predicate.
|
|
395
|
-
*
|
|
396
|
-
* True iff `latestFeedback[ref]` is defined AND either no prior proposal
|
|
397
|
-
* exists for this (ref, source) OR `latestFeedback[ref] > lastProposal[ref]`.
|
|
398
|
-
*
|
|
399
|
-
* Refs with no feedback signal at all are ineligible by definition — the
|
|
400
|
-
* high-retrieval fallback path (see `noFeedbackCandidates` later in the
|
|
401
|
-
* planner) handles never-touched-but-frequently-read assets separately.
|
|
402
|
-
*/
|
|
403
|
-
function isSignalDeltaEligible(ref, latestFeedback, lastProposal) {
|
|
404
|
-
const fb = latestFeedback.get(ref);
|
|
405
|
-
if (!fb)
|
|
406
|
-
return false;
|
|
407
|
-
const lp = lastProposal.get(ref);
|
|
408
|
-
if (!lp)
|
|
409
|
-
return true;
|
|
410
|
-
return fb > lp;
|
|
411
|
-
}
|
|
412
|
-
/**
|
|
413
|
-
* H7 (#566): cooperative budget watchdog with a captured, RAII-cleared hard-kill.
|
|
414
|
-
*
|
|
415
|
-
* When the wall-clock budget expires, `onExhausted` (normally an
|
|
416
|
-
* `AbortController.abort`) signals cooperative cancellation so the run can drain
|
|
417
|
-
* its in-flight log/`state.db` flush and unwind naturally. A second hard-kill
|
|
418
|
-
* timer is then armed as a watchdog: it only `exit(0)`s if the drain itself
|
|
419
|
-
* overruns `hardKillGraceMs`, preventing the process from outliving the task
|
|
420
|
-
* timeout window (lock-cascade fix).
|
|
421
|
-
*
|
|
422
|
-
* Both timers are captured; the returned dispose() clears whichever is still
|
|
423
|
-
* pending. Callers invoke it from a `finally`, so a *clean* drain reaches the
|
|
424
|
-
* `finally` and cancels the pending hard-kill before it can fire — the previous
|
|
425
|
-
* detached `setTimeout(() => process.exit(0), 5000)` always fired, truncating a
|
|
426
|
-
* clean flush. The hard-kill timer is `unref()`-ed so it never keeps the event
|
|
427
|
-
* loop alive on its own: once the run drains it exits with its own code, not the
|
|
428
|
-
* forced 0.
|
|
429
|
-
*
|
|
430
|
-
* Dependencies are injectable purely so the concurrency-sensitive timing
|
|
431
|
-
* contract can be exercised deterministically in unit tests.
|
|
432
|
-
*/
|
|
433
62
|
export function armBudgetWatchdog(budgetMs, controller, deps) {
|
|
434
63
|
const setTimeoutFn = deps?.setTimeoutFn ?? setTimeout;
|
|
435
64
|
const clearTimeoutFn = deps?.clearTimeoutFn ?? clearTimeout;
|
|
@@ -461,6 +90,11 @@ export async function akmImprove(options = {}) {
|
|
|
461
90
|
const ensureIndexFn = options.ensureIndexFn ?? ensureIndex;
|
|
462
91
|
const reindexFn = options.reindexFn ?? akmIndex;
|
|
463
92
|
const drainProposalsFn = options.drainProposalsFn ?? drainProposals;
|
|
93
|
+
// #616 multi-cycle test seams. Default to the real module-local fns.
|
|
94
|
+
const collectEligibleRefsImpl = options.collectEligibleRefsFn ?? collectEligibleRefs;
|
|
95
|
+
const runImprovePreparationStageImpl = options.runImprovePreparationStageFn ?? runImprovePreparationStage;
|
|
96
|
+
const runImproveLoopStageImpl = options.runImproveLoopStageFn ?? runImproveLoopStage;
|
|
97
|
+
const runImprovePostLoopStageImpl = options.runImprovePostLoopStageFn ?? runImprovePostLoopStage;
|
|
464
98
|
// Resolve the improve profile for this run. Profile drives type filtering,
|
|
465
99
|
// process gating, and default autoAccept/limit values.
|
|
466
100
|
const _earlyConfig = options.config ?? loadConfig();
|
|
@@ -475,6 +109,9 @@ export async function akmImprove(options = {}) {
|
|
|
475
109
|
// CLI --limit takes precedence over both.
|
|
476
110
|
limit: options.limit ?? improveProfile?.processes?.reflect?.limit ?? improveProfile.limit,
|
|
477
111
|
};
|
|
112
|
+
// #616 — bounded multi-cycle phasing. CLI/programmatic override wins over
|
|
113
|
+
// profile.maxCycles; default 1 => single pass (byte-identical to pre-#616).
|
|
114
|
+
const maxCycles = Math.max(1, Math.trunc(options.maxCycles ?? improveProfile.maxCycles ?? 1));
|
|
478
115
|
let primaryStashDir;
|
|
479
116
|
try {
|
|
480
117
|
primaryStashDir = resolveSourceEntries(options.stashDir)[0]?.path;
|
|
@@ -491,174 +128,51 @@ export async function akmImprove(options = {}) {
|
|
|
491
128
|
// timeout root cause). Because beforeEach runs synchronously, env is still the
|
|
492
129
|
// calling test's own at this point; we capture it before yielding the loop.
|
|
493
130
|
const resolvedStateDbPath = getStateDbPathInDataDir();
|
|
494
|
-
//
|
|
495
|
-
//
|
|
496
|
-
//
|
|
497
|
-
//
|
|
498
|
-
//
|
|
499
|
-
//
|
|
500
|
-
|
|
501
|
-
|
|
502
|
-
|
|
503
|
-
|
|
504
|
-
|
|
505
|
-
|
|
506
|
-
|
|
507
|
-
|
|
508
|
-
|
|
509
|
-
|
|
510
|
-
|
|
511
|
-
|
|
512
|
-
|
|
513
|
-
|
|
514
|
-
|
|
515
|
-
|
|
516
|
-
|
|
517
|
-
|
|
518
|
-
|
|
519
|
-
|
|
520
|
-
|
|
521
|
-
})()
|
|
522
|
-
: null;
|
|
523
|
-
if (probe.state === "stale") {
|
|
524
|
-
// O-7 / #394: Emit improve_lock_recovered event before recovery so the
|
|
525
|
-
// audit trail records the abnormal prior-run exit (Temporal/Airflow pattern).
|
|
526
|
-
try {
|
|
527
|
-
appendEvent({
|
|
528
|
-
eventType: "improve_lock_recovered",
|
|
529
|
-
metadata: {
|
|
530
|
-
stalePid: lock?.pid ?? null,
|
|
531
|
-
lockedAt: lock?.startedAt ?? null,
|
|
532
|
-
recoveredAt: new Date().toISOString(),
|
|
533
|
-
lockAgeMs: probe.ageMs ?? null,
|
|
534
|
-
reason: probe.reason === "pid_dead" ? "pid_not_alive" : probe.reason,
|
|
535
|
-
},
|
|
536
|
-
});
|
|
537
|
-
}
|
|
538
|
-
catch {
|
|
539
|
-
/* event emission is best-effort; never block lock recovery */
|
|
540
|
-
}
|
|
541
|
-
releaseLock(resolvedLockPath);
|
|
542
|
-
if (tryAcquireLockSync(resolvedLockPath, lockPayload()))
|
|
543
|
-
return "acquired";
|
|
544
|
-
// Lost the race to another run that grabbed the freed stale lock.
|
|
545
|
-
if (options.skipIfLocked) {
|
|
546
|
-
warn("[improve] another run acquired the lock during stale recovery; skipping (--skip-if-locked)");
|
|
547
|
-
return "skipped";
|
|
548
|
-
}
|
|
549
|
-
throw new ConfigError(`akm improve is already running. Delete ${resolvedLockPath} to force.`, "INVALID_CONFIG_FILE");
|
|
550
|
-
}
|
|
551
|
-
// Lock is held by a live run within the staleness window.
|
|
552
|
-
if (options.skipIfLocked) {
|
|
553
|
-
warn(`[improve] another improve run holds the lock (PID ${lock?.pid}, started ${lock?.startedAt}); skipping (--skip-if-locked)`);
|
|
554
|
-
return "skipped";
|
|
555
|
-
}
|
|
556
|
-
throw new ConfigError(`akm improve is already running (PID ${lock?.pid}, started ${lock?.startedAt}). Delete ${resolvedLockPath} to force.`, "INVALID_CONFIG_FILE");
|
|
557
|
-
};
|
|
558
|
-
// Phase 4 lock-leak guard (§7 ordering hazard): hoisting `improve.lock` above
|
|
559
|
-
// the pre-index region (so the triage pre-pass runs under it) means the lock is
|
|
560
|
-
// held while ensureIndex / collectEligibleRefs / contradiction-detection /
|
|
561
|
-
// memory-cleanup analysis run — but the main protecting `try { … } finally {
|
|
562
|
-
// unlinkSync(resolvedLockPath) }` does not begin until after them. A throw in
|
|
563
|
-
// any of those steps would leak the lock. We close that window by wrapping the
|
|
564
|
-
// whole region in a try whose catch releases the lock (when held) and
|
|
565
|
-
// re-throws. The values this region computes are declared in the outer scope so
|
|
566
|
-
// they remain visible to the main run below. The dry-run path never sets
|
|
567
|
-
// `lockAcquired`, so its early return releases nothing.
|
|
568
|
-
let lockAcquired = false;
|
|
569
|
-
const releaseLockOnError = () => {
|
|
570
|
-
if (!lockAcquired)
|
|
571
|
-
return;
|
|
572
|
-
try {
|
|
573
|
-
fs.unlinkSync(resolvedLockPath);
|
|
574
|
-
}
|
|
575
|
-
catch {
|
|
576
|
-
// best-effort release on the error path
|
|
577
|
-
}
|
|
578
|
-
lockAcquired = false;
|
|
579
|
-
};
|
|
580
|
-
// Signal-safe lock release. The SIGTERM/SIGINT/SIGHUP handler in improve-cli.ts
|
|
581
|
-
// calls `process.exit()`, which does NOT run the `finally` below that owns lock
|
|
582
|
-
// release — so a cron-timeout SIGTERM leaked `improve.lock` every run.
|
|
583
|
-
// `process.exit()` DOES fire `'exit'` listeners, so we release the lock from
|
|
584
|
-
// one. `releaseLockIfOwned` only unlinks a lock still owned by this PID, so it
|
|
585
|
-
// is safe even if a later run re-acquired it. The listener is removed in the
|
|
586
|
-
// `finally` so the normal path stays single-release and repeated in-process
|
|
587
|
-
// `akmImprove` calls (tests) do not accumulate listeners.
|
|
588
|
-
const releaseLockOnExit = () => {
|
|
589
|
-
releaseLockIfOwned(resolvedLockPath, process.pid);
|
|
590
|
-
};
|
|
131
|
+
// #612 / WS-4 — bounded, OPT-IN per-phase auto-accept threshold auto-tune.
|
|
132
|
+
// DEFAULT OFF: `autoTune: false` (or absent) is a complete no-op.
|
|
133
|
+
//
|
|
134
|
+
// WS-4 change: thresholds are now PER PHASE. The old single global mutation
|
|
135
|
+
// of `options.autoAccept` is retired — it caused every phase to share one
|
|
136
|
+
// calibration signal, so a reflect-dominated run could tighten the consolidate
|
|
137
|
+
// gate (or vice-versa). Instead:
|
|
138
|
+
// - Each `makeGateConfig` call reads the phase's stored threshold from
|
|
139
|
+
// state.db (Migration 012) and uses it as `phaseThreshold`, overriding
|
|
140
|
+
// the `globalThreshold` (= options.autoAccept) for that phase.
|
|
141
|
+
// - Per-phase `maybeAutoTuneThreshold` calls fire AFTER each phase's gate
|
|
142
|
+
// has run and persist the new threshold to state.db for the NEXT run.
|
|
143
|
+
// - `options.autoAccept` stays unchanged (it is the operator-supplied
|
|
144
|
+
// baseline, not a mutable run-time state).
|
|
145
|
+
//
|
|
146
|
+
// The global tune call is intentionally removed here. See per-phase calls
|
|
147
|
+
// below (near each makeGateConfig / runAutoAcceptGate block).
|
|
148
|
+
// #607 Lock decomposition: three per-process locks replace the single
|
|
149
|
+
// `improve.lock`. Each process acquires only the lock(s) it needs, so
|
|
150
|
+
// quick-shredder consolidate can run alongside daily reflect+distill.
|
|
151
|
+
//
|
|
152
|
+
// consolidate.lock — protects consolidate + memoryInference + graphExtraction (index.db writers)
|
|
153
|
+
// reflect-distill.lock — protects reflect + distill (state.db proposal writers)
|
|
154
|
+
// triage.lock — protects triage pre-pass (state.db proposal promotions)
|
|
155
|
+
//
|
|
156
|
+
// Lock base directory — same `.akm/` under the primary stash dir.
|
|
157
|
+
const lockBaseDir = primaryStashDir ? path.join(primaryStashDir, ".akm") : path.join(options.stashDir ?? ".", ".akm");
|
|
591
158
|
const preEnsureCleanupWarnings = [];
|
|
592
|
-
|
|
593
|
-
|
|
594
|
-
|
|
159
|
+
// #616: assigned by runIndexAndCollect() (closure) so TS cannot prove definite
|
|
160
|
+
// assignment — seed with empty values; the first runIndexAndCollect() call
|
|
161
|
+
// (cycle 1, in the first try) always overwrites them before any read.
|
|
162
|
+
let plannedRefs = [];
|
|
163
|
+
let memorySummary = { eligible: 0, derived: 0 };
|
|
164
|
+
let profileFilteredRefs = [];
|
|
595
165
|
let memoryCleanupPlan;
|
|
596
166
|
let guidance;
|
|
597
167
|
let triageDrain;
|
|
598
|
-
|
|
599
|
-
|
|
600
|
-
|
|
601
|
-
|
|
602
|
-
|
|
603
|
-
|
|
604
|
-
|
|
605
|
-
|
|
606
|
-
// work — never registered the exit listener, never set lockAcquired,
|
|
607
|
-
// so we release nothing belonging to the run that owns the lock.
|
|
608
|
-
return {
|
|
609
|
-
schemaVersion: 1,
|
|
610
|
-
ok: true,
|
|
611
|
-
scope,
|
|
612
|
-
dryRun: false,
|
|
613
|
-
skipped: { reason: "lock-held" },
|
|
614
|
-
memorySummary: { eligible: 0, derived: 0 },
|
|
615
|
-
plannedRefs: [],
|
|
616
|
-
};
|
|
617
|
-
}
|
|
618
|
-
lockAcquired = true;
|
|
619
|
-
// Backstop release on process.exit() (signal handler / budget watchdog),
|
|
620
|
-
// which skips the finally below. Removed in that finally on the normal path.
|
|
621
|
-
process.on("exit", releaseLockOnExit);
|
|
622
|
-
// Phase 4 triage pre-pass (§7, §13): drain the standing pending backlog
|
|
623
|
-
// BEFORE ensureIndex so improve generates fresh proposals against a cleared
|
|
624
|
-
// queue (no `duplicate_pending` collisions) and ensureIndex absorbs triage's
|
|
625
|
-
// promotions for free. Gated on the triage process being enabled (opt-in,
|
|
626
|
-
// defaults off) and on a whole-stash / type-scoped run — a single-ref
|
|
627
|
-
// `akm improve skill:x` must never drain the whole queue. Best-effort: a
|
|
628
|
-
// triage failure is a non-fatal warning, never an abort (mirrors the
|
|
629
|
-
// contradiction-detection pass below).
|
|
630
|
-
if (primaryStashDir && resolveProcessEnabled("triage", improveProfile)) {
|
|
631
|
-
if (scope.mode === "ref") {
|
|
632
|
-
warn("[improve] triage pre-pass skipped (single-ref scope never drains the whole queue)");
|
|
633
|
-
}
|
|
634
|
-
else {
|
|
635
|
-
try {
|
|
636
|
-
const triageConfig = improveProfile.processes?.triage;
|
|
637
|
-
const policy = resolveDrainPolicy(triageConfig?.policy);
|
|
638
|
-
const applyMode = triageConfig?.applyMode ?? "queue";
|
|
639
|
-
const maxAccepts = triageConfig?.maxAcceptsPerRun ?? 25;
|
|
640
|
-
const judgment = triageConfig?.judgment
|
|
641
|
-
? resolveTriageJudgmentRunner(triageConfig.judgment, _earlyConfig)
|
|
642
|
-
: null;
|
|
643
|
-
triageDrain = await drainProposalsFn({
|
|
644
|
-
stashDir: primaryStashDir,
|
|
645
|
-
policy,
|
|
646
|
-
applyMode,
|
|
647
|
-
maxAccepts,
|
|
648
|
-
dryRun: false,
|
|
649
|
-
// No fresh ids exist yet — triage runs before improve generates any.
|
|
650
|
-
excludeIds: new Set(),
|
|
651
|
-
...(triageConfig?.maxDiffLines !== undefined ? { maxDiffLines: triageConfig.maxDiffLines } : {}),
|
|
652
|
-
judgment,
|
|
653
|
-
});
|
|
654
|
-
}
|
|
655
|
-
catch (err) {
|
|
656
|
-
// Non-fatal: triage is a best-effort pre-pass and must never abort improve.
|
|
657
|
-
warn(`[improve] triage pre-pass failed (non-fatal): ${err instanceof Error ? err.message : String(err)}`);
|
|
658
|
-
}
|
|
659
|
-
}
|
|
660
|
-
}
|
|
661
|
-
}
|
|
168
|
+
// #616 — ensureIndex + collectEligibleRefs + memory-cleanup recompute, lifted
|
|
169
|
+
// into a helper so the SAME sequence runs once for cycle 1 (below, in the
|
|
170
|
+
// first try) and is re-run at the top of each subsequent multi-cycle cycle.
|
|
171
|
+
// Re-running ensureIndex between cycles makes cycle N's gate-promoted
|
|
172
|
+
// proposals visible to cycle N+1's collectEligibleRefs. Mutates the
|
|
173
|
+
// outer-scope plannedRefs/memorySummary/profileFilteredRefs/memoryCleanupPlan/
|
|
174
|
+
// guidance so for maxCycles:1 the body is byte-identical to pre-#616.
|
|
175
|
+
const runIndexAndCollect = async () => {
|
|
662
176
|
// #339 fix: ensureIndex MUST run BEFORE collectEligibleRefs. The eligible-ref
|
|
663
177
|
// query reads the `entries` table; if a DB version upgrade just dropped that
|
|
664
178
|
// table (or the index is otherwise empty), the prior run order silently
|
|
@@ -686,10 +200,10 @@ export async function akmImprove(options = {}) {
|
|
|
686
200
|
// best-effort; leave preEnsureEntryCount undefined
|
|
687
201
|
}
|
|
688
202
|
try {
|
|
689
|
-
await ensureIndexFn(primaryStashDir);
|
|
203
|
+
await ensureIndexFn(primaryStashDir, { mode: "blocking" });
|
|
690
204
|
}
|
|
691
205
|
catch (err) {
|
|
692
|
-
preEnsureCleanupWarnings.push(`ensureIndex failed: ${
|
|
206
|
+
preEnsureCleanupWarnings.push(`ensureIndex failed: ${errMessage(err)}`);
|
|
693
207
|
}
|
|
694
208
|
// #339 loud-fail: if the index was empty pre-ensureIndex but is now
|
|
695
209
|
// populated, a version-upgrade-triggered rebuild just happened. Surface
|
|
@@ -716,7 +230,7 @@ export async function akmImprove(options = {}) {
|
|
|
716
230
|
}
|
|
717
231
|
}
|
|
718
232
|
}
|
|
719
|
-
({ plannedRefs, memorySummary, profileFilteredRefs } = await
|
|
233
|
+
({ plannedRefs, memorySummary, profileFilteredRefs } = await collectEligibleRefsImpl(scope, options.stashDir, improveProfile));
|
|
720
234
|
const cleanupParentRef = memoryCleanupParentRef(scope, options.stashDir);
|
|
721
235
|
// M-1 (#367): Run contradiction-detection BEFORE analyzeMemoryCleanup so
|
|
722
236
|
// the SCC resolver in resolveFamilyContradictions has edges to work on.
|
|
@@ -728,7 +242,7 @@ export async function akmImprove(options = {}) {
|
|
|
728
242
|
}
|
|
729
243
|
catch (err) {
|
|
730
244
|
// Non-fatal: contradiction detection is a best-effort pass.
|
|
731
|
-
warn(`[improve] contradiction detection failed (non-fatal): ${
|
|
245
|
+
warn(`[improve] contradiction detection failed (non-fatal): ${errMessage(err)}`);
|
|
732
246
|
}
|
|
733
247
|
}
|
|
734
248
|
memoryCleanupPlan = shouldAnalyzeMemoryCleanup(scope, memorySummary.eligible, primaryStashDir)
|
|
@@ -738,6 +252,70 @@ export async function akmImprove(options = {}) {
|
|
|
738
252
|
memorySummary.eligible > 0
|
|
739
253
|
? "Improve folds memory cleanup into the same proposal queue: speculative promotions still go through reflect/distill proposals, while high-confidence redundant derived memories are moved into a recoverable cleanup archive instead of being left active in the stash."
|
|
740
254
|
: undefined;
|
|
255
|
+
};
|
|
256
|
+
// Holds our own process.on("exit") backstop so the finally can remove EXACTLY
|
|
257
|
+
// that handler (not every exit listener in the process). Declared in the scope
|
|
258
|
+
// shared by the try and its finally; assigned when the backstop is registered.
|
|
259
|
+
let exitBackstop;
|
|
260
|
+
try {
|
|
261
|
+
// #607: Per-process lock acquisition. Each process acquires only the lock(s)
|
|
262
|
+
// it needs. The dry-run branch produces plannedRefs/memorySummary WITHOUT any
|
|
263
|
+
// locks (decision: dry-run never mutates the queue).
|
|
264
|
+
if (!options.dryRun) {
|
|
265
|
+
// Backstop release on process.exit() (signal handler / budget watchdog),
|
|
266
|
+
// which skips the finally below. Removed in that finally on the normal path.
|
|
267
|
+
const releaseAllOnExit = () => releaseHeldLocksIfOwned(process.pid);
|
|
268
|
+
exitBackstop = releaseAllOnExit;
|
|
269
|
+
process.on("exit", releaseAllOnExit);
|
|
270
|
+
// #607 triage pre-pass: acquire triage.lock, drain the standing pending
|
|
271
|
+
// backlog BEFORE ensureIndex so improve generates fresh proposals against
|
|
272
|
+
// a cleared queue (no `duplicate_pending` collisions) and ensureIndex
|
|
273
|
+
// absorbs triage's promotions for free. Release immediately after —
|
|
274
|
+
// triage.lock is not needed again until the next improve run.
|
|
275
|
+
if (primaryStashDir && resolveProcessEnabled("triage", improveProfile)) {
|
|
276
|
+
if (scope.mode === "ref") {
|
|
277
|
+
warn("[improve] triage pre-pass skipped (single-ref scope never drains the whole queue)");
|
|
278
|
+
}
|
|
279
|
+
else {
|
|
280
|
+
const triageLPath = processLockPath(lockBaseDir, "triage");
|
|
281
|
+
const triageResult = tryAcquireProcessLock(triageLPath, PROCESS_LOCK_DEFS.triage.staleAfterMs, options.skipIfLocked, "triage");
|
|
282
|
+
if (triageResult === "skipped") {
|
|
283
|
+
triageDrain = undefined;
|
|
284
|
+
}
|
|
285
|
+
else {
|
|
286
|
+
try {
|
|
287
|
+
const triageConfig = improveProfile.processes?.triage;
|
|
288
|
+
const policy = resolveDrainPolicy(triageConfig?.policy);
|
|
289
|
+
const applyMode = triageConfig?.applyMode ?? "queue";
|
|
290
|
+
const maxAccepts = triageConfig?.maxAcceptsPerRun ?? 25;
|
|
291
|
+
const judgment = triageConfig?.judgment
|
|
292
|
+
? resolveTriageJudgmentRunner(triageConfig.judgment, _earlyConfig)
|
|
293
|
+
: null;
|
|
294
|
+
triageDrain = await drainProposalsFn({
|
|
295
|
+
stashDir: primaryStashDir,
|
|
296
|
+
policy,
|
|
297
|
+
applyMode,
|
|
298
|
+
maxAccepts,
|
|
299
|
+
dryRun: false,
|
|
300
|
+
excludeIds: new Set(),
|
|
301
|
+
...(triageConfig?.maxDiffLines !== undefined ? { maxDiffLines: triageConfig.maxDiffLines } : {}),
|
|
302
|
+
judgment,
|
|
303
|
+
});
|
|
304
|
+
}
|
|
305
|
+
catch (err) {
|
|
306
|
+
warn(`[improve] triage pre-pass failed (non-fatal): ${errMessage(err)}`);
|
|
307
|
+
}
|
|
308
|
+
finally {
|
|
309
|
+
releaseProcessLock(triageLPath);
|
|
310
|
+
}
|
|
311
|
+
}
|
|
312
|
+
}
|
|
313
|
+
}
|
|
314
|
+
}
|
|
315
|
+
// #339 fix: ensureIndex MUST run BEFORE collectEligibleRefs (now inside the
|
|
316
|
+
// helper). Cycle 1 runs it here; subsequent multi-cycle cycles re-run it via
|
|
317
|
+
// the same helper at the top of each cycle below.
|
|
318
|
+
await runIndexAndCollect();
|
|
741
319
|
if (options.dryRun) {
|
|
742
320
|
const result = {
|
|
743
321
|
schemaVersion: 1,
|
|
@@ -754,17 +332,14 @@ export async function akmImprove(options = {}) {
|
|
|
754
332
|
}
|
|
755
333
|
}
|
|
756
334
|
catch (err) {
|
|
757
|
-
|
|
335
|
+
releaseAllProcessLocks();
|
|
758
336
|
throw err;
|
|
759
337
|
}
|
|
760
|
-
//
|
|
761
|
-
//
|
|
762
|
-
//
|
|
763
|
-
//
|
|
764
|
-
//
|
|
765
|
-
// any of them used to leak the lock (blocking the next improve up to 4h);
|
|
766
|
-
// now the finally releases it exactly once. The dry-run path already returned
|
|
767
|
-
// above without acquiring the lock, so it never reaches this finally; the
|
|
338
|
+
// #607: per-process locks are acquired/released around each stage below.
|
|
339
|
+
// The triage pre-pass already ran under triage.lock (released). The
|
|
340
|
+
// preparation stage runs under consolidate.lock, the loop stage under
|
|
341
|
+
// reflect-distill.lock, and the post-loop stage under consolidate.lock again.
|
|
342
|
+
// Each stage acquires its lock just before starting and releases in finally.
|
|
768
343
|
// best-effort `unlinkSync` is a no-op when no lock file exists.
|
|
769
344
|
const startMs = Date.now();
|
|
770
345
|
const budgetMs = options.timeoutMs ?? 2 * 60 * 60 * 1000; // default 2 hours
|
|
@@ -773,6 +348,16 @@ export async function akmImprove(options = {}) {
|
|
|
773
348
|
// run past the declared budget.
|
|
774
349
|
// References: Anthropic *Building Effective Agents* (2024); CoALA §5 (arXiv:2309.02427).
|
|
775
350
|
const budgetAbortController = new AbortController();
|
|
351
|
+
// Attach a live `remainingBudgetMs` getter to the signal so sub-callers
|
|
352
|
+
// (e.g. consolidate.ts cold-start budget estimation) can read the remaining
|
|
353
|
+
// wall-clock budget without needing an extra plumbing parameter. The property
|
|
354
|
+
// is computed at access time via a getter so it always reflects the actual
|
|
355
|
+
// elapsed time rather than a stale snapshot taken at arm time.
|
|
356
|
+
Object.defineProperty(budgetAbortController.signal, "remainingBudgetMs", {
|
|
357
|
+
get: () => Math.max(0, budgetMs - (Date.now() - startMs)),
|
|
358
|
+
enumerable: false,
|
|
359
|
+
configurable: true,
|
|
360
|
+
});
|
|
776
361
|
// Declared in the outer scope so the `finally` can clear the timer even if a
|
|
777
362
|
// throw occurs before/after it is armed. Defaults to a no-op until armed.
|
|
778
363
|
let clearBudgetTimer = () => { };
|
|
@@ -787,6 +372,63 @@ export async function akmImprove(options = {}) {
|
|
|
787
372
|
// #576: clears the per-run LLM usage sink. Defaults to a no-op until the sink
|
|
788
373
|
// is installed inside the try; the `finally` always calls it.
|
|
789
374
|
let disposeLlmUsageSink = () => { };
|
|
375
|
+
// ── Crash-safe / incremental stash sync (#662) ──────────────────────────────
|
|
376
|
+
// The primary stash writes as a filesystem source DURING the run
|
|
377
|
+
// (write-source.ts case-3); those writes become a git commit only when this
|
|
378
|
+
// closure runs. Historically the only call site was a single BATCH commit at
|
|
379
|
+
// the very end of the happy path, so a run interrupted AFTER writing but
|
|
380
|
+
// BEFORE finishing — a mid-cycle crash, a budget abort, or an external
|
|
381
|
+
// SIGTERM/`process.exit` — left every write uncommitted until some LATER run
|
|
382
|
+
// happened to finish cleanly and swept the whole backlog up. We now call this
|
|
383
|
+
// from THREE places: between cycles (bank each completed cycle), at end-of-run
|
|
384
|
+
// (the converged commit), and from the catch path (commit what was written
|
|
385
|
+
// before the crash). That shrinks the worst-case loss from "the entire run" to
|
|
386
|
+
// "the in-flight cycle".
|
|
387
|
+
//
|
|
388
|
+
// Declared in the OUTER scope (not inside the try) so the catch block can reach
|
|
389
|
+
// it. Idempotent + NON-FATAL: `saveGitStash` short-circuits a clean working
|
|
390
|
+
// tree ("nothing to commit") and a thrown sync error is swallowed here, so a
|
|
391
|
+
// repeat call after a no-op cycle is cheap and a failed push never fails the
|
|
392
|
+
// run. Gated identically to the original end-of-run block (git-backed primary
|
|
393
|
+
// stash, sync not disabled). `eventsCtx` is captured by reference, so calls
|
|
394
|
+
// after the db-backed context is installed inside the try use the live handle.
|
|
395
|
+
const effectiveSync = { ...improveProfile.sync, ...options.sync };
|
|
396
|
+
const commitStashBatch = (messageContext) => {
|
|
397
|
+
if (!primaryStashDir || effectiveSync.enabled === false || !isGitBackedStash(primaryStashDir)) {
|
|
398
|
+
return undefined;
|
|
399
|
+
}
|
|
400
|
+
const saveGitStashFn = options.saveGitStashFn ?? saveGitStash;
|
|
401
|
+
const writableOverride = resolveWritableOverride(_earlyConfig);
|
|
402
|
+
const push = effectiveSync.push !== false;
|
|
403
|
+
const message = renderSyncCommitMessage(effectiveSync.message ?? "akm improve auto-sync", messageContext, Date.now());
|
|
404
|
+
try {
|
|
405
|
+
const syncResult = saveGitStashFn(undefined, message, writableOverride, { push, repoDir: primaryStashDir });
|
|
406
|
+
appendEvent({
|
|
407
|
+
eventType: "stash_synced",
|
|
408
|
+
metadata: {
|
|
409
|
+
committed: syncResult.committed,
|
|
410
|
+
pushed: syncResult.pushed,
|
|
411
|
+
skipped: syncResult.skipped,
|
|
412
|
+
reason: syncResult.reason ?? null,
|
|
413
|
+
},
|
|
414
|
+
}, eventsCtx);
|
|
415
|
+
return {
|
|
416
|
+
committed: syncResult.committed,
|
|
417
|
+
pushed: syncResult.pushed,
|
|
418
|
+
skipped: syncResult.skipped,
|
|
419
|
+
...(syncResult.reason !== undefined ? { reason: syncResult.reason } : {}),
|
|
420
|
+
};
|
|
421
|
+
}
|
|
422
|
+
catch (syncErr) {
|
|
423
|
+
const reason = errMessage(syncErr);
|
|
424
|
+
warn(`improve: stash sync failed (non-fatal): ${reason}`);
|
|
425
|
+
appendEvent({
|
|
426
|
+
eventType: "stash_synced",
|
|
427
|
+
metadata: { committed: false, pushed: false, skipped: true, reason },
|
|
428
|
+
}, eventsCtx);
|
|
429
|
+
return { committed: false, pushed: false, skipped: true, reason };
|
|
430
|
+
}
|
|
431
|
+
};
|
|
790
432
|
try {
|
|
791
433
|
// H7 (#566): arm the budget watchdog. `armBudgetWatchdog` captures both the
|
|
792
434
|
// budget timer and the hard-kill timer it schedules on exhaustion, returning
|
|
@@ -828,71 +470,244 @@ export async function akmImprove(options = {}) {
|
|
|
828
470
|
},
|
|
829
471
|
}, eventsCtx);
|
|
830
472
|
}
|
|
831
|
-
|
|
832
|
-
|
|
833
|
-
|
|
834
|
-
|
|
835
|
-
|
|
836
|
-
|
|
837
|
-
|
|
838
|
-
|
|
839
|
-
|
|
840
|
-
|
|
841
|
-
|
|
842
|
-
|
|
843
|
-
|
|
844
|
-
|
|
845
|
-
|
|
846
|
-
|
|
847
|
-
|
|
848
|
-
|
|
849
|
-
|
|
850
|
-
|
|
851
|
-
|
|
852
|
-
|
|
853
|
-
|
|
473
|
+
// #616 — bounded multi-cycle phasing. The prep->loop->post-loop sequence is
|
|
474
|
+
// wrapped in an N-cycle loop. Each cycle re-runs ensureIndex +
|
|
475
|
+
// collectEligibleRefs (via runIndexAndCollect) so gate-accepted output of
|
|
476
|
+
// cycle N becomes selectable input to cycle N+1. The per-stage process locks
|
|
477
|
+
// (consolidate / reflect-distill) are acquired+released INSIDE each cycle,
|
|
478
|
+
// exactly as the single-pass path did. For maxCycles:1 the loop runs once and
|
|
479
|
+
// every accumulator below collapses to the single-cycle value (sum-of-one,
|
|
480
|
+
// concat-of-one, last==only) => BYTE-IDENTICAL to pre-#616.
|
|
481
|
+
//
|
|
482
|
+
// Accumulators (see CONSTRAINTS / aggregation plan in #616): SUM the count
|
|
483
|
+
// fields and durations; CONCAT the array fields; LAST-WINS for point-in-time
|
|
484
|
+
// objects (the final cycle's value reflects the converged state).
|
|
485
|
+
let cyclesRun = 0;
|
|
486
|
+
// Last-wins point-in-time values (assigned every cycle; the final cycle wins).
|
|
487
|
+
let preparation;
|
|
488
|
+
let memoryRefsForInference = new Set();
|
|
489
|
+
let consolidation;
|
|
490
|
+
let memoryInference;
|
|
491
|
+
let graphExtraction;
|
|
492
|
+
let recombination;
|
|
493
|
+
let proceduralCompilation;
|
|
494
|
+
let cycleMetrics;
|
|
495
|
+
// Summed counters/durations.
|
|
496
|
+
let prepGateCount = 0;
|
|
497
|
+
let prepGateFailedCount = 0;
|
|
498
|
+
let reflectsWithErrorContext = 0;
|
|
499
|
+
let loopGateCount = 0;
|
|
500
|
+
let loopGateFailedCount = 0;
|
|
501
|
+
let postLoopGateCount = 0;
|
|
502
|
+
let postLoopGateFailedCount = 0;
|
|
503
|
+
let memoryInferenceDurationMs = 0;
|
|
504
|
+
let graphExtractionDurationMs = 0;
|
|
505
|
+
let orphansPurged;
|
|
506
|
+
let proposalsExpired;
|
|
507
|
+
// Concatenated arrays.
|
|
508
|
+
const allWarnings = [];
|
|
509
|
+
let deadUrls;
|
|
510
|
+
const finalActions = [];
|
|
511
|
+
for (let cycleIndex = 0; cycleIndex < maxCycles; cycleIndex++) {
|
|
512
|
+
// #616 budget gate: never start a NEW cycle once the run's wall-clock
|
|
513
|
+
// budget is exhausted (or the run was aborted). Cycle 0 ALWAYS runs so
|
|
514
|
+
// maxCycles:1 is byte-identical regardless of budget.
|
|
515
|
+
if (cycleIndex > 0) {
|
|
516
|
+
const remaining = budgetAbortController.signal.remainingBudgetMs;
|
|
517
|
+
if (budgetAbortController.signal.aborted || (remaining !== undefined && remaining <= 0)) {
|
|
518
|
+
break;
|
|
519
|
+
}
|
|
520
|
+
}
|
|
521
|
+
// #662 incremental sync: bank the PREVIOUS cycle's writes before starting a
|
|
522
|
+
// new one, so a crash/abort/timeout mid-run loses at most the in-flight
|
|
523
|
+
// cycle rather than the whole run. Guarded on `cycleIndex > 0`, so the
|
|
524
|
+
// common maxCycles:1 path never calls this — its single end-of-run commit
|
|
525
|
+
// below stays the only sync and the serialized envelope is byte-identical
|
|
526
|
+
// to pre-#662. `saveGitStash` no-ops a clean tree, so a cycle that wrote
|
|
527
|
+
// nothing costs only a `git status`.
|
|
528
|
+
if (cycleIndex > 0) {
|
|
529
|
+
commitStashBatch({ scope, plannedRefs, runId: options.runId });
|
|
530
|
+
}
|
|
531
|
+
// Re-run ensureIndex + collectEligibleRefs + memory-cleanup recompute for
|
|
532
|
+
// cycles 2+ (cycle 1 already ran them in the first try above). This makes
|
|
533
|
+
// cycle N's gate-promoted proposals visible to this cycle's ref selection.
|
|
534
|
+
if (cycleIndex > 0) {
|
|
535
|
+
await runIndexAndCollect();
|
|
536
|
+
// Re-emit the profile-filtered audit summary for this cycle's selection.
|
|
537
|
+
if (profileFilteredRefs.length > 0) {
|
|
538
|
+
appendEvent({
|
|
539
|
+
eventType: "improve_skipped",
|
|
540
|
+
ref: undefined,
|
|
541
|
+
metadata: { reason: "profile_filtered_all_passes", count: profileFilteredRefs.length },
|
|
542
|
+
}, eventsCtx);
|
|
543
|
+
}
|
|
854
544
|
}
|
|
545
|
+
// #607: acquire consolidate.lock for the preparation stage (consolidate,
|
|
546
|
+
// ensureIndex, extract all write index.db). Released immediately after.
|
|
547
|
+
const consolidateLPath = processLockPath(lockBaseDir, "consolidate");
|
|
548
|
+
preparation = await withOptionalProcessLock({
|
|
549
|
+
lockPath: consolidateLPath,
|
|
550
|
+
staleAfterMs: PROCESS_LOCK_DEFS.consolidate.staleAfterMs,
|
|
551
|
+
skipIfLocked: options.skipIfLocked,
|
|
552
|
+
label: "consolidate",
|
|
553
|
+
}, () => runImprovePreparationStageImpl({
|
|
554
|
+
scope,
|
|
555
|
+
options,
|
|
556
|
+
plannedRefs,
|
|
557
|
+
memoryCleanupPlan,
|
|
558
|
+
primaryStashDir,
|
|
559
|
+
memorySummary,
|
|
560
|
+
reindexFn,
|
|
561
|
+
startMs,
|
|
562
|
+
budgetMs,
|
|
563
|
+
eventsCtx,
|
|
564
|
+
initialCleanupWarnings: preEnsureCleanupWarnings,
|
|
565
|
+
improveProfile,
|
|
566
|
+
budgetSignal: budgetAbortController.signal,
|
|
567
|
+
}));
|
|
568
|
+
prepGateCount += preparation.gateAutoAcceptedCount;
|
|
569
|
+
prepGateFailedCount += preparation.gateAutoAcceptFailedCount;
|
|
570
|
+
// D6: pre-load all proposal_rejected events from the last 30 days once,
|
|
571
|
+
// so the per-asset loop can use a Map lookup instead of N DB round trips.
|
|
572
|
+
const REJECTED_PROPOSAL_WINDOW_MS = daysToMs(30);
|
|
573
|
+
const rejectedProposalSince = new Date(Date.now() - REJECTED_PROPOSAL_WINDOW_MS).toISOString();
|
|
574
|
+
const allRejectedProposalEvents = readEvents({ type: "proposal_rejected", since: rejectedProposalSince }).events;
|
|
575
|
+
const rejectedProposalsByRef = new Map();
|
|
576
|
+
for (const e of allRejectedProposalEvents) {
|
|
577
|
+
if (e.ref && (!rejectedProposalsByRef.has(e.ref) || e.ts > (rejectedProposalsByRef.get(e.ref)?.ts ?? ""))) {
|
|
578
|
+
rejectedProposalsByRef.set(e.ref, e);
|
|
579
|
+
}
|
|
580
|
+
}
|
|
581
|
+
// #607: acquire reflect-distill.lock for the loop stage (reflect + distill
|
|
582
|
+
// both write proposals to state.db). Released immediately after.
|
|
583
|
+
const reflectDistillLPath = processLockPath(lockBaseDir, "reflectDistill");
|
|
584
|
+
const loopResult = await withOptionalProcessLock({
|
|
585
|
+
lockPath: reflectDistillLPath,
|
|
586
|
+
staleAfterMs: PROCESS_LOCK_DEFS.reflectDistill.staleAfterMs,
|
|
587
|
+
skipIfLocked: options.skipIfLocked,
|
|
588
|
+
label: "reflect-distill",
|
|
589
|
+
}, () => {
|
|
590
|
+
// Post-lock cooldown re-filter for proactive refs (#SELECT-TIME-LEAK).
|
|
591
|
+
// Planning built `lastReflectProposalTs` BEFORE acquiring this lock, so a
|
|
592
|
+
// concurrent run's `reflect_invoked` writes are invisible to it. Now that
|
|
593
|
+
// we hold the lock, re-read fresh timestamp maps for the proactive subset
|
|
594
|
+
// and drop any ref whose cooldown has been consumed by the concurrent run.
|
|
595
|
+
const proactiveLoopRefs = preparation.loopRefs.filter((r) => r.eligibilitySource === "proactive");
|
|
596
|
+
let postLockLoopRefs = preparation.loopRefs;
|
|
597
|
+
if (proactiveLoopRefs.length > 0) {
|
|
598
|
+
const proactiveRefStrs = proactiveLoopRefs.map((r) => r.ref);
|
|
599
|
+
const freshReflectTs = buildLatestProposalTsMap(proactiveRefStrs, "reflect");
|
|
600
|
+
const freshDistillTs = buildLatestProposalTsMap(proactiveRefStrs, "distill");
|
|
601
|
+
const pmDueDays = improveProfile.processes?.proactiveMaintenance?.dueDays ?? DEFAULT_DUE_DAYS;
|
|
602
|
+
const stillDue = new Set(filterProactiveDue(proactiveLoopRefs, freshReflectTs, freshDistillTs, pmDueDays, Date.now()).map((r) => r.ref));
|
|
603
|
+
const dropped = proactiveLoopRefs.filter((r) => !stillDue.has(r.ref));
|
|
604
|
+
if (dropped.length > 0) {
|
|
605
|
+
info(`[improve] post-lock cooldown re-filter: dropped ${dropped.length} proactive ref(s) claimed by concurrent run (${dropped.map((r) => r.ref).join(", ")})`);
|
|
606
|
+
postLockLoopRefs = preparation.loopRefs.filter((r) => r.eligibilitySource !== "proactive" || stillDue.has(r.ref));
|
|
607
|
+
}
|
|
608
|
+
}
|
|
609
|
+
return runImproveLoopStageImpl({
|
|
610
|
+
scope,
|
|
611
|
+
options,
|
|
612
|
+
primaryStashDir,
|
|
613
|
+
reflectFn,
|
|
614
|
+
distillFn,
|
|
615
|
+
loopRefs: postLockLoopRefs,
|
|
616
|
+
actions: preparation.actions,
|
|
617
|
+
signalBearingSet: preparation.signalBearingSet,
|
|
618
|
+
distillCooledRefs: preparation.distillCooledRefs,
|
|
619
|
+
distillOnlyRefs: preparation.distillOnlyRefs,
|
|
620
|
+
recentErrors: preparation.recentErrors,
|
|
621
|
+
rejectedProposalsByRef,
|
|
622
|
+
utilityMap: preparation.utilityMap,
|
|
623
|
+
startMs,
|
|
624
|
+
budgetMs,
|
|
625
|
+
eventsCtx,
|
|
626
|
+
improveProfile,
|
|
627
|
+
budgetSignal: budgetAbortController.signal,
|
|
628
|
+
});
|
|
629
|
+
});
|
|
630
|
+
const loopGateCountThisCycle = loopResult.gateAutoAcceptedCount;
|
|
631
|
+
reflectsWithErrorContext += loopResult.reflectsWithErrorContext;
|
|
632
|
+
loopGateCount += loopResult.gateAutoAcceptedCount;
|
|
633
|
+
loopGateFailedCount += loopResult.gateAutoAcceptFailedCount;
|
|
634
|
+
memoryRefsForInference = loopResult.memoryRefsForInference;
|
|
635
|
+
// #551: consolidation now runs in the preparation stage (before extract);
|
|
636
|
+
// its result and run-flag are read from `preparation`, not the post-loop.
|
|
637
|
+
consolidation = preparation.consolidation;
|
|
638
|
+
// #607: acquire consolidate.lock for the post-loop stage (memoryInference +
|
|
639
|
+
// graphExtraction both write index.db). Released immediately after.
|
|
640
|
+
const consolidatePostLPath = processLockPath(lockBaseDir, "consolidate");
|
|
641
|
+
const postLoopResult = await withOptionalProcessLock({
|
|
642
|
+
lockPath: consolidatePostLPath,
|
|
643
|
+
staleAfterMs: PROCESS_LOCK_DEFS.consolidate.staleAfterMs,
|
|
644
|
+
skipIfLocked: options.skipIfLocked,
|
|
645
|
+
label: "consolidate",
|
|
646
|
+
}, () => runImprovePostLoopStageImpl({
|
|
647
|
+
scope,
|
|
648
|
+
options,
|
|
649
|
+
primaryStashDir,
|
|
650
|
+
actionableRefs: preparation.actionableRefs,
|
|
651
|
+
appliedCleanup: preparation.appliedCleanup,
|
|
652
|
+
cleanupWarnings: preparation.cleanupWarnings,
|
|
653
|
+
memoryRefsForInference,
|
|
654
|
+
reindexFn,
|
|
655
|
+
eventsCtx,
|
|
656
|
+
budgetSignal: budgetAbortController.signal,
|
|
657
|
+
improveProfile,
|
|
658
|
+
consolidationRan: preparation.consolidationRan,
|
|
659
|
+
// R5: floor violations from this run's consolidate pass + the
|
|
660
|
+
// auto-accepted volume so far (prep + loop gates) for churn detection.
|
|
661
|
+
consolidationMergeFloorViolations: preparation.consolidation.mergeFloorViolations ?? 0,
|
|
662
|
+
acceptedActions: preparation.gateAutoAcceptedCount + loopGateCountThisCycle,
|
|
663
|
+
}));
|
|
664
|
+
const postLoopGateCountThisCycle = postLoopResult.gateAutoAcceptedCount;
|
|
665
|
+
// Last-wins point-in-time objects.
|
|
666
|
+
memoryInference = postLoopResult.memoryInference;
|
|
667
|
+
graphExtraction = postLoopResult.graphExtraction;
|
|
668
|
+
recombination = postLoopResult.recombination;
|
|
669
|
+
proceduralCompilation = postLoopResult.proceduralCompilation;
|
|
670
|
+
// Keep the last QUALIFYING cycle's snapshot — a later non-qualifying
|
|
671
|
+
// cycle in a maxCycles>1 run must not clobber it with undefined.
|
|
672
|
+
if (postLoopResult.cycleMetrics)
|
|
673
|
+
cycleMetrics = postLoopResult.cycleMetrics;
|
|
674
|
+
// Summed counters/durations.
|
|
675
|
+
postLoopGateCount += postLoopResult.gateAutoAcceptedCount;
|
|
676
|
+
postLoopGateFailedCount += postLoopResult.gateAutoAcceptFailedCount;
|
|
677
|
+
memoryInferenceDurationMs += postLoopResult.memoryInferenceDurationMs;
|
|
678
|
+
graphExtractionDurationMs += postLoopResult.graphExtractionDurationMs;
|
|
679
|
+
if (postLoopResult.orphansPurged !== undefined) {
|
|
680
|
+
orphansPurged = (orphansPurged ?? 0) + postLoopResult.orphansPurged;
|
|
681
|
+
}
|
|
682
|
+
if (postLoopResult.proposalsExpired !== undefined) {
|
|
683
|
+
proposalsExpired = (proposalsExpired ?? 0) + postLoopResult.proposalsExpired;
|
|
684
|
+
}
|
|
685
|
+
// Concatenated arrays.
|
|
686
|
+
allWarnings.push(...postLoopResult.allWarnings);
|
|
687
|
+
if (postLoopResult.deadUrls !== undefined) {
|
|
688
|
+
deadUrls = [...(deadUrls ?? []), ...postLoopResult.deadUrls];
|
|
689
|
+
}
|
|
690
|
+
const maintenanceActions = postLoopResult.maintenanceActions;
|
|
691
|
+
if (maintenanceActions && maintenanceActions.length > 0) {
|
|
692
|
+
finalActions.push(...preparation.actions, ...maintenanceActions);
|
|
693
|
+
}
|
|
694
|
+
else {
|
|
695
|
+
finalActions.push(...preparation.actions);
|
|
696
|
+
}
|
|
697
|
+
cyclesRun++;
|
|
698
|
+
// #616 fixed-point stop: a cycle that produced ZERO gate-accepted proposals
|
|
699
|
+
// (summed across prep + loop + post-loop) would feed cycle N+1 an identical
|
|
700
|
+
// ref set, so end the loop here rather than spin a pointless next cycle.
|
|
701
|
+
const gateAcceptedThisCycle = preparation.gateAutoAcceptedCount + loopGateCountThisCycle + postLoopGateCountThisCycle;
|
|
702
|
+
if (gateAcceptedThisCycle === 0)
|
|
703
|
+
break;
|
|
855
704
|
}
|
|
856
|
-
|
|
857
|
-
|
|
858
|
-
|
|
859
|
-
|
|
860
|
-
|
|
861
|
-
|
|
862
|
-
loopRefs: preparation.loopRefs,
|
|
863
|
-
actions: preparation.actions,
|
|
864
|
-
signalBearingSet: preparation.signalBearingSet,
|
|
865
|
-
distillCooledRefs: preparation.distillCooledRefs,
|
|
866
|
-
distillOnlyRefs: preparation.distillOnlyRefs,
|
|
867
|
-
recentErrors: preparation.recentErrors,
|
|
868
|
-
rejectedProposalsByRef,
|
|
869
|
-
utilityMap: preparation.utilityMap,
|
|
870
|
-
startMs,
|
|
871
|
-
budgetMs,
|
|
872
|
-
eventsCtx,
|
|
873
|
-
improveProfile,
|
|
874
|
-
});
|
|
875
|
-
// #551: consolidation now runs in the preparation stage (before extract);
|
|
876
|
-
// its result and run-flag are read from `preparation`, not the post-loop.
|
|
877
|
-
const consolidation = preparation.consolidation;
|
|
878
|
-
const { allWarnings, deadUrls, memoryInference, graphExtraction, stalenessDetection, maintenanceActions, memoryInferenceDurationMs, graphExtractionDurationMs, orphansPurged, proposalsExpired, gateAutoAcceptedCount: postLoopGateCount, gateAutoAcceptFailedCount: postLoopGateFailedCount, } = await runImprovePostLoopStage({
|
|
879
|
-
scope,
|
|
880
|
-
options,
|
|
881
|
-
primaryStashDir,
|
|
882
|
-
actionableRefs: preparation.actionableRefs,
|
|
883
|
-
appliedCleanup: preparation.appliedCleanup,
|
|
884
|
-
cleanupWarnings: preparation.cleanupWarnings,
|
|
885
|
-
memoryRefsForInference,
|
|
886
|
-
reindexFn,
|
|
887
|
-
eventsCtx,
|
|
888
|
-
// O-1 (#364): propagate wall-clock budget signal to post-loop maintenance.
|
|
889
|
-
budgetSignal: budgetAbortController.signal,
|
|
890
|
-
improveProfile,
|
|
891
|
-
consolidationRan: preparation.consolidationRan,
|
|
892
|
-
});
|
|
893
|
-
const finalActions = maintenanceActions && maintenanceActions.length > 0
|
|
894
|
-
? [...preparation.actions, ...maintenanceActions]
|
|
895
|
-
: preparation.actions;
|
|
705
|
+
// C1 (13-bus-factor): fold the per-ref `distill-skipped` rows (~13k/run,
|
|
706
|
+
// ~91% of result_json bytes) into a bounded aggregate BEFORE persistence.
|
|
707
|
+
// The metric total + per-reason breakdown are preserved on `distillSkipped`;
|
|
708
|
+
// the unbounded row list never reaches result_json. Reflect skip counters
|
|
709
|
+
// below still read `finalActions` (reflect skips are not folded).
|
|
710
|
+
const { actions: persistedActions, aggregate: distillSkippedAggregate } = foldDistillSkipped(finalActions);
|
|
896
711
|
const result = {
|
|
897
712
|
schemaVersion: 1,
|
|
898
713
|
ok: true,
|
|
@@ -923,7 +738,8 @@ export async function akmImprove(options = {}) {
|
|
|
923
738
|
: {}),
|
|
924
739
|
plannedRefs: preparation.actionableRefs,
|
|
925
740
|
...(profileFilteredRefs.length > 0 ? { profileFilteredRefs } : {}),
|
|
926
|
-
actions:
|
|
741
|
+
actions: persistedActions,
|
|
742
|
+
...(distillSkippedAggregate ? { distillSkipped: distillSkippedAggregate } : {}),
|
|
927
743
|
...(preparation.validationFailures.length > 0 ? { validationFailures: preparation.validationFailures } : {}),
|
|
928
744
|
...(preparation.schemaRepairs.length > 0 ? { schemaRepairs: preparation.schemaRepairs } : {}),
|
|
929
745
|
...(consolidation.processed > 0 || consolidation.warnings.length > 0 ? { consolidation } : {}),
|
|
@@ -950,18 +766,20 @@ export async function akmImprove(options = {}) {
|
|
|
950
766
|
// `/tmp/akm-health-investigations/metrics-taxonomy-review.md` §1k / §3.
|
|
951
767
|
...(memoryInferenceDurationMs > 0 ? { memoryInferenceDurationMs } : {}),
|
|
952
768
|
...(graphExtractionDurationMs > 0 ? { graphExtractionDurationMs } : {}),
|
|
953
|
-
...(
|
|
769
|
+
...(recombination ? { recombination } : {}),
|
|
770
|
+
...(proceduralCompilation ? { proceduralCompilation } : {}),
|
|
771
|
+
...(cycleMetrics ? { cycleMetrics } : {}),
|
|
954
772
|
...(orphansPurged !== undefined ? { orphansPurged } : {}),
|
|
955
773
|
...(proposalsExpired !== undefined && proposalsExpired > 0 ? { proposalsExpired } : {}),
|
|
956
774
|
reflectCooldownActions: finalActions.filter((a) => a.mode === "reflect-cooldown").length,
|
|
957
775
|
reflectSkippedActions: finalActions.filter((a) => a.mode === "reflect-skipped").length,
|
|
958
776
|
reflectGuardRejectedActions: finalActions.filter((a) => a.mode === "reflect-guard-rejected").length,
|
|
959
777
|
...(() => {
|
|
960
|
-
const t =
|
|
778
|
+
const t = prepGateCount + loopGateCount + postLoopGateCount;
|
|
961
779
|
return t > 0 ? { gateAutoAcceptedCount: t } : {};
|
|
962
780
|
})(),
|
|
963
781
|
...(() => {
|
|
964
|
-
const f =
|
|
782
|
+
const f = prepGateFailedCount + loopGateFailedCount + postLoopGateFailedCount;
|
|
965
783
|
return f > 0 ? { gateAutoAcceptFailedCount: f } : {};
|
|
966
784
|
})(),
|
|
967
785
|
...(triageDrain
|
|
@@ -974,6 +792,10 @@ export async function akmImprove(options = {}) {
|
|
|
974
792
|
},
|
|
975
793
|
}
|
|
976
794
|
: {}),
|
|
795
|
+
...(preparation.proactiveMaintenance ? { proactiveMaintenance: preparation.proactiveMaintenance } : {}),
|
|
796
|
+
// #616 — report cycles run only when >1 so the default single-pass
|
|
797
|
+
// serialized envelope stays byte-identical to pre-#616 (AC1).
|
|
798
|
+
...(cyclesRun > 1 ? { cyclesRun } : {}),
|
|
977
799
|
...(options.runId !== undefined ? { runId: options.runId } : {}),
|
|
978
800
|
};
|
|
979
801
|
if (!result.dryRun)
|
|
@@ -984,57 +806,18 @@ export async function akmImprove(options = {}) {
|
|
|
984
806
|
warningCount: allWarnings.length,
|
|
985
807
|
orphansPurged: orphansPurged ?? 0,
|
|
986
808
|
}, eventsCtx);
|
|
987
|
-
// End-of-run BATCH auto-sync
|
|
988
|
-
// path (see write-source.ts case-3): the primary
|
|
989
|
-
// filesystem source during the run, then is committed
|
|
990
|
-
// the same `saveGitStash` that `akm sync` calls.
|
|
991
|
-
//
|
|
992
|
-
//
|
|
993
|
-
//
|
|
994
|
-
|
|
995
|
-
|
|
996
|
-
|
|
997
|
-
|
|
998
|
-
|
|
999
|
-
// `akm sync` via resolveWritableOverride().
|
|
1000
|
-
const writableOverride = resolveWritableOverride(_earlyConfig);
|
|
1001
|
-
const push = effectiveSync.push !== false;
|
|
1002
|
-
// `sync.message` may contain `{token}` placeholders (timestamp/date/time/
|
|
1003
|
-
// scope/refs/accepted) expanded against this run's results; the default
|
|
1004
|
-
// template has no tokens so it renders verbatim.
|
|
1005
|
-
const message = renderSyncCommitMessage(effectiveSync.message ?? "akm improve auto-sync", result, Date.now());
|
|
1006
|
-
try {
|
|
1007
|
-
// Pass primaryStashDir as the explicit commit target so the gate above
|
|
1008
|
-
// (which validated primaryStashDir via isGitBackedStash) and the commit
|
|
1009
|
-
// operate on the SAME directory — avoids divergence when a caller passes
|
|
1010
|
-
// a non-default options.stashDir (FIX 9).
|
|
1011
|
-
const syncResult = saveGitStashFn(undefined, message, writableOverride, { push, repoDir: primaryStashDir });
|
|
1012
|
-
result.sync = {
|
|
1013
|
-
committed: syncResult.committed,
|
|
1014
|
-
pushed: syncResult.pushed,
|
|
1015
|
-
skipped: syncResult.skipped,
|
|
1016
|
-
...(syncResult.reason !== undefined ? { reason: syncResult.reason } : {}),
|
|
1017
|
-
};
|
|
1018
|
-
appendEvent({
|
|
1019
|
-
eventType: "stash_synced",
|
|
1020
|
-
metadata: {
|
|
1021
|
-
committed: syncResult.committed,
|
|
1022
|
-
pushed: syncResult.pushed,
|
|
1023
|
-
skipped: syncResult.skipped,
|
|
1024
|
-
reason: syncResult.reason ?? null,
|
|
1025
|
-
},
|
|
1026
|
-
}, eventsCtx);
|
|
1027
|
-
}
|
|
1028
|
-
catch (syncErr) {
|
|
1029
|
-
const reason = syncErr instanceof Error ? syncErr.message : String(syncErr);
|
|
1030
|
-
warn(`improve: end-of-run stash sync failed (non-fatal): ${reason}`);
|
|
1031
|
-
result.sync = { committed: false, pushed: false, skipped: true, reason };
|
|
1032
|
-
appendEvent({
|
|
1033
|
-
eventType: "stash_synced",
|
|
1034
|
-
metadata: { committed: false, pushed: false, skipped: true, reason },
|
|
1035
|
-
}, eventsCtx);
|
|
1036
|
-
}
|
|
1037
|
-
}
|
|
809
|
+
// End-of-run BATCH auto-sync — the converged commit. Recognition is
|
|
810
|
+
// decoupled from the per-write path (see write-source.ts case-3): the primary
|
|
811
|
+
// stash writes as a filesystem source during the run, then is committed via
|
|
812
|
+
// the same `saveGitStash` that `akm sync` calls. The gating (git-backed
|
|
813
|
+
// primary stash, sync not disabled) and the NON-FATAL guarantee now live in
|
|
814
|
+
// `commitStashBatch` (#662); the inter-cycle and catch-path calls reuse it.
|
|
815
|
+
// dry-run already returned above, so this always runs on a completed live
|
|
816
|
+
// run. `result.sync` reflects this final commit (for a one-cycle run it is
|
|
817
|
+
// the only commit; for a multi-cycle run the earlier cycles were banked by
|
|
818
|
+
// the inter-cycle calls and this records the last batch). `result` carries
|
|
819
|
+
// the full `{accepted}`/`{refs}`/`{triage_*}` token data for the message.
|
|
820
|
+
result.sync = commitStashBatch(result);
|
|
1038
821
|
return result;
|
|
1039
822
|
}
|
|
1040
823
|
catch (err) {
|
|
@@ -1043,10 +826,17 @@ export async function akmImprove(options = {}) {
|
|
|
1043
826
|
eventType: "improve_failed",
|
|
1044
827
|
ref: scope.mode === "ref" ? scope.value : `improve:${scope.mode}:${scope.value ?? "all"}`,
|
|
1045
828
|
metadata: {
|
|
1046
|
-
error:
|
|
829
|
+
error: errMessage(err),
|
|
1047
830
|
durationMs: Date.now() - startMs,
|
|
1048
831
|
},
|
|
1049
832
|
}, eventsCtx);
|
|
833
|
+
// #662 crash/abort safety net: commit whatever this run already wrote to the
|
|
834
|
+
// primary stash BEFORE rethrowing, so an interrupted run (mid-cycle crash or
|
|
835
|
+
// a cooperative budget abort that surfaces as a throw) does not leave its
|
|
836
|
+
// writes uncommitted until a later clean run sweeps them up. Best-effort —
|
|
837
|
+
// `commitStashBatch` swallows its own errors and no-ops a clean tree, so this
|
|
838
|
+
// never masks or supersedes the original failure being rethrown below.
|
|
839
|
+
commitStashBatch({ scope, plannedRefs, runId: options.runId });
|
|
1050
840
|
throw err;
|
|
1051
841
|
}
|
|
1052
842
|
finally {
|
|
@@ -1056,15 +846,18 @@ export async function akmImprove(options = {}) {
|
|
|
1056
846
|
// O-1 (#364): Clear the budget abort timer so it does not keep the event
|
|
1057
847
|
// loop alive after the run completes.
|
|
1058
848
|
clearBudgetTimer();
|
|
1059
|
-
|
|
1060
|
-
|
|
849
|
+
// #607: release any per-process locks still held (backstop for error paths;
|
|
850
|
+
// the normal path already released each lock after its stage completed).
|
|
851
|
+
releaseAllProcessLocks();
|
|
852
|
+
// Drop ONLY our own process.exit backstop so it does not fire later (or
|
|
853
|
+
// accumulate across repeated in-process calls). Must NOT use
|
|
854
|
+
// removeAllListeners("exit") here: in the in-process model (tests and
|
|
855
|
+
// programmatic callers import cli.ts) that would silently destroy exit
|
|
856
|
+
// handlers owned by the host or other commands.
|
|
857
|
+
if (exitBackstop) {
|
|
858
|
+
process.removeListener("exit", exitBackstop);
|
|
859
|
+
exitBackstop = undefined;
|
|
1061
860
|
}
|
|
1062
|
-
catch {
|
|
1063
|
-
// ignore
|
|
1064
|
-
}
|
|
1065
|
-
// The normal path released the lock above; drop the process.exit backstop so
|
|
1066
|
-
// it does not fire later (or accumulate across repeated in-process calls).
|
|
1067
|
-
process.removeListener("exit", releaseLockOnExit);
|
|
1068
861
|
// I1: close the long-lived state.db connection opened at the top of the run.
|
|
1069
862
|
try {
|
|
1070
863
|
eventsDb?.close();
|
|
@@ -1091,7 +884,7 @@ function emitImproveCompletedEvent(result, durations, eventsCtx) {
|
|
|
1091
884
|
// Coarse audit buckets, derived from the SAME classifyImproveAction the
|
|
1092
885
|
// persisted metrics_json uses (state-db.ts#computeImproveRunMetrics) so the
|
|
1093
886
|
// emitted event and the stored row can never disagree.
|
|
1094
|
-
const classCounts = { accepted: 0, rejected: 0, error: 0, noop: 0 };
|
|
887
|
+
const classCounts = { accepted: 0, rejected: 0, skipped: 0, error: 0, noop: 0 };
|
|
1095
888
|
for (const action of result.actions ?? []) {
|
|
1096
889
|
classCounts[classifyImproveAction(action.mode)] += 1;
|
|
1097
890
|
// Per-variant counters for the event metadata. The default arm makes any
|
|
@@ -1136,6 +929,13 @@ function emitImproveCompletedEvent(result, durations, eventsCtx) {
|
|
|
1136
929
|
assertNever(action.mode);
|
|
1137
930
|
}
|
|
1138
931
|
}
|
|
932
|
+
// C1: distill-skipped rows are no longer in `result.actions` (folded into the
|
|
933
|
+
// bounded `distillSkipped` aggregate at assembly). Add the aggregate total to
|
|
934
|
+
// the per-variant counter AND the coarse `skipped` bucket so the emitted event
|
|
935
|
+
// still reports the true skipped volume.
|
|
936
|
+
const distillSkippedTotal = result.distillSkipped?.total ?? 0;
|
|
937
|
+
actionCounts.distillSkipped += distillSkippedTotal;
|
|
938
|
+
classCounts.skipped += distillSkippedTotal;
|
|
1139
939
|
appendEvent({
|
|
1140
940
|
eventType: "improve_completed",
|
|
1141
941
|
ref: result.scope.mode === "ref"
|
|
@@ -1158,6 +958,7 @@ function emitImproveCompletedEvent(result, durations, eventsCtx) {
|
|
|
1158
958
|
reflectGuardRejectedActions: actionCounts.reflectGuardRejected,
|
|
1159
959
|
acceptedActions: classCounts.accepted,
|
|
1160
960
|
rejectedActions: classCounts.rejected,
|
|
961
|
+
skippedActions: classCounts.skipped,
|
|
1161
962
|
noopActions: classCounts.noop,
|
|
1162
963
|
reflectsWithErrorContext: result.reflectsWithErrorContext ?? 0,
|
|
1163
964
|
coverageGapCount: result.coverageGaps?.length ?? 0,
|
|
@@ -1177,6 +978,11 @@ function emitImproveCompletedEvent(result, durations, eventsCtx) {
|
|
|
1177
978
|
memoryInferenceDurationMs: durations.memoryInferenceDurationMs,
|
|
1178
979
|
graphExtractionExtractedFiles: result.graphExtraction?.quality.extractedFiles ?? 0,
|
|
1179
980
|
graphExtractionDurationMs: durations.graphExtractionDurationMs,
|
|
981
|
+
// Layer-2 proactive-maintenance coverage (0 when the process is disabled
|
|
982
|
+
// or the run was ref-scoped) so a scheduled sweep's reach is trackable.
|
|
983
|
+
proactiveSelected: result.proactiveMaintenance?.selected ?? 0,
|
|
984
|
+
proactiveDueTotal: result.proactiveMaintenance?.dueTotal ?? 0,
|
|
985
|
+
proactiveNeverReflected: result.proactiveMaintenance?.neverReflected ?? 0,
|
|
1180
986
|
// New metrics for tuning the improve loop.
|
|
1181
987
|
...(durations.totalDurationMs !== undefined ? { durationMs: durations.totalDurationMs } : {}),
|
|
1182
988
|
...(durations.warningCount !== undefined ? { warningCount: durations.warningCount } : {}),
|
|
@@ -1191,1794 +997,6 @@ function emitImproveCompletedEvent(result, durations, eventsCtx) {
|
|
|
1191
997
|
},
|
|
1192
998
|
}, eventsCtx);
|
|
1193
999
|
}
|
|
1194
|
-
/**
|
|
1195
|
-
* Run (or gate-skip) the memory consolidation pass.
|
|
1196
|
-
*
|
|
1197
|
-
* #551 — two coordinated changes live here:
|
|
1198
|
-
*
|
|
1199
|
-
* 1. STRUCTURAL: this runs before extract in the improve pipeline (see
|
|
1200
|
-
* `runImprovePreparationStage`). Consolidation therefore only ever judges
|
|
1201
|
-
* PRIOR-run memories; current-run extract promotions are invisible to it.
|
|
1202
|
-
*
|
|
1203
|
-
* 2. SMARTER POOL-DELTA GATE: even among on-disk files, a memory whose only
|
|
1204
|
-
* post-`lastConsolidateTs` mtime bump came from its OWN auto-accept
|
|
1205
|
-
* promotion (i.e. it was just promoted by extract in the immediately
|
|
1206
|
-
* preceding run and has not had a full improve cycle to settle) does NOT
|
|
1207
|
-
* count as "work to do". We exclude those paths from the pool-delta check
|
|
1208
|
-
* using the `promoted` events already emitted with each promotion's
|
|
1209
|
-
* `assetPath`. A genuinely-settled prior memory — one edited by feedback,
|
|
1210
|
-
* reflect, manual edit, or simply older than the last consolidate — still
|
|
1211
|
-
* triggers the run. This is gate-option (a) from the issue (same-run /
|
|
1212
|
-
* adjacent-run promotion exclusion), chosen over option (b) because there
|
|
1213
|
-
* is no `extract_completed` event in the data model to gate against;
|
|
1214
|
-
* `promoted` events with `assetPath` already carry exactly the signal we
|
|
1215
|
-
* need, so the fix is non-invasive and provably correct.
|
|
1216
|
-
*/
|
|
1217
|
-
async function runConsolidationPass(args) {
|
|
1218
|
-
const { options, primaryStashDir, memorySummary, improveProfile, eventsCtx } = args;
|
|
1219
|
-
const baseConfig = options.config ?? loadConfig();
|
|
1220
|
-
const MEMORY_VOLUME_THRESHOLD = options.memoryVolumeConsolidationThreshold ?? 100;
|
|
1221
|
-
const hasLlm = !!(baseConfig.defaults?.llm || baseConfig.defaults?.agent);
|
|
1222
|
-
const volumeTriggered = typeof memorySummary.eligible === "number" && memorySummary.eligible > MEMORY_VOLUME_THRESHOLD && hasLlm;
|
|
1223
|
-
// When volume triggers a consolidation pass, force-enable the consolidate
|
|
1224
|
-
// process on the default improve profile so the gate accepts the run even
|
|
1225
|
-
// if the user's config disabled it. We synthesise a new profile override
|
|
1226
|
-
// rather than mutating connection settings.
|
|
1227
|
-
const consolidationConfig = volumeTriggered
|
|
1228
|
-
? {
|
|
1229
|
-
...baseConfig,
|
|
1230
|
-
profiles: {
|
|
1231
|
-
...(baseConfig.profiles ?? {}),
|
|
1232
|
-
improve: {
|
|
1233
|
-
...(baseConfig.profiles?.improve ?? {}),
|
|
1234
|
-
default: {
|
|
1235
|
-
...(baseConfig.profiles?.improve?.default ?? {}),
|
|
1236
|
-
processes: {
|
|
1237
|
-
...(baseConfig.profiles?.improve?.default?.processes ?? {}),
|
|
1238
|
-
consolidate: {
|
|
1239
|
-
...(baseConfig.profiles?.improve?.default?.processes?.consolidate ?? {}),
|
|
1240
|
-
enabled: true,
|
|
1241
|
-
},
|
|
1242
|
-
},
|
|
1243
|
-
},
|
|
1244
|
-
},
|
|
1245
|
-
},
|
|
1246
|
-
}
|
|
1247
|
-
: baseConfig;
|
|
1248
|
-
// 0.8.0 pool-delta gate for consolidate: re-eligible iff at least one
|
|
1249
|
-
// memory file has been updated since the most recent successful
|
|
1250
|
-
// consolidate_completed event. Time-based cooldowns produced the same
|
|
1251
|
-
// synchronised-wave failure mode the reflect/distill cooldowns did; the
|
|
1252
|
-
// pool-delta gate ties consolidation to actual work-to-do.
|
|
1253
|
-
const recentConsolidations = readEvents({ type: "consolidate_completed" });
|
|
1254
|
-
const lastConsolidation = recentConsolidations.events
|
|
1255
|
-
.filter((e) => e.metadata?.processed && Number(e.metadata.processed) > 0)
|
|
1256
|
-
.sort((a, b) => new Date(b.ts ?? 0).getTime() - new Date(a.ts ?? 0).getTime())[0];
|
|
1257
|
-
const lastConsolidateTs = lastConsolidation?.ts;
|
|
1258
|
-
// #551 smarter gate: build the set of memory asset paths whose only delta
|
|
1259
|
-
// since the last consolidate is their OWN auto-accept promotion. Those files
|
|
1260
|
-
// have not had a full improve cycle to settle, so they offer no merge /
|
|
1261
|
-
// contradiction candidates yet — excluding them stops the gate firing on
|
|
1262
|
-
// freshly-promoted single-source memories. We read `promoted` events emitted
|
|
1263
|
-
// after the last consolidate; each carries the written `assetPath`.
|
|
1264
|
-
const promotedSinceConsolidate = (() => {
|
|
1265
|
-
const paths = new Set();
|
|
1266
|
-
try {
|
|
1267
|
-
const promoted = readEvents({
|
|
1268
|
-
type: "promoted",
|
|
1269
|
-
...(lastConsolidateTs ? { since: lastConsolidateTs } : {}),
|
|
1270
|
-
}).events;
|
|
1271
|
-
for (const e of promoted) {
|
|
1272
|
-
const ap = e.metadata?.assetPath;
|
|
1273
|
-
if (typeof ap === "string" && ap.length > 0)
|
|
1274
|
-
paths.add(path.resolve(ap));
|
|
1275
|
-
}
|
|
1276
|
-
}
|
|
1277
|
-
catch {
|
|
1278
|
-
// best-effort: if the events query fails, fall back to no exclusions
|
|
1279
|
-
// (preserves pre-#551 behaviour rather than over-skipping).
|
|
1280
|
-
}
|
|
1281
|
-
return paths;
|
|
1282
|
-
})();
|
|
1283
|
-
// Pool-delta: any memory file with mtime > lastConsolidateTs flags work to do,
|
|
1284
|
-
// EXCEPT files whose only post-consolidate change was their own promotion.
|
|
1285
|
-
// Using file mtime keeps this query DB-free and matches what the indexer
|
|
1286
|
-
// already uses as the canonical `memory.updated_at` proxy.
|
|
1287
|
-
//
|
|
1288
|
-
// Bootstrap: when no successful consolidate_completed event has ever been
|
|
1289
|
-
// recorded, we cannot evaluate the pool-delta — treat as eligible so a
|
|
1290
|
-
// fresh stash runs consolidate once before the steady-state gate kicks in.
|
|
1291
|
-
const memoryUpdatedAfterLastConsolidate = (() => {
|
|
1292
|
-
if (volumeTriggered)
|
|
1293
|
-
return true; // volume override forces the run regardless.
|
|
1294
|
-
if (!lastConsolidateTs)
|
|
1295
|
-
return true; // bootstrap path: never consolidated.
|
|
1296
|
-
if (!primaryStashDir)
|
|
1297
|
-
return false;
|
|
1298
|
-
const memoriesDir = path.join(primaryStashDir, "memories");
|
|
1299
|
-
if (!fs.existsSync(memoriesDir))
|
|
1300
|
-
return false;
|
|
1301
|
-
try {
|
|
1302
|
-
return fs.readdirSync(memoriesDir).some((f) => {
|
|
1303
|
-
if (!f.endsWith(".md"))
|
|
1304
|
-
return false;
|
|
1305
|
-
const filePath = path.join(memoriesDir, f);
|
|
1306
|
-
// #551: skip files that were only touched by their own promotion this
|
|
1307
|
-
// cohort — they have no settled merge/contradiction candidates yet.
|
|
1308
|
-
if (promotedSinceConsolidate.has(path.resolve(filePath)))
|
|
1309
|
-
return false;
|
|
1310
|
-
try {
|
|
1311
|
-
return fs.statSync(filePath).mtime.toISOString() > lastConsolidateTs;
|
|
1312
|
-
}
|
|
1313
|
-
catch {
|
|
1314
|
-
return false;
|
|
1315
|
-
}
|
|
1316
|
-
});
|
|
1317
|
-
}
|
|
1318
|
-
catch {
|
|
1319
|
-
return false;
|
|
1320
|
-
}
|
|
1321
|
-
})();
|
|
1322
|
-
const consolidationOnCooldown = !volumeTriggered && !memoryUpdatedAfterLastConsolidate;
|
|
1323
|
-
// Profile gate: if profile explicitly disables consolidate, skip the entire pass.
|
|
1324
|
-
const consolidateDisabledByProfile = improveProfile?.processes?.consolidate?.enabled === false;
|
|
1325
|
-
// #553 minPoolSize guard: skip consolidation when the eligible memory pool is
|
|
1326
|
-
// below a minimum size, rather than spending an LLM pass on a handful of
|
|
1327
|
-
// memories. This is an INDEPENDENT skip condition from #551's mtime pool-delta
|
|
1328
|
-
// gate — either can skip. Default 500; `minPoolSize: 0` disables the guard.
|
|
1329
|
-
// Evaluated against the eligible-pool count BEFORE entering the LLM loop so a
|
|
1330
|
-
// skip costs ZERO LLM calls.
|
|
1331
|
-
const CONSOLIDATE_DEFAULT_MIN_POOL_SIZE = 500;
|
|
1332
|
-
const configuredMinPoolSize = improveProfile?.processes?.consolidate?.minPoolSize;
|
|
1333
|
-
const minPoolSize = typeof configuredMinPoolSize === "number" ? configuredMinPoolSize : CONSOLIDATE_DEFAULT_MIN_POOL_SIZE;
|
|
1334
|
-
const eligiblePoolSize = typeof memorySummary.eligible === "number" ? memorySummary.eligible : 0;
|
|
1335
|
-
// volumeTriggered means the pool already exceeds the volume threshold (100),
|
|
1336
|
-
// so a force-triggered run never trips the pool-size guard. The guard only
|
|
1337
|
-
// engages when minPoolSize > 0 and the eligible pool is strictly below it.
|
|
1338
|
-
const poolBelowMinSize = !volumeTriggered && minPoolSize > 0 && eligiblePoolSize < minPoolSize;
|
|
1339
|
-
let consolidation = {
|
|
1340
|
-
schemaVersion: 1,
|
|
1341
|
-
ok: true,
|
|
1342
|
-
shape: "consolidate-result",
|
|
1343
|
-
dryRun: false,
|
|
1344
|
-
previewOnly: false,
|
|
1345
|
-
target: "",
|
|
1346
|
-
processed: 0,
|
|
1347
|
-
merged: 0,
|
|
1348
|
-
deleted: 0,
|
|
1349
|
-
promoted: [],
|
|
1350
|
-
contradicted: 0,
|
|
1351
|
-
warnings: [],
|
|
1352
|
-
durationMs: 0,
|
|
1353
|
-
};
|
|
1354
|
-
let gateAutoAcceptedCount = 0;
|
|
1355
|
-
let gateAutoAcceptFailedCount = 0;
|
|
1356
|
-
const consolidateGateCfg = makeGateConfig("consolidate", {
|
|
1357
|
-
globalThreshold: options.autoAccept,
|
|
1358
|
-
dryRun: options.dryRun ?? false,
|
|
1359
|
-
stashDir: primaryStashDir,
|
|
1360
|
-
config: consolidationConfig,
|
|
1361
|
-
eventsCtx,
|
|
1362
|
-
}, { minimumThreshold: 95 });
|
|
1363
|
-
if (consolidateDisabledByProfile) {
|
|
1364
|
-
info("[improve] consolidation skipped (disabled by improve profile)");
|
|
1365
|
-
}
|
|
1366
|
-
else if (poolBelowMinSize) {
|
|
1367
|
-
// #553: eligible pool below the configured minimum — skip with zero LLM
|
|
1368
|
-
// calls. Reuse the #551 `improve_skipped` emission path so health surfaces
|
|
1369
|
-
// it via the dynamic skipReasons aggregation under `pool_below_min_size`.
|
|
1370
|
-
appendEvent({
|
|
1371
|
-
eventType: "improve_skipped",
|
|
1372
|
-
ref: "memory:_consolidation",
|
|
1373
|
-
metadata: {
|
|
1374
|
-
reason: "pool_below_min_size",
|
|
1375
|
-
poolSize: eligiblePoolSize,
|
|
1376
|
-
minPoolSize,
|
|
1377
|
-
},
|
|
1378
|
-
}, eventsCtx);
|
|
1379
|
-
info(`[improve] consolidation skipped (pool ${eligiblePoolSize} < minPoolSize ${minPoolSize})`);
|
|
1380
|
-
}
|
|
1381
|
-
else if (!consolidationOnCooldown) {
|
|
1382
|
-
consolidation = await withLlmStage("consolidate", () => akmConsolidate({
|
|
1383
|
-
...options.consolidateOptions,
|
|
1384
|
-
config: consolidationConfig,
|
|
1385
|
-
stashDir: options.stashDir,
|
|
1386
|
-
autoTriggered: volumeTriggered,
|
|
1387
|
-
// Tie consolidate proposals back to this improve invocation so
|
|
1388
|
-
// accept-rate-per-run aggregation works. Mirrors reflect/propose/extract.
|
|
1389
|
-
sourceRun: `consolidate-${Date.now()}`,
|
|
1390
|
-
// Pass profile-configured options. incrementalSince narrows the pool to
|
|
1391
|
-
// recently-changed memories + graph neighbours — use this for frequent
|
|
1392
|
-
// passes (quick-shredder). Leave absent in the nightly default profile for
|
|
1393
|
-
// a full-pool sweep that catches stale-but-unmerged duplicates.
|
|
1394
|
-
incrementalSince: improveProfile?.processes?.consolidate?.incrementalSince,
|
|
1395
|
-
limit: improveProfile?.processes?.consolidate?.limit,
|
|
1396
|
-
neighborsPerChanged: improveProfile?.processes?.consolidate?.neighborsPerChanged,
|
|
1397
|
-
maxChunkSize: improveProfile?.processes?.consolidate?.maxChunkSize,
|
|
1398
|
-
// Honor profile.autoAccept (already merged into options.autoAccept at the
|
|
1399
|
-
// top of akmImprove). The CLI parser always supplies 90 when --auto-accept
|
|
1400
|
-
// is absent, so ?? 90 is not needed here and would prevent --auto-accept=false
|
|
1401
|
-
// (which maps to undefined) from disabling consolidation auto-accept.
|
|
1402
|
-
// options.consolidateOptions.autoAccept (if explicitly provided by caller)
|
|
1403
|
-
// still wins because the spread above runs first.
|
|
1404
|
-
autoAccept: options.consolidateOptions?.autoAccept ?? options.autoAccept,
|
|
1405
|
-
}));
|
|
1406
|
-
{
|
|
1407
|
-
const consolidateGr = await runAutoAcceptGate(consolidation.promoted.map((proposalId) => {
|
|
1408
|
-
try {
|
|
1409
|
-
if (!primaryStashDir)
|
|
1410
|
-
return { proposalId, confidence: undefined };
|
|
1411
|
-
const proposal = getProposal(primaryStashDir, proposalId);
|
|
1412
|
-
return { proposalId, confidence: proposal.confidence };
|
|
1413
|
-
}
|
|
1414
|
-
catch {
|
|
1415
|
-
return { proposalId, confidence: undefined };
|
|
1416
|
-
}
|
|
1417
|
-
}), consolidateGateCfg);
|
|
1418
|
-
gateAutoAcceptedCount += consolidateGr.promoted.length;
|
|
1419
|
-
gateAutoAcceptFailedCount += consolidateGr.failed.length;
|
|
1420
|
-
}
|
|
1421
|
-
if (consolidation.processed > 0) {
|
|
1422
|
-
appendEvent({
|
|
1423
|
-
eventType: "consolidate_completed",
|
|
1424
|
-
ref: "memory:_consolidation",
|
|
1425
|
-
metadata: { processed: consolidation.processed, merged: consolidation.merged },
|
|
1426
|
-
}, eventsCtx);
|
|
1427
|
-
}
|
|
1428
|
-
}
|
|
1429
|
-
else {
|
|
1430
|
-
appendEvent({
|
|
1431
|
-
eventType: "improve_skipped",
|
|
1432
|
-
ref: "memory:_consolidation",
|
|
1433
|
-
metadata: {
|
|
1434
|
-
reason: "consolidation_no_memory_updates",
|
|
1435
|
-
lastEventTs: lastConsolidation?.ts ?? null,
|
|
1436
|
-
},
|
|
1437
|
-
}, eventsCtx);
|
|
1438
|
-
info("[improve] consolidation skipped (no memory updates since last run)");
|
|
1439
|
-
}
|
|
1440
|
-
// D9: track whether consolidation wrote any data so graph extraction can reindex if needed
|
|
1441
|
-
const consolidationRan = !consolidateDisabledByProfile && !poolBelowMinSize && !consolidationOnCooldown && consolidation.processed > 0;
|
|
1442
|
-
return { consolidation, consolidationRan, gateAutoAcceptedCount, gateAutoAcceptFailedCount };
|
|
1443
|
-
}
|
|
1444
|
-
async function runImprovePreparationStage(args) {
|
|
1445
|
-
const { scope, options, plannedRefs, memoryCleanupPlan, primaryStashDir, memorySummary, reindexFn, startMs, budgetMs, eventsCtx, initialCleanupWarnings, improveProfile, } = args;
|
|
1446
|
-
const actions = [];
|
|
1447
|
-
const cleanupWarnings = initialCleanupWarnings ? [...initialCleanupWarnings] : [];
|
|
1448
|
-
// Phase 0 — MEMORY.md budget check (200-line cap; warn at 180)
|
|
1449
|
-
let memoryIndexHealth;
|
|
1450
|
-
if (primaryStashDir) {
|
|
1451
|
-
const memoryMdPath = path.join(primaryStashDir, "memories", "MEMORY.md");
|
|
1452
|
-
if (fs.existsSync(memoryMdPath)) {
|
|
1453
|
-
try {
|
|
1454
|
-
const lines = fs.readFileSync(memoryMdPath, "utf8").split("\n").length;
|
|
1455
|
-
const overBudget = lines >= 180;
|
|
1456
|
-
memoryIndexHealth = { lineCount: lines, overBudget };
|
|
1457
|
-
if (overBudget) {
|
|
1458
|
-
cleanupWarnings.push(`MEMORY.md has ${lines} lines (budget: 200). Consolidation strongly recommended.`);
|
|
1459
|
-
}
|
|
1460
|
-
}
|
|
1461
|
-
catch {
|
|
1462
|
-
// best-effort
|
|
1463
|
-
}
|
|
1464
|
-
}
|
|
1465
|
-
}
|
|
1466
|
-
// Phase 0.3 — memory consolidation pass (#551).
|
|
1467
|
-
//
|
|
1468
|
-
// Consolidation runs BEFORE the session-extract pass. This is the structural
|
|
1469
|
-
// half of the #551 fix: extract auto-accept writes brand-new memory .md files
|
|
1470
|
-
// on every run, which previously made the consolidation pool-delta gate fire
|
|
1471
|
-
// unconditionally (any new file => "memory updated since last consolidate").
|
|
1472
|
-
// By running consolidation first, the gate and akmConsolidate only ever see
|
|
1473
|
-
// memories that existed at the start of the run — current-run extract
|
|
1474
|
-
// promotions are not on disk yet. The complementary smarter-gate logic
|
|
1475
|
-
// (excluding adjacent-run promotions) lives in `runConsolidationPass`.
|
|
1476
|
-
const consolidationPass = await runConsolidationPass({
|
|
1477
|
-
options,
|
|
1478
|
-
primaryStashDir,
|
|
1479
|
-
memorySummary,
|
|
1480
|
-
improveProfile,
|
|
1481
|
-
eventsCtx,
|
|
1482
|
-
});
|
|
1483
|
-
// Phase 0.4 — session-extract pass.
|
|
1484
|
-
//
|
|
1485
|
-
// Reads native session files (claude-code JSONL, opencode storage tree)
|
|
1486
|
-
// through the SessionLogHarness registry, pre-filters noise, and asks a
|
|
1487
|
-
// bounded in-tree LLM to produce candidate memory/lesson/knowledge
|
|
1488
|
-
// proposals for content the agent did NOT preserve via inline `akm remember`
|
|
1489
|
-
// / `akm feedback` invocations. Replaces the akm-plugin session-checkpoint
|
|
1490
|
-
// hook with an on-demand pull pipeline.
|
|
1491
|
-
//
|
|
1492
|
-
// Default-on; opt out via the ACTIVE profile's `processes.extract.enabled: false`
|
|
1493
|
-
// (#593: the gate respects the resolved improve profile, not just the
|
|
1494
|
-
// hardcoded `default` profile path the legacy feature flag reads).
|
|
1495
|
-
// Each available harness gets one call with the default --since window;
|
|
1496
|
-
// already-seen sessions (tracked in state.db.extract_sessions_seen) are
|
|
1497
|
-
// skipped automatically so re-runs don't burn LLM calls on unchanged data.
|
|
1498
|
-
//
|
|
1499
|
-
// Failures are non-fatal — one harness throwing doesn't abort improve.
|
|
1500
|
-
// The extract envelope's own `warnings` field surfaces what went wrong.
|
|
1501
|
-
let extractResults;
|
|
1502
|
-
// Seed the preparation-stage gate counters with consolidation's auto-accept
|
|
1503
|
-
// gate results (#551: consolidation now runs in this stage), then accumulate
|
|
1504
|
-
// extract's gate results on top.
|
|
1505
|
-
let gateAutoAcceptedCount = consolidationPass.gateAutoAcceptedCount;
|
|
1506
|
-
let gateAutoAcceptFailedCount = consolidationPass.gateAutoAcceptFailedCount;
|
|
1507
|
-
const extractConfig = options.config ?? loadConfig();
|
|
1508
|
-
const extractGateCfg = makeGateConfig("extract", {
|
|
1509
|
-
globalThreshold: options.autoAccept,
|
|
1510
|
-
dryRun: options.dryRun ?? false,
|
|
1511
|
-
stashDir: primaryStashDir,
|
|
1512
|
-
config: extractConfig,
|
|
1513
|
-
eventsCtx,
|
|
1514
|
-
});
|
|
1515
|
-
// #554 minNewSessions gate: skip the entire extract pass (ensureIndex was
|
|
1516
|
-
// already done upstream; here we elide every akmExtract/processSession call)
|
|
1517
|
-
// when the NEW (unseen, in-window) candidate-session pool is below a minimum.
|
|
1518
|
-
// 22% of improve runs produce zero memory-inference writes because extract
|
|
1519
|
-
// finds no new sessions, yet still burns the full extract pipeline. Default 0
|
|
1520
|
-
// (disabled) preserves existing always-run behaviour; only opted-in profiles
|
|
1521
|
-
// (e.g. `frequent`) set it. Evaluated BEFORE any LLM call so a skip costs zero
|
|
1522
|
-
// LLM work AND writes nothing — which also means no extract auto-accept bumps
|
|
1523
|
-
// memory mtimes, so a skipped extract never flags work for the NEXT run's
|
|
1524
|
-
// consolidation mtime-gate (the downstream trigger #554 asks us to suppress).
|
|
1525
|
-
const EXTRACT_DEFAULT_MIN_NEW_SESSIONS = 0;
|
|
1526
|
-
const configuredMinNewSessions = extractConfig.profiles?.improve?.default?.processes?.extract?.minNewSessions;
|
|
1527
|
-
const minNewSessions = typeof configuredMinNewSessions === "number" ? configuredMinNewSessions : EXTRACT_DEFAULT_MIN_NEW_SESSIONS;
|
|
1528
|
-
// #593: gate on BOTH the legacy feature flag (which only reads
|
|
1529
|
-
// `profiles.improve.default.processes.extract.enabled` — kept for back-compat
|
|
1530
|
-
// with users who disable extract via the default-profile path) AND the active
|
|
1531
|
-
// resolved profile. Without the second check a non-default profile setting
|
|
1532
|
-
// `extract.enabled: false` (e.g. the built-in `quick`) was silently ignored
|
|
1533
|
-
// and extract ran on every improve call regardless.
|
|
1534
|
-
if (isLlmFeatureEnabled(extractConfig, "session_extraction") && resolveProcessEnabled("extract", improveProfile)) {
|
|
1535
|
-
const availableHarnesses = options.extractHarnesses ?? getAvailableHarnesses();
|
|
1536
|
-
// The guard engages only when minNewSessions > 0; 0 disables it entirely.
|
|
1537
|
-
let belowMinNewSessions = false;
|
|
1538
|
-
if (minNewSessions > 0 && availableHarnesses.length > 0) {
|
|
1539
|
-
const countFn = options.extractCandidateCountFn ?? countNewExtractCandidates;
|
|
1540
|
-
const newCandidateCount = countFn(extractConfig, {
|
|
1541
|
-
...(options.extractHarnesses ? { harnesses: options.extractHarnesses } : {}),
|
|
1542
|
-
// C2: pin the candidate-count state.db open to the boundary-resolved path.
|
|
1543
|
-
...(eventsCtx?.dbPath ? { stateDbPath: eventsCtx.dbPath } : {}),
|
|
1544
|
-
});
|
|
1545
|
-
if (newCandidateCount < minNewSessions) {
|
|
1546
|
-
belowMinNewSessions = true;
|
|
1547
|
-
// Reuse the #551/#553 `improve_skipped` emission path so health's dynamic
|
|
1548
|
-
// skipReasons aggregation surfaces this under `below_min_new_sessions`.
|
|
1549
|
-
appendEvent({
|
|
1550
|
-
eventType: "improve_skipped",
|
|
1551
|
-
ref: "memory:_extract",
|
|
1552
|
-
metadata: {
|
|
1553
|
-
reason: "below_min_new_sessions",
|
|
1554
|
-
newSessions: newCandidateCount,
|
|
1555
|
-
minNewSessions,
|
|
1556
|
-
},
|
|
1557
|
-
}, eventsCtx);
|
|
1558
|
-
info(`[improve] extract skipped (new sessions ${newCandidateCount} < minNewSessions ${minNewSessions})`);
|
|
1559
|
-
}
|
|
1560
|
-
}
|
|
1561
|
-
if (!belowMinNewSessions && availableHarnesses.length > 0) {
|
|
1562
|
-
extractResults = [];
|
|
1563
|
-
for (const h of availableHarnesses) {
|
|
1564
|
-
try {
|
|
1565
|
-
const result = await withLlmStage("session-extraction", () => akmExtract({
|
|
1566
|
-
type: h.name,
|
|
1567
|
-
...(primaryStashDir !== undefined ? { stashDir: primaryStashDir } : {}),
|
|
1568
|
-
config: extractConfig,
|
|
1569
|
-
dryRun: options.dryRun ?? false,
|
|
1570
|
-
...(options.extractHarnesses ? { harnesses: options.extractHarnesses } : {}),
|
|
1571
|
-
// C2: pin extract's skip-tracking state.db open to the boundary path.
|
|
1572
|
-
...(eventsCtx?.dbPath ? { stateDbPath: eventsCtx.dbPath } : {}),
|
|
1573
|
-
}));
|
|
1574
|
-
extractResults.push(result);
|
|
1575
|
-
{
|
|
1576
|
-
const gr = await runAutoAcceptGate(primaryStashDir
|
|
1577
|
-
? result.proposals.map((proposalId) => {
|
|
1578
|
-
const proposal = getProposal(primaryStashDir, proposalId);
|
|
1579
|
-
return { proposalId, confidence: resolveExtractConfidence(proposal) };
|
|
1580
|
-
})
|
|
1581
|
-
: [], extractGateCfg);
|
|
1582
|
-
gateAutoAcceptedCount += gr.promoted.length;
|
|
1583
|
-
gateAutoAcceptFailedCount += gr.failed.length;
|
|
1584
|
-
}
|
|
1585
|
-
}
|
|
1586
|
-
catch (err) {
|
|
1587
|
-
const msg = err instanceof Error ? err.message : String(err);
|
|
1588
|
-
cleanupWarnings.push(`extract(${h.name}) failed: ${msg}`);
|
|
1589
|
-
}
|
|
1590
|
-
}
|
|
1591
|
-
if (extractResults.length === 0) {
|
|
1592
|
-
// All harnesses threw — clear so the envelope's `extract` field is
|
|
1593
|
-
// absent rather than misleadingly empty.
|
|
1594
|
-
extractResults = undefined;
|
|
1595
|
-
}
|
|
1596
|
-
}
|
|
1597
|
-
}
|
|
1598
|
-
// Backlog drain: gate any pending extract proposals that weren't created in
|
|
1599
|
-
// this run (i.e. pre-date the gate or were produced by a run that timed out
|
|
1600
|
-
// before the gate fired). Without this, eligible proposals accumulate
|
|
1601
|
-
// indefinitely — the fresh-gate only covers the current run's output.
|
|
1602
|
-
if (primaryStashDir && !options.dryRun && options.autoAccept !== undefined) {
|
|
1603
|
-
const freshIds = new Set((extractResults ?? []).flatMap((r) => r.proposals));
|
|
1604
|
-
const backlog = listProposals(primaryStashDir, { status: "pending" }).filter((p) => p.source === "extract" && !freshIds.has(p.id));
|
|
1605
|
-
if (backlog.length > 0) {
|
|
1606
|
-
const backlogCandidates = backlog.map((p) => ({
|
|
1607
|
-
proposalId: p.id,
|
|
1608
|
-
confidence: resolveExtractConfidence(p),
|
|
1609
|
-
}));
|
|
1610
|
-
const backlogGr = await runAutoAcceptGate(backlogCandidates, extractGateCfg);
|
|
1611
|
-
gateAutoAcceptedCount += backlogGr.promoted.length;
|
|
1612
|
-
gateAutoAcceptFailedCount += backlogGr.failed.length;
|
|
1613
|
-
}
|
|
1614
|
-
}
|
|
1615
|
-
// eligibleCount = raw pre-filter count (before cooldown/signal/cleanup filters).
|
|
1616
|
-
// improve_completed.plannedRefs = post-filter count of refs that actually entered the loop.
|
|
1617
|
-
appendEvent({
|
|
1618
|
-
eventType: "improve_invoked",
|
|
1619
|
-
ref: scope.mode === "ref" ? scope.value : `improve:${scope.mode}:${scope.value ?? "all"}`,
|
|
1620
|
-
metadata: { scope, dryRun: options.dryRun ?? false, eligibleCount: plannedRefs.length },
|
|
1621
|
-
}, eventsCtx);
|
|
1622
|
-
// ensureIndex now runs in akmImprove() BEFORE collectEligibleRefs so the
|
|
1623
|
-
// eligible-ref query sees a populated `entries` table on the very first
|
|
1624
|
-
// pass after a DB version upgrade (#339). Any failure messages from that
|
|
1625
|
-
// earlier call were threaded in via args.initialCleanupWarnings.
|
|
1626
|
-
let appliedCleanup;
|
|
1627
|
-
try {
|
|
1628
|
-
appliedCleanup =
|
|
1629
|
-
primaryStashDir && memoryCleanupPlan ? applyMemoryCleanup(primaryStashDir, memoryCleanupPlan) : undefined;
|
|
1630
|
-
}
|
|
1631
|
-
catch (err) {
|
|
1632
|
-
cleanupWarnings.push(`applyMemoryCleanup failed: ${err instanceof Error ? err.message : String(err)}`);
|
|
1633
|
-
}
|
|
1634
|
-
const archivedRefs = appliedCleanup?.archived.map((record) => record.ref) ?? [];
|
|
1635
|
-
const removed = new Set(archivedRefs);
|
|
1636
|
-
const postCleanupRefs = archivedRefs.length === 0 ? plannedRefs : plannedRefs.filter((r) => !removed.has(r.ref));
|
|
1637
|
-
// ── Phase 1: validation pass + schema repair (run on full postCleanupRefs) ──
|
|
1638
|
-
// Identifies refs whose on-disk asset has structural problems. Validation
|
|
1639
|
-
// failures are excluded from every downstream bucket. Run early so the
|
|
1640
|
-
// cooldown partition operates on a clean set.
|
|
1641
|
-
if (appliedCleanup) {
|
|
1642
|
-
for (const candidate of memoryCleanupPlan?.pruneCandidates ?? []) {
|
|
1643
|
-
const archived = appliedCleanup.archived.find((record) => record.ref === candidate.ref);
|
|
1644
|
-
if (!archived)
|
|
1645
|
-
continue;
|
|
1646
|
-
actions.push({
|
|
1647
|
-
ref: candidate.ref,
|
|
1648
|
-
mode: "memory-prune",
|
|
1649
|
-
result: { ok: true, pruned: true, reason: candidate.reason },
|
|
1650
|
-
});
|
|
1651
|
-
}
|
|
1652
|
-
if ((appliedCleanup.archived.length > 0 || appliedCleanup.beliefStateTransitions.length > 0) && primaryStashDir) {
|
|
1653
|
-
try {
|
|
1654
|
-
await reindexFn({ stashDir: primaryStashDir });
|
|
1655
|
-
}
|
|
1656
|
-
catch (err) {
|
|
1657
|
-
cleanupWarnings.push(`reindex after cleanup failed: ${err instanceof Error ? err.message : String(err)}`);
|
|
1658
|
-
}
|
|
1659
|
-
}
|
|
1660
|
-
}
|
|
1661
|
-
const validationFailures = [];
|
|
1662
|
-
for (const candidate of postCleanupRefs) {
|
|
1663
|
-
try {
|
|
1664
|
-
// #591: use the path pre-resolved at planning time when it is still on
|
|
1665
|
-
// disk — a serial async DB lookup per ref cost ~500 s on a 9 000-ref
|
|
1666
|
-
// stash. Fall back to findAssetFilePath only for refs that bypassed
|
|
1667
|
-
// collectEligibleRefs' index scan or whose file moved since planning.
|
|
1668
|
-
const filePath = candidate.filePath && fs.existsSync(candidate.filePath)
|
|
1669
|
-
? candidate.filePath
|
|
1670
|
-
: await findAssetFilePath(candidate.ref, options.stashDir);
|
|
1671
|
-
if (!filePath) {
|
|
1672
|
-
validationFailures.push({ ref: candidate.ref, reason: "file not found on disk" });
|
|
1673
|
-
continue;
|
|
1674
|
-
}
|
|
1675
|
-
if (path.extname(filePath).toLowerCase() !== ".md") {
|
|
1676
|
-
continue;
|
|
1677
|
-
}
|
|
1678
|
-
if (isLessonCandidate(candidate.ref)) {
|
|
1679
|
-
const raw = fs.readFileSync(filePath, "utf8");
|
|
1680
|
-
const fm = parseFrontmatter(raw).data;
|
|
1681
|
-
if (!fm.description)
|
|
1682
|
-
validationFailures.push({ ref: candidate.ref, reason: "missing description" });
|
|
1683
|
-
}
|
|
1684
|
-
}
|
|
1685
|
-
catch (e) {
|
|
1686
|
-
validationFailures.push({ ref: candidate.ref, reason: String(e) });
|
|
1687
|
-
}
|
|
1688
|
-
}
|
|
1689
|
-
if (validationFailures.length > 0) {
|
|
1690
|
-
info(`[improve] ${validationFailures.length} assets have validation issues (will attempt schema repair):`);
|
|
1691
|
-
for (const f of validationFailures)
|
|
1692
|
-
info(` ${f.ref}: ${f.reason}`);
|
|
1693
|
-
}
|
|
1694
|
-
let schemaRepairs = [];
|
|
1695
|
-
let repairedRefs = new Set();
|
|
1696
|
-
// Schema repair pass: attempt to fix validation failures via LLM before skipping.
|
|
1697
|
-
if (validationFailures.length > 0 && options.repairValidationFailures !== false) {
|
|
1698
|
-
const baseConfigForRepair = options.config ?? loadConfig();
|
|
1699
|
-
const llmCfg = getDefaultLlmConfig(baseConfigForRepair);
|
|
1700
|
-
if (llmCfg) {
|
|
1701
|
-
const result = await runSchemaRepairPass(validationFailures, {
|
|
1702
|
-
startMs,
|
|
1703
|
-
budgetMs,
|
|
1704
|
-
llmConfig: llmCfg,
|
|
1705
|
-
stashDir: options.stashDir,
|
|
1706
|
-
findFilePath: findAssetFilePath,
|
|
1707
|
-
isLessonCandidateFn: isLessonCandidate,
|
|
1708
|
-
});
|
|
1709
|
-
schemaRepairs = result.repairs;
|
|
1710
|
-
repairedRefs = result.repairedRefs;
|
|
1711
|
-
}
|
|
1712
|
-
}
|
|
1713
|
-
const validationFailureRefs = new Set(validationFailures.filter((f) => !repairedRefs.has(f.ref)).map((f) => f.ref));
|
|
1714
|
-
if (repairedRefs.size > 0) {
|
|
1715
|
-
info(`[improve] schema repair fixed ${repairedRefs.size}/${validationFailures.length} validation failures; ${validationFailureRefs.size} remain`);
|
|
1716
|
-
}
|
|
1717
|
-
// Phase 0.5 — structural hygiene pass
|
|
1718
|
-
let lintSummary;
|
|
1719
|
-
if (primaryStashDir) {
|
|
1720
|
-
try {
|
|
1721
|
-
const lintResult = akmLint({ fix: true, dir: primaryStashDir });
|
|
1722
|
-
lintSummary = { fixed: lintResult.summary.fixed, flagged: lintResult.summary.flagged };
|
|
1723
|
-
}
|
|
1724
|
-
catch {
|
|
1725
|
-
// lint is best-effort; never block improve
|
|
1726
|
-
}
|
|
1727
|
-
}
|
|
1728
|
-
// O-5 / #378: Per-originator rolling error windows.
|
|
1729
|
-
// Reflexion (arXiv:2303.11366) warns that cross-task verbal critique
|
|
1730
|
-
// contamination degrades below single-shot baseline. Each originator key
|
|
1731
|
-
// ("schema-repair", "reflect") maintains its own rolling window so that
|
|
1732
|
-
// schema-repair failures are not injected as avoidPatterns into reflect calls.
|
|
1733
|
-
const recentErrors = {};
|
|
1734
|
-
const RECENT_ERRORS_CAP = 3;
|
|
1735
|
-
// Helper: push an error onto an originator's rolling window.
|
|
1736
|
-
function pushRecentError(originator, msg) {
|
|
1737
|
-
if (!recentErrors[originator])
|
|
1738
|
-
recentErrors[originator] = [];
|
|
1739
|
-
recentErrors[originator].push(msg);
|
|
1740
|
-
if (recentErrors[originator].length > RECENT_ERRORS_CAP)
|
|
1741
|
-
recentErrors[originator].shift();
|
|
1742
|
-
}
|
|
1743
|
-
// Seed schema-repair originator window from any schema-repair errors.
|
|
1744
|
-
for (const repair of schemaRepairs) {
|
|
1745
|
-
if (repair.outcome === "error") {
|
|
1746
|
-
const errMsg = repair.error ?? `schema repair error: ${repair.reason}`;
|
|
1747
|
-
pushRecentError("schema-repair", errMsg);
|
|
1748
|
-
}
|
|
1749
|
-
}
|
|
1750
|
-
// ── Phase 2: signal-delta eligibility sets built EARLY ────────────────────
|
|
1751
|
-
// 0.8.0 replaces the flat time-based cooldowns (which produced synchronised
|
|
1752
|
-
// waves whenever many refs cooled at the same instant — see the 2026-05-26
|
|
1753
|
-
// 54-ref simultaneous-reflect incident) with a *signal-delta* gate:
|
|
1754
|
-
//
|
|
1755
|
-
// reflectEligible(ref) ≡ latestFeedbackTs(ref) > lastReflectProposalTs(ref)
|
|
1756
|
-
// distillEligible(ref) ≡ latestFeedbackTs(ref) > lastDistillProposalTs(ref)
|
|
1757
|
-
//
|
|
1758
|
-
// i.e. a ref is re-eligible iff new feedback has landed since the last
|
|
1759
|
-
// proposal was generated for it. Stable content with no new signal stays
|
|
1760
|
-
// out of the queue regardless of clock time; a sudden burst of feedback
|
|
1761
|
-
// surfaces only the refs that the burst actually touches.
|
|
1762
|
-
//
|
|
1763
|
-
// The 30-day FEEDBACK_SIGNAL_WINDOW_DAYS bound still applies — only feedback
|
|
1764
|
-
// events newer than that count as "current signal". Ancient one-off
|
|
1765
|
-
// negatives don't permanently lock a ref into every run.
|
|
1766
|
-
//
|
|
1767
|
-
// High-retrieval refs (P0-A path) use a simpler "eligible once" rule: a
|
|
1768
|
-
// ref with no feedback signal but retrievalCount ≥ threshold is eligible
|
|
1769
|
-
// exactly once (no prior reflect proposal). Subsequent re-eligibility for
|
|
1770
|
-
// those refs requires either a new feedback event (then the normal
|
|
1771
|
-
// signal-delta gate applies) or human action. Documented limitation: this
|
|
1772
|
-
// path does not re-fire on retrieval-count growth alone in 0.8.0; storing
|
|
1773
|
-
// the retrieval count in proposal metadata for proper delta-tracking is
|
|
1774
|
-
// captured as future work.
|
|
1775
|
-
const FEEDBACK_SIGNAL_WINDOW_DAYS = 30;
|
|
1776
|
-
const feedbackSinceCutoff = new Date(Date.now() - daysToMs(FEEDBACK_SIGNAL_WINDOW_DAYS)).toISOString();
|
|
1777
|
-
// Build the three timestamp maps once across the entire postCleanupRefs set.
|
|
1778
|
-
// Per-ref queries would be N+1 and the planner is already the hottest path
|
|
1779
|
-
// in `akm improve`.
|
|
1780
|
-
const candidateRefs = postCleanupRefs.filter((r) => !validationFailureRefs.has(r.ref)).map((r) => r.ref);
|
|
1781
|
-
const latestFeedbackTs = buildLatestFeedbackTsMap(candidateRefs, feedbackSinceCutoff);
|
|
1782
|
-
const lastReflectProposalTs = buildLatestProposalTsMap(candidateRefs, "reflect");
|
|
1783
|
-
const lastDistillProposalTs = buildLatestProposalTsMap(candidateRefs, "distill");
|
|
1784
|
-
// Refs the distill signal-delta gate rejected at planning time. The main
|
|
1785
|
-
// loop reads this to skip distill for these refs without re-checking
|
|
1786
|
-
// eligibility per iteration.
|
|
1787
|
-
const distillCooledRefs = new Set();
|
|
1788
|
-
const preCooldownCount = postCleanupRefs.length;
|
|
1789
|
-
// ── Phase 3: partition postCleanupRefs by signal-delta eligibility ────────
|
|
1790
|
-
// Three buckets (validation failures are excluded entirely):
|
|
1791
|
-
// eligibleRefs — reflect signal-delta passes (full reflect+distill
|
|
1792
|
-
// loop path; distill guard remains in the loop for
|
|
1793
|
-
// refs that fail the distill signal-delta gate).
|
|
1794
|
-
// distillOnlyRefs — reflect blocked but distill signal-delta passes
|
|
1795
|
-
// AND ref is a distill candidate.
|
|
1796
|
-
// fullySkippedCount — neither gate passes → synthetic skip action
|
|
1797
|
-
// + improve_skipped event, excluded from sort.
|
|
1798
|
-
const eligibleRefs = [];
|
|
1799
|
-
const distillOnlyRefs = [];
|
|
1800
|
-
let fullySkippedCount = 0;
|
|
1801
|
-
// O-2 (#365): explicit --scope <ref> bypasses every gate (user intent wins).
|
|
1802
|
-
const scopeRefBypass = scope.mode === "ref";
|
|
1803
|
-
for (const r of postCleanupRefs) {
|
|
1804
|
-
if (validationFailureRefs.has(r.ref))
|
|
1805
|
-
continue;
|
|
1806
|
-
if (scopeRefBypass) {
|
|
1807
|
-
eligibleRefs.push(r);
|
|
1808
|
-
continue;
|
|
1809
|
-
}
|
|
1810
|
-
const reflectOk = isSignalDeltaEligible(r.ref, latestFeedbackTs, lastReflectProposalTs);
|
|
1811
|
-
const distillOk = isSignalDeltaEligible(r.ref, latestFeedbackTs, lastDistillProposalTs);
|
|
1812
|
-
const isDistillCandidate = isDistillCandidateRef(r.ref, options.stashDir);
|
|
1813
|
-
if (reflectOk) {
|
|
1814
|
-
if (!distillOk && isDistillCandidate) {
|
|
1815
|
-
// Reflect passes the gate, distill does not — emit the synthetic
|
|
1816
|
-
// distill-skipped action and event up-front so the in-loop guard
|
|
1817
|
-
// does not have to re-derive eligibility.
|
|
1818
|
-
distillCooledRefs.add(r.ref);
|
|
1819
|
-
actions.push({ ref: r.ref, mode: "distill-skipped", result: { ok: true, reason: "distill signal-delta" } });
|
|
1820
|
-
appendEvent({
|
|
1821
|
-
eventType: "improve_skipped",
|
|
1822
|
-
ref: r.ref,
|
|
1823
|
-
metadata: { reason: "distill_no_new_signal" },
|
|
1824
|
-
}, eventsCtx);
|
|
1825
|
-
}
|
|
1826
|
-
else if (!distillOk) {
|
|
1827
|
-
// Not a distill candidate AND distill gate doesn't pass — just mark
|
|
1828
|
-
// distillCooled so the loop's distill section is a no-op.
|
|
1829
|
-
distillCooledRefs.add(r.ref);
|
|
1830
|
-
}
|
|
1831
|
-
eligibleRefs.push(r);
|
|
1832
|
-
}
|
|
1833
|
-
else if (distillOk && isDistillCandidate) {
|
|
1834
|
-
// Reflect blocked but distill passes → distill-only bucket.
|
|
1835
|
-
distillOnlyRefs.push(r);
|
|
1836
|
-
}
|
|
1837
|
-
else {
|
|
1838
|
-
// Neither gate passes — fully skipped.
|
|
1839
|
-
fullySkippedCount++;
|
|
1840
|
-
actions.push({
|
|
1841
|
-
ref: r.ref,
|
|
1842
|
-
mode: "distill-skipped",
|
|
1843
|
-
result: { ok: true, reason: "no new signal since last proposal" },
|
|
1844
|
-
});
|
|
1845
|
-
appendEvent({ eventType: "improve_skipped", ref: r.ref, metadata: { reason: "no_new_signal" } }, eventsCtx);
|
|
1846
|
-
}
|
|
1847
|
-
}
|
|
1848
|
-
// ── Phase 4: signal/feedback/utility/sort on the reduced set ──────────────
|
|
1849
|
-
// Everything from here works only on (eligibleRefs ∪ distillOnlyRefs). The
|
|
1850
|
-
// fully-skipped bucket has already been routed and emitted; we deliberately
|
|
1851
|
-
// avoid spending DB/CPU on refs that cannot enter the loop.
|
|
1852
|
-
const processableRefs = [...eligibleRefs, ...distillOnlyRefs];
|
|
1853
|
-
// Gap 6: only surface feedback signals from the last 30 days so that
|
|
1854
|
-
// ancient one-off feedback events don't permanently lock an asset into
|
|
1855
|
-
// every improve run. Assets with only stale signals fall through to the
|
|
1856
|
-
// high-retrieval path (P0-A) or are skipped until new signals arrive.
|
|
1857
|
-
// (FEEDBACK_SIGNAL_WINDOW_DAYS / feedbackSinceCutoff are already defined in
|
|
1858
|
-
// Phase 2 above for the signal-delta gate; we reuse them here.)
|
|
1859
|
-
// Pre-compute feedback summary per ref in a single pass so we don't issue
|
|
1860
|
-
// two readEvents({type:"feedback", ref}) per asset (one for signal filtering,
|
|
1861
|
-
// one for ratio computation).
|
|
1862
|
-
const feedbackSummary = new Map();
|
|
1863
|
-
for (const candidate of processableRefs) {
|
|
1864
|
-
const { events } = readEvents({ type: "feedback", ref: candidate.ref });
|
|
1865
|
-
let hasSignal = false;
|
|
1866
|
-
let positive = 0;
|
|
1867
|
-
let negative = 0;
|
|
1868
|
-
for (const e of events) {
|
|
1869
|
-
if (!hasSignal &&
|
|
1870
|
-
(e.ts ?? "") >= feedbackSinceCutoff &&
|
|
1871
|
-
e.metadata !== undefined &&
|
|
1872
|
-
(typeof e.metadata.signal === "string" || typeof e.metadata.note === "string")) {
|
|
1873
|
-
hasSignal = true;
|
|
1874
|
-
}
|
|
1875
|
-
if (e.metadata?.signal === "positive")
|
|
1876
|
-
positive++;
|
|
1877
|
-
else if (e.metadata?.signal === "negative")
|
|
1878
|
-
negative++;
|
|
1879
|
-
}
|
|
1880
|
-
feedbackSummary.set(candidate.ref, { hasSignal, positive, negative });
|
|
1881
|
-
}
|
|
1882
|
-
const signalFiltered = processableRefs.filter((candidate) => feedbackSummary.get(candidate.ref)?.hasSignal === true);
|
|
1883
|
-
// P0-A: also surface zero-feedback assets that have been retrieved many times.
|
|
1884
|
-
const RETRIEVAL_COUNT_THRESHOLD = options.minRetrievalCount ?? 5;
|
|
1885
|
-
const signalBearingSet = new Set(signalFiltered.map((r) => r.ref));
|
|
1886
|
-
const noFeedbackCandidates = processableRefs.filter((r) => !signalBearingSet.has(r.ref));
|
|
1887
|
-
let highRetrievalRefs = [];
|
|
1888
|
-
let dbForRetrieval;
|
|
1889
|
-
try {
|
|
1890
|
-
dbForRetrieval = openExistingDatabase();
|
|
1891
|
-
const showEventCount = countUsageEventsByType(dbForRetrieval, "show");
|
|
1892
|
-
if (showEventCount === 0) {
|
|
1893
|
-
warn("Warning: show events not yet in usage_events — zero-feedback fallback will match only search-retrieved assets.");
|
|
1894
|
-
}
|
|
1895
|
-
const retrievalCounts = getRetrievalCounts(dbForRetrieval, noFeedbackCandidates.map((r) => r.ref));
|
|
1896
|
-
// High-retrieval signal-delta (simplified rule, 0.8.0): a no-feedback
|
|
1897
|
-
// ref qualifies exactly once — when retrievalCount ≥ threshold AND no
|
|
1898
|
-
// prior reflect proposal exists for it. Once a reflect proposal is on
|
|
1899
|
-
// record, subsequent re-eligibility requires explicit feedback (which
|
|
1900
|
-
// flows through the normal signal-delta gate above). Tracking growth in
|
|
1901
|
-
// retrieval count would require persisting the count in proposal
|
|
1902
|
-
// metadata; deferred to a follow-up.
|
|
1903
|
-
highRetrievalRefs = noFeedbackCandidates.filter((r) => (retrievalCounts.get(r.ref) ?? 0) >= RETRIEVAL_COUNT_THRESHOLD && !lastReflectProposalTs.has(r.ref));
|
|
1904
|
-
}
|
|
1905
|
-
catch (err) {
|
|
1906
|
-
rethrowIfTestIsolationError(err);
|
|
1907
|
-
// best-effort: if DB unavailable, highRetrievalRefs stays empty
|
|
1908
|
-
}
|
|
1909
|
-
finally {
|
|
1910
|
-
if (dbForRetrieval)
|
|
1911
|
-
closeDatabase(dbForRetrieval);
|
|
1912
|
-
}
|
|
1913
|
-
// If the user explicitly scoped to a single ref, always act on it —
|
|
1914
|
-
// skip the signal/retrieval filter entirely. The filter exists to avoid
|
|
1915
|
-
// noisy "improve everything" runs; it should not gate an intentional
|
|
1916
|
-
// per-ref invocation where the user's explicit choice is the signal.
|
|
1917
|
-
//
|
|
1918
|
-
// For type/all scope: only process refs with usage signals (recent feedback
|
|
1919
|
-
// or sufficient retrievals). A stash with no signals has 0 eligible refs —
|
|
1920
|
-
// usage is the gate. Run `akm feedback <ref> --positive` or retrieve assets
|
|
1921
|
-
// to bring them into the eligible pool.
|
|
1922
|
-
const signalAndRetrievalRefs = [...signalFiltered, ...highRetrievalRefs];
|
|
1923
|
-
const mergedRefs = scope.mode === "ref" ? processableRefs : options.requireFeedbackSignal ? signalFiltered : signalAndRetrievalRefs;
|
|
1924
|
-
const utilityMap = buildUtilityMap(mergedRefs);
|
|
1925
|
-
// Load feedback ratio per ref from the pre-computed summary (no extra DB pass).
|
|
1926
|
-
const feedbackRatios = new Map();
|
|
1927
|
-
for (const ref of mergedRefs) {
|
|
1928
|
-
const summary = feedbackSummary.get(ref.ref);
|
|
1929
|
-
const positive = summary?.positive ?? 0;
|
|
1930
|
-
const negative = summary?.negative ?? 0;
|
|
1931
|
-
const total = positive + negative;
|
|
1932
|
-
// ratio = negative proportion (high = needs more improvement)
|
|
1933
|
-
feedbackRatios.set(ref.ref, total > 0 ? negative / total : 0);
|
|
1934
|
-
}
|
|
1935
|
-
// Sort: combine utility (desc) with feedback negativity (desc) — high-negative assets rank higher
|
|
1936
|
-
const sorted = [...mergedRefs].sort((a, b) => {
|
|
1937
|
-
const utilA = utilityMap.get(a.ref) ?? 0;
|
|
1938
|
-
const utilB = utilityMap.get(b.ref) ?? 0;
|
|
1939
|
-
const ratioA = feedbackRatios.get(a.ref) ?? 0;
|
|
1940
|
-
const ratioB = feedbackRatios.get(b.ref) ?? 0;
|
|
1941
|
-
// Combined score: 70% utility, 30% negative ratio
|
|
1942
|
-
const scoreA = utilA * 0.7 + ratioA * 0.3;
|
|
1943
|
-
const scoreB = utilB * 0.7 + ratioB * 0.3;
|
|
1944
|
-
return scoreB - scoreA;
|
|
1945
|
-
});
|
|
1946
|
-
// Phase 0: surface coverage gaps from zero-result search queries
|
|
1947
|
-
let coverageGaps = [];
|
|
1948
|
-
try {
|
|
1949
|
-
const dbForGaps = openExistingDatabase();
|
|
1950
|
-
try {
|
|
1951
|
-
coverageGaps = getZeroResultSearches(dbForGaps);
|
|
1952
|
-
}
|
|
1953
|
-
finally {
|
|
1954
|
-
closeDatabase(dbForGaps);
|
|
1955
|
-
}
|
|
1956
|
-
}
|
|
1957
|
-
catch (err) {
|
|
1958
|
-
rethrowIfTestIsolationError(err);
|
|
1959
|
-
// best-effort
|
|
1960
|
-
}
|
|
1961
|
-
// actionableRefs is the post-cooldown, post-validation, post-signal, post-sort
|
|
1962
|
-
// set — i.e. the genuinely processable refs in priority order. Note: this is
|
|
1963
|
-
// a semantic shift from earlier code where actionableRefs was the pre-cooldown
|
|
1964
|
-
// sorted set; the new meaning matches reality and is documented on
|
|
1965
|
-
// ImprovePreparationResult.actionableRefs.
|
|
1966
|
-
//
|
|
1967
|
-
// Final guard: drop any candidate whose backing file is no longer on disk.
|
|
1968
|
-
// Phase 1 validation captures missing files at the start of preparation, but
|
|
1969
|
-
// the gap between that check and dispatch can be minutes on large stashes —
|
|
1970
|
-
// long enough for a checkpoint / git checkout / external cleanup to delete
|
|
1971
|
-
// the asset. Empirically (improve-critical-review 2026-05-20) the single
|
|
1972
|
-
// biggest reject category was "Asset no longer exists on disk" (604/1407 =
|
|
1973
|
-
// 43%), meaning reflect/distill was producing proposals against deleted refs.
|
|
1974
|
-
// A cheap existsSync per surviving candidate eliminates that wasted work.
|
|
1975
|
-
const assetMissingOnDisk = [];
|
|
1976
|
-
const existsCheckedActionable = [];
|
|
1977
|
-
for (const candidate of sorted) {
|
|
1978
|
-
// #591: prefer the path pre-resolved at planning time (synchronous
|
|
1979
|
-
// existsSync) over a serial async DB lookup per ref.
|
|
1980
|
-
const filePath = candidate.filePath && fs.existsSync(candidate.filePath)
|
|
1981
|
-
? candidate.filePath
|
|
1982
|
-
: await findAssetFilePath(candidate.ref, options.stashDir);
|
|
1983
|
-
if (filePath && fs.existsSync(filePath)) {
|
|
1984
|
-
existsCheckedActionable.push(candidate);
|
|
1985
|
-
}
|
|
1986
|
-
else {
|
|
1987
|
-
assetMissingOnDisk.push(candidate.ref);
|
|
1988
|
-
}
|
|
1989
|
-
}
|
|
1990
|
-
// #592 audit: one summary event instead of one per missing ref. Normally
|
|
1991
|
-
// tiny, but a stash deletion racing the run could make this O(n) sequential
|
|
1992
|
-
// state.db writes. `refs` is capped so the metadata row stays bounded.
|
|
1993
|
-
if (assetMissingOnDisk.length > 0) {
|
|
1994
|
-
appendEvent({
|
|
1995
|
-
eventType: "improve_skipped",
|
|
1996
|
-
ref: undefined,
|
|
1997
|
-
metadata: {
|
|
1998
|
-
reason: "asset_missing_on_disk",
|
|
1999
|
-
count: assetMissingOnDisk.length,
|
|
2000
|
-
refs: assetMissingOnDisk.slice(0, 50),
|
|
2001
|
-
},
|
|
2002
|
-
}, eventsCtx);
|
|
2003
|
-
}
|
|
2004
|
-
const actionableRefs = existsCheckedActionable;
|
|
2005
|
-
// Re-split actionableRefs (sorted) into reflect-path vs distill-only-path while
|
|
2006
|
-
// preserving sort order. distillOnlyRefs participate in the sort so --limit
|
|
2007
|
-
// picks them by score, not by arbitrary position.
|
|
2008
|
-
const distillOnlyRefSetForSort = new Set(distillOnlyRefs.map((r) => r.ref));
|
|
2009
|
-
const reflectAndDistillRefsAfterSort = [];
|
|
2010
|
-
const distillOnlyRefsAfterSort = [];
|
|
2011
|
-
for (const r of actionableRefs) {
|
|
2012
|
-
if (distillOnlyRefSetForSort.has(r.ref)) {
|
|
2013
|
-
distillOnlyRefsAfterSort.push(r);
|
|
2014
|
-
}
|
|
2015
|
-
else {
|
|
2016
|
-
reflectAndDistillRefsAfterSort.push(r);
|
|
2017
|
-
}
|
|
2018
|
-
}
|
|
2019
|
-
// ── Phase 5: --limit applies to the post-cooldown actionable set ──────────
|
|
2020
|
-
const allLoopRefs = [...reflectAndDistillRefsAfterSort, ...distillOnlyRefsAfterSort];
|
|
2021
|
-
const loopRefs = options.limit ? allLoopRefs.slice(0, options.limit) : allLoopRefs;
|
|
2022
|
-
// Update the returned distillOnlyRefs to the sorted order so callers see the
|
|
2023
|
-
// ranked view (loop stage uses it as a Set so order is irrelevant, but the
|
|
2024
|
-
// shape change keeps downstream consumers consistent).
|
|
2025
|
-
const distillOnlyRefsResult = distillOnlyRefsAfterSort;
|
|
2026
|
-
const totalReflectBlocked = fullySkippedCount + distillOnlyRefs.length;
|
|
2027
|
-
if (totalReflectBlocked > 0) {
|
|
2028
|
-
info(`[improve] ${totalReflectBlocked} of ${preCooldownCount} indexed refs blocked by reflect signal-delta ` +
|
|
2029
|
-
`(${fullySkippedCount} fully skipped, ${distillOnlyRefs.length} routed to distill-only)`);
|
|
2030
|
-
}
|
|
2031
|
-
if (signalAndRetrievalRefs.length > 0) {
|
|
2032
|
-
info(`[improve] ${signalAndRetrievalRefs.length} refs with usage signals (${signalFiltered.length} feedback, ${highRetrievalRefs.length} high-retrieval)`);
|
|
2033
|
-
}
|
|
2034
|
-
if (validationFailureRefs.size > 0) {
|
|
2035
|
-
info(`[improve] ${validationFailureRefs.size} with validation failures excluded`);
|
|
2036
|
-
}
|
|
2037
|
-
if (assetMissingOnDisk.length > 0) {
|
|
2038
|
-
info(`[improve] ${assetMissingOnDisk.length} candidates dropped — file not on disk`);
|
|
2039
|
-
}
|
|
2040
|
-
const deferredCount = actionableRefs.length - loopRefs.length;
|
|
2041
|
-
info(`[improve] ${actionableRefs.length} actionable; ${loopRefs.length} will be processed` +
|
|
2042
|
-
(options.limit && deferredCount > 0 ? ` (--limit ${options.limit} applied; ${deferredCount} deferred)` : ""));
|
|
2043
|
-
return {
|
|
2044
|
-
actions,
|
|
2045
|
-
cleanupWarnings,
|
|
2046
|
-
appliedCleanup,
|
|
2047
|
-
memoryIndexHealth,
|
|
2048
|
-
extract: extractResults,
|
|
2049
|
-
actionableRefs,
|
|
2050
|
-
signalBearingSet,
|
|
2051
|
-
validationFailures,
|
|
2052
|
-
schemaRepairs,
|
|
2053
|
-
lintSummary,
|
|
2054
|
-
loopRefs,
|
|
2055
|
-
distillCooledRefs,
|
|
2056
|
-
distillOnlyRefs: distillOnlyRefsResult,
|
|
2057
|
-
coverageGaps,
|
|
2058
|
-
recentErrors,
|
|
2059
|
-
utilityMap,
|
|
2060
|
-
gateAutoAcceptedCount,
|
|
2061
|
-
gateAutoAcceptFailedCount,
|
|
2062
|
-
consolidation: consolidationPass.consolidation,
|
|
2063
|
-
consolidationRan: consolidationPass.consolidationRan,
|
|
2064
|
-
};
|
|
2065
|
-
}
|
|
2066
|
-
async function runImproveLoopStage(args) {
|
|
2067
|
-
const { scope, options, primaryStashDir, reflectFn, distillFn, loopRefs, actions, signalBearingSet, distillCooledRefs, distillOnlyRefs, recentErrors, rejectedProposalsByRef, utilityMap, startMs, budgetMs, eventsCtx, improveProfile, } = args;
|
|
2068
|
-
// O-1 (#364): compute remaining budget at call time so each sub-call
|
|
2069
|
-
// receives only its fair share of the wall-clock budget.
|
|
2070
|
-
const remainingBudgetMs = () => Math.max(0, budgetMs - (Date.now() - startMs));
|
|
2071
|
-
const RECENT_ERRORS_CAP = 3;
|
|
2072
|
-
// requirePlannedRefs guard: when the distill profile sets this flag, skip
|
|
2073
|
-
// distill for distill-only refs if the reflect phase produced no planned refs.
|
|
2074
|
-
// Prevents the distill loop from generating hundreds of distill-skipped events
|
|
2075
|
-
// on quiet passes (all refs on reflect cooldown, no new signal to distill).
|
|
2076
|
-
const requirePlannedRefs = improveProfile?.processes?.distill?.requirePlannedRefs === true;
|
|
2077
|
-
const _distillOnlyRefNames = new Set(distillOnlyRefs.map((r) => r.ref));
|
|
2078
|
-
const hasReflectEligibleRefs = loopRefs.some((r) => !_distillOnlyRefNames.has(r.ref));
|
|
2079
|
-
const skipDistillDueToRequirePlannedRefs = requirePlannedRefs && !hasReflectEligibleRefs;
|
|
2080
|
-
// R-2 / #389: Self-Consistency multi-sample voting helpers.
|
|
2081
|
-
// Wang et al. arXiv:2203.11171 — N=3 samples beat single-shot on reasoning tasks.
|
|
2082
|
-
const SC_THRESHOLD = options.selfConsistencyThreshold ?? 0.7;
|
|
2083
|
-
const SC_N = Math.min(Math.max(2, options.selfConsistencyN ?? 3), 5);
|
|
2084
|
-
/**
|
|
2085
|
-
* Compute Jaccard token overlap between two strings.
|
|
2086
|
-
* Tokenizes by whitespace; returns 0 when both are empty.
|
|
2087
|
-
*/
|
|
2088
|
-
function jaccardSimilarity(a, b) {
|
|
2089
|
-
const tokensA = new Set(a.split(/\s+/).filter(Boolean));
|
|
2090
|
-
const tokensB = new Set(b.split(/\s+/).filter(Boolean));
|
|
2091
|
-
if (tokensA.size === 0 && tokensB.size === 0)
|
|
2092
|
-
return 1;
|
|
2093
|
-
let intersection = 0;
|
|
2094
|
-
for (const t of tokensA) {
|
|
2095
|
-
if (tokensB.has(t))
|
|
2096
|
-
intersection++;
|
|
2097
|
-
}
|
|
2098
|
-
const union = tokensA.size + tokensB.size - intersection;
|
|
2099
|
-
return union > 0 ? intersection / union : 0;
|
|
2100
|
-
}
|
|
2101
|
-
/**
|
|
2102
|
-
* Given N reflect results, return the one with the highest average Jaccard
|
|
2103
|
-
* similarity to all other successful results (majority-vote winner).
|
|
2104
|
-
* Falls back to the first successful result when N < 2.
|
|
2105
|
-
*/
|
|
2106
|
-
function pickMajorityVote(results) {
|
|
2107
|
-
const successful = results.filter((r) => r.ok);
|
|
2108
|
-
if (successful.length === 0)
|
|
2109
|
-
return (results[0] ?? {
|
|
2110
|
-
schemaVersion: 1,
|
|
2111
|
-
ok: false,
|
|
2112
|
-
reason: "non_zero_exit",
|
|
2113
|
-
error: "all samples failed",
|
|
2114
|
-
exitCode: null,
|
|
2115
|
-
});
|
|
2116
|
-
if (successful.length === 1)
|
|
2117
|
-
return successful[0];
|
|
2118
|
-
let bestIdx = 0;
|
|
2119
|
-
let bestScore = -1;
|
|
2120
|
-
for (let i = 0; i < successful.length; i++) {
|
|
2121
|
-
let totalSim = 0;
|
|
2122
|
-
for (let j = 0; j < successful.length; j++) {
|
|
2123
|
-
if (i === j)
|
|
2124
|
-
continue;
|
|
2125
|
-
totalSim += jaccardSimilarity(successful[i].proposal.payload.content ?? "", successful[j].proposal.payload.content ?? "");
|
|
2126
|
-
}
|
|
2127
|
-
const avgSim = totalSim / (successful.length - 1);
|
|
2128
|
-
if (avgSim > bestScore) {
|
|
2129
|
-
bestScore = avgSim;
|
|
2130
|
-
bestIdx = i;
|
|
2131
|
-
}
|
|
2132
|
-
}
|
|
2133
|
-
return successful[bestIdx] ?? successful[0];
|
|
2134
|
-
}
|
|
2135
|
-
// O-5 / #378: helper to push per-originator errors into the rolling window.
|
|
2136
|
-
function pushRecentError(originator, msg) {
|
|
2137
|
-
if (!recentErrors[originator])
|
|
2138
|
-
recentErrors[originator] = [];
|
|
2139
|
-
recentErrors[originator].push(msg);
|
|
2140
|
-
if (recentErrors[originator].length > RECENT_ERRORS_CAP)
|
|
2141
|
-
recentErrors[originator].shift();
|
|
2142
|
-
}
|
|
2143
|
-
// Build a Set for O(1) membership test — these refs skip the reflect call (Bug D2).
|
|
2144
|
-
const distillOnlyRefSet = new Set(distillOnlyRefs.map((r) => r.ref));
|
|
2145
|
-
let completedCount = 0;
|
|
2146
|
-
let reflectsWithErrorContext = 0;
|
|
2147
|
-
const memoryRefsForInference = new Set();
|
|
2148
|
-
// Pre-load all pending proposals once instead of querying per asset in the loop.
|
|
2149
|
-
const dedupeStashDirForProposals = primaryStashDir ?? options.stashDir;
|
|
2150
|
-
const pendingProposalRefSet = new Set(dedupeStashDirForProposals
|
|
2151
|
-
? listProposals(dedupeStashDirForProposals, { status: "pending" }).map((p) => p.ref)
|
|
2152
|
-
: []);
|
|
2153
|
-
let gateAutoAcceptedCount = 0;
|
|
2154
|
-
let gateAutoAcceptFailedCount = 0;
|
|
2155
|
-
const reflectGateCfg = makeGateConfig("reflect", {
|
|
2156
|
-
globalThreshold: options.autoAccept,
|
|
2157
|
-
dryRun: options.dryRun ?? false,
|
|
2158
|
-
stashDir: primaryStashDir,
|
|
2159
|
-
config: options.config ?? loadConfig(),
|
|
2160
|
-
eventsCtx,
|
|
2161
|
-
});
|
|
2162
|
-
const distillGateCfg = makeGateConfig("distill", {
|
|
2163
|
-
globalThreshold: options.autoAccept,
|
|
2164
|
-
dryRun: options.dryRun ?? false,
|
|
2165
|
-
stashDir: primaryStashDir,
|
|
2166
|
-
config: options.config ?? loadConfig(),
|
|
2167
|
-
eventsCtx,
|
|
2168
|
-
});
|
|
2169
|
-
for (const planned of loopRefs) {
|
|
2170
|
-
if (Date.now() - startMs >= budgetMs) {
|
|
2171
|
-
const remaining = loopRefs.length - completedCount;
|
|
2172
|
-
info(`[improve] budget exhausted after ${Math.round((Date.now() - startMs) / 60000)}min — ${remaining} assets skipped`);
|
|
2173
|
-
appendEvent({
|
|
2174
|
-
eventType: "improve_skipped",
|
|
2175
|
-
ref: planned.ref,
|
|
2176
|
-
metadata: {
|
|
2177
|
-
reason: "budget_exhausted",
|
|
2178
|
-
remaining,
|
|
2179
|
-
},
|
|
2180
|
-
}, eventsCtx);
|
|
2181
|
-
// B11: Emit improve_skipped for all remaining assets that will not be processed.
|
|
2182
|
-
for (const remainingRef of loopRefs.slice(completedCount + 1)) {
|
|
2183
|
-
appendEvent({
|
|
2184
|
-
eventType: "improve_skipped",
|
|
2185
|
-
ref: remainingRef.ref,
|
|
2186
|
-
metadata: { reason: "budget_exhausted_batch", remaining: loopRefs.length - completedCount - 1 },
|
|
2187
|
-
}, eventsCtx);
|
|
2188
|
-
}
|
|
2189
|
-
actions.push({
|
|
2190
|
-
ref: planned.ref,
|
|
2191
|
-
mode: "error",
|
|
2192
|
-
result: { ok: false, error: "timeout: improve wall-clock budget exhausted" },
|
|
2193
|
-
});
|
|
2194
|
-
break;
|
|
2195
|
-
}
|
|
2196
|
-
try {
|
|
2197
|
-
// Bug D2: distillOnlyRefs skip the reflect call but still run the distill path.
|
|
2198
|
-
// Bug D1: in-loop distill-cooldown check removed — distill-cooled candidates
|
|
2199
|
-
// have their synthetic actions emitted in runImprovePreparationStage.
|
|
2200
|
-
const isDistillOnly = distillOnlyRefSet.has(planned.ref);
|
|
2201
|
-
const parsedPlannedRef = parseAssetRef(planned.ref);
|
|
2202
|
-
// B6: derived memories are machine-generated; skip reflect to avoid noisy proposals.
|
|
2203
|
-
// shouldDistillMemoryRef already returns false for .derived refs, so the distill
|
|
2204
|
-
// path is also a no-op for them — we just avoid unnecessary agent spawns.
|
|
2205
|
-
// D2: distillOnlyRefs also skip the reflect call (reflect-cooled, distill path only).
|
|
2206
|
-
if (!isDistillOnly && !planned.ref.endsWith(".derived")) {
|
|
2207
|
-
// Type guard: skip reflect for unsupported types (script, env, task, etc.)
|
|
2208
|
-
// and raw wiki directories, driven by the active improve profile.
|
|
2209
|
-
const reflectSkip = shouldSkipRef(planned.ref, "reflect", improveProfile);
|
|
2210
|
-
if (reflectSkip.skip) {
|
|
2211
|
-
actions.push({
|
|
2212
|
-
ref: planned.ref,
|
|
2213
|
-
mode: "reflect-skipped",
|
|
2214
|
-
result: { ok: true, reason: reflectSkip.reason },
|
|
2215
|
-
});
|
|
2216
|
-
}
|
|
2217
|
-
else {
|
|
2218
|
-
// O-5 / #378: only inject reflect-originator errors into the reflect call.
|
|
2219
|
-
// Cross-task errors (e.g. schema-repair) must NOT contaminate reflect prompts.
|
|
2220
|
-
const reflectErrors = recentErrors.reflect ?? [];
|
|
2221
|
-
if (reflectErrors.length > 0)
|
|
2222
|
-
reflectsWithErrorContext++;
|
|
2223
|
-
// O-1 (#364): pass remaining budget as timeoutMs so the agent spawn is
|
|
2224
|
-
// bounded by the wall-clock deadline rather than the default per-profile timeout.
|
|
2225
|
-
const reflectBudgetMs = remainingBudgetMs();
|
|
2226
|
-
// Wire profile.processes.reflect.{mode, profile, timeoutMs} into the reflect
|
|
2227
|
-
// dispatch when present. Falls back to akmReflect's own config-based resolution
|
|
2228
|
-
// (profiles.improve.<name>.processes.reflect → defaults.llm) when the profile
|
|
2229
|
-
// does not specify.
|
|
2230
|
-
const reflectProfileRunner = resolveImproveProcessRunnerFromProfile(improveProfile.processes?.reflect, options.config ?? loadConfig());
|
|
2231
|
-
const reflectCallArgs = {
|
|
2232
|
-
ref: planned.ref,
|
|
2233
|
-
task: options.task,
|
|
2234
|
-
...(options.stashDir ? { stashDir: options.stashDir } : {}),
|
|
2235
|
-
...(reflectErrors.length > 0 ? { avoidPatterns: [...reflectErrors] } : {}),
|
|
2236
|
-
agentProcess: options.agentProcess ?? "reflect",
|
|
2237
|
-
eventSource: "improve",
|
|
2238
|
-
...(reflectBudgetMs > 0 ? { timeoutMs: reflectBudgetMs } : {}),
|
|
2239
|
-
...(reflectProfileRunner ? { runner: reflectProfileRunner } : {}),
|
|
2240
|
-
};
|
|
2241
|
-
// R-2 / #389: Self-consistency multi-sample voting for high-utility refs.
|
|
2242
|
-
// Self-Consistency arXiv:2203.11171 — N=3 samples beat single-shot quality.
|
|
2243
|
-
const refUtility = utilityMap.get(planned.ref) ?? 0;
|
|
2244
|
-
const useConsistency = refUtility >= SC_THRESHOLD && SC_N >= 2;
|
|
2245
|
-
let reflectResult;
|
|
2246
|
-
if (useConsistency) {
|
|
2247
|
-
const samples = [];
|
|
2248
|
-
for (let s = 0; s < SC_N; s++) {
|
|
2249
|
-
if (remainingBudgetMs() <= 0)
|
|
2250
|
-
break;
|
|
2251
|
-
// draftMode: skip DB write so each sample doesn't create a proposal.
|
|
2252
|
-
samples.push(await withLlmStage("reflect", () => reflectFn({ ...reflectCallArgs, draftMode: true })));
|
|
2253
|
-
}
|
|
2254
|
-
const winner = pickMajorityVote(samples.length > 0
|
|
2255
|
-
? samples
|
|
2256
|
-
: [await withLlmStage("reflect", () => reflectFn({ ...reflectCallArgs, draftMode: true }))]);
|
|
2257
|
-
// Persist only the majority-vote winner as a single real proposal.
|
|
2258
|
-
if (winner.ok && primaryStashDir) {
|
|
2259
|
-
const persistResult = createProposal(primaryStashDir, {
|
|
2260
|
-
ref: winner.proposal.ref,
|
|
2261
|
-
source: "reflect",
|
|
2262
|
-
sourceRun: `reflect-sc-${Date.now()}`,
|
|
2263
|
-
payload: winner.proposal.payload,
|
|
2264
|
-
});
|
|
2265
|
-
reflectResult = isProposalSkipped(persistResult)
|
|
2266
|
-
? {
|
|
2267
|
-
schemaVersion: 1,
|
|
2268
|
-
ok: false,
|
|
2269
|
-
reason: "cooldown",
|
|
2270
|
-
error: `SC proposal skipped: ${persistResult.message}`,
|
|
2271
|
-
ref: winner.ref,
|
|
2272
|
-
exitCode: null,
|
|
2273
|
-
}
|
|
2274
|
-
: { ...winner, proposal: persistResult };
|
|
2275
|
-
}
|
|
2276
|
-
else {
|
|
2277
|
-
reflectResult = winner;
|
|
2278
|
-
}
|
|
2279
|
-
}
|
|
2280
|
-
else {
|
|
2281
|
-
reflectResult = await withLlmStage("reflect", () => reflectFn(reflectCallArgs));
|
|
2282
|
-
}
|
|
2283
|
-
const isCooldown = !reflectResult.ok && reflectResult.reason === "cooldown";
|
|
2284
|
-
// Content-policy guard hits (reflect size-rail rejections) are NOT
|
|
2285
|
-
// LLM faults — the agent responded fine, the downstream guard
|
|
2286
|
-
// blocked the output. Route them to a distinct `reflect-guard-rejected`
|
|
2287
|
-
// mode so health metrics can split deterministic guard hits out of
|
|
2288
|
-
// true LLM failures. See
|
|
2289
|
-
// `/tmp/akm-health-investigations/metrics-taxonomy-review.md` §1a.
|
|
2290
|
-
const isGuardReject = !reflectResult.ok && reflectResult.reason === "content_policy_reject";
|
|
2291
|
-
// Type-guard rejection (reflect refused a script/env/task ref) is
|
|
2292
|
-
// also NOT an LLM failure — the LLM is never invoked. Route to the
|
|
2293
|
-
// existing `reflect-skipped` bucket so it does not inflate the
|
|
2294
|
-
// failure-rate numerator. ~9% of `reflect-failed` events in the
|
|
2295
|
-
// user's stack were this case; see review §1a row "Reflect refused
|
|
2296
|
-
// asset type".
|
|
2297
|
-
const isTypeRefused = !reflectResult.ok && reflectResult.reason === "unsupported_type";
|
|
2298
|
-
// Noise-gate suppression (#580): the candidate edit was an empty
|
|
2299
|
-
// diff or a cosmetic-only reformat of the current asset. Like
|
|
2300
|
-
// `unsupported_type`, this is a deterministic skip — not an LLM
|
|
2301
|
-
// fault — so it routes to the `reflect-skipped` bucket and stays
|
|
2302
|
-
// out of recentErrors/avoidPatterns.
|
|
2303
|
-
const isNoChange = !reflectResult.ok && reflectResult.reason === "no_change";
|
|
2304
|
-
actions.push({
|
|
2305
|
-
ref: planned.ref,
|
|
2306
|
-
mode: reflectResult.ok
|
|
2307
|
-
? "reflect"
|
|
2308
|
-
: isCooldown
|
|
2309
|
-
? "reflect-cooldown"
|
|
2310
|
-
: isGuardReject
|
|
2311
|
-
? "reflect-guard-rejected"
|
|
2312
|
-
: isTypeRefused || isNoChange
|
|
2313
|
-
? "reflect-skipped"
|
|
2314
|
-
: "reflect-failed",
|
|
2315
|
-
result: reflectResult,
|
|
2316
|
-
});
|
|
2317
|
-
// Cooldown skips, guard rejects, type-refused skips, and noise-gate
|
|
2318
|
-
// skips are not failures — do not pollute recentErrors with them
|
|
2319
|
-
// (those get injected as `avoidPatterns` into the next reflect
|
|
2320
|
-
// prompt). Guard rejects ARE worth showing the LLM as a learn-signal
|
|
2321
|
-
// so the next iteration sees "your last expansion was too large";
|
|
2322
|
-
// type-refused and no-change are deterministic and add no learning
|
|
2323
|
-
// signal.
|
|
2324
|
-
if (!reflectResult.ok && !isCooldown && !isTypeRefused && !isNoChange) {
|
|
2325
|
-
const errMsg = reflectResult.error ?? reflectResult.reason ?? "unknown reflect error";
|
|
2326
|
-
pushRecentError("reflect", errMsg);
|
|
2327
|
-
}
|
|
2328
|
-
// improve_reflect_outcome — per-asset metric for tuning the reflect path.
|
|
2329
|
-
appendEvent({
|
|
2330
|
-
eventType: "improve_reflect_outcome",
|
|
2331
|
-
ref: planned.ref,
|
|
2332
|
-
metadata: {
|
|
2333
|
-
ok: reflectResult.ok,
|
|
2334
|
-
durationMs: reflectResult.ok ? reflectResult.durationMs : undefined,
|
|
2335
|
-
agentProfile: reflectResult.ok ? reflectResult.agentProfile : undefined,
|
|
2336
|
-
reason: reflectResult.ok ? undefined : reflectResult.reason,
|
|
2337
|
-
},
|
|
2338
|
-
}, eventsCtx);
|
|
2339
|
-
if (reflectResult.ok) {
|
|
2340
|
-
const reflectGr = await runAutoAcceptGate([{ proposalId: reflectResult.proposal.id, confidence: reflectResult.proposal.confidence }], reflectGateCfg);
|
|
2341
|
-
gateAutoAcceptedCount += reflectGr.promoted.length;
|
|
2342
|
-
gateAutoAcceptFailedCount += reflectGr.failed.length;
|
|
2343
|
-
}
|
|
2344
|
-
} // end else (reflect type/profile check)
|
|
2345
|
-
}
|
|
2346
|
-
else if (!isDistillOnly && planned.ref.endsWith(".derived")) {
|
|
2347
|
-
// B6: .derived refs skip reflect; record synthetic skip action.
|
|
2348
|
-
actions.push({
|
|
2349
|
-
ref: planned.ref,
|
|
2350
|
-
mode: "distill-skipped",
|
|
2351
|
-
result: { ok: true, reason: "derived-memory-reflect-skipped" },
|
|
2352
|
-
});
|
|
2353
|
-
appendEvent({
|
|
2354
|
-
eventType: "improve_skipped",
|
|
2355
|
-
ref: planned.ref,
|
|
2356
|
-
metadata: { reason: "derived_memory_reflect_skipped" },
|
|
2357
|
-
}, eventsCtx);
|
|
2358
|
-
}
|
|
2359
|
-
// isDistillOnly refs: no reflect action emitted — proceed directly to distill path below.
|
|
2360
|
-
const hasRecentFeedbackSignal = signalBearingSet.has(planned.ref);
|
|
2361
|
-
const explicitRefScope = scope.mode === "ref";
|
|
2362
|
-
// Profile gate: apply the full type-filter / raw-wiki / disabled rules to
|
|
2363
|
-
// distill so callers who configure `profile.processes.distill.allowedTypes`
|
|
2364
|
-
// or land on raw-wiki refs get a recorded skip action instead of silently
|
|
2365
|
-
// proceeding.
|
|
2366
|
-
const distillSkip = shouldSkipRef(planned.ref, "distill", improveProfile);
|
|
2367
|
-
if (distillSkip.skip) {
|
|
2368
|
-
actions.push({
|
|
2369
|
-
ref: planned.ref,
|
|
2370
|
-
mode: "distill-skipped",
|
|
2371
|
-
result: { ok: true, reason: distillSkip.reason },
|
|
2372
|
-
});
|
|
2373
|
-
completedCount++;
|
|
2374
|
-
info(`[improve] ${completedCount}/${loopRefs.length} ${planned.ref}`);
|
|
2375
|
-
continue;
|
|
2376
|
-
}
|
|
2377
|
-
// requirePlannedRefs guard: skip distill for distill-only refs when no
|
|
2378
|
-
// reflect-eligible refs were planned this run, preventing mass skip events.
|
|
2379
|
-
if (skipDistillDueToRequirePlannedRefs && isDistillOnly) {
|
|
2380
|
-
actions.push({
|
|
2381
|
-
ref: planned.ref,
|
|
2382
|
-
mode: "distill-skipped",
|
|
2383
|
-
result: { ok: true, reason: "require_planned_refs" },
|
|
2384
|
-
});
|
|
2385
|
-
completedCount++;
|
|
2386
|
-
info(`[improve] ${completedCount}/${loopRefs.length} ${planned.ref}`);
|
|
2387
|
-
continue;
|
|
2388
|
-
}
|
|
2389
|
-
// See `isDistillCandidateRef` — excludes `lesson:*` (and anything else in
|
|
2390
|
-
// DISTILL_REFUSED_INPUT_TYPES) so distill never gets queued for an input
|
|
2391
|
-
// it will refuse.
|
|
2392
|
-
const shouldAttemptDistill = isDistillCandidateRef(planned.ref, options.stashDir);
|
|
2393
|
-
const skipMemoryDistillForWeakSignal = !isDistillOnly && parsedPlannedRef.type === "memory" && !hasRecentFeedbackSignal && !explicitRefScope;
|
|
2394
|
-
// distillCooledRefs guard: pre-filter emitted synthetic actions for distill-candidate
|
|
2395
|
-
// refs; non-candidate refs in the set are blocked here.
|
|
2396
|
-
// O-2 (#365): bypass the distill cooldown when the user explicitly targeted
|
|
2397
|
-
// this ref via --scope — their intent overrides unattended-run policies.
|
|
2398
|
-
if (shouldAttemptDistill &&
|
|
2399
|
-
!skipMemoryDistillForWeakSignal &&
|
|
2400
|
-
(!distillCooledRefs.has(planned.ref) || explicitRefScope)) {
|
|
2401
|
-
// TODO(refactor): single call site needs both lesson+knowledge refs for proposal dedup. If a third target ref type is added, extract deriveAllTargetRefs(inputRef): string[].
|
|
2402
|
-
const lessonRef = deriveLessonRef(planned.ref);
|
|
2403
|
-
const knowledgeRef = deriveKnowledgeRef(planned.ref);
|
|
2404
|
-
const dedupeStashDir = primaryStashDir ?? options.stashDir;
|
|
2405
|
-
if (dedupeStashDir) {
|
|
2406
|
-
// B2: check both lesson ref and knowledge ref since auto-promoted memories
|
|
2407
|
-
// create knowledge: proposals, not lesson: proposals.
|
|
2408
|
-
const hasExistingPending = pendingProposalRefSet.has(lessonRef) || pendingProposalRefSet.has(knowledgeRef);
|
|
2409
|
-
if (hasExistingPending) {
|
|
2410
|
-
actions.push({
|
|
2411
|
-
ref: planned.ref,
|
|
2412
|
-
mode: "distill-skipped",
|
|
2413
|
-
result: { ok: true, reason: "pending proposal exists" },
|
|
2414
|
-
});
|
|
2415
|
-
appendEvent({
|
|
2416
|
-
eventType: "improve_skipped",
|
|
2417
|
-
ref: planned.ref,
|
|
2418
|
-
metadata: { reason: "pending_proposal_exists" },
|
|
2419
|
-
}, eventsCtx);
|
|
2420
|
-
completedCount++;
|
|
2421
|
-
info(`[improve] ${completedCount}/${loopRefs.length} ${planned.ref}`);
|
|
2422
|
-
continue;
|
|
2423
|
-
}
|
|
2424
|
-
// D-2 (#370): reject-aware cooldown for distill. When the reviewer
|
|
2425
|
-
// recently rejected a distilled lesson or knowledge proposal for this
|
|
2426
|
-
// asset, skip re-distillation for a 1-day grace window. Prevents the
|
|
2427
|
-
// same rejected proposal from being regenerated immediately. The
|
|
2428
|
-
// window is fixed (the 0.8.0 redesign moved per-ref cooldowns to
|
|
2429
|
-
// signal-delta gates and dropped --distill-cooldown-days; a short
|
|
2430
|
-
// reject grace is preserved here so a fresh rejection isn't
|
|
2431
|
-
// overridden by the same run).
|
|
2432
|
-
// References: ExpeL arXiv:2308.10144, STaR arXiv:2203.14465.
|
|
2433
|
-
const DISTILL_REJECT_COOLDOWN_MS = daysToMs(1);
|
|
2434
|
-
const recentlyRejectedLesson = !explicitRefScope && // O-2: bypass when --scope <ref> is explicit
|
|
2435
|
-
(rejectedProposalsByRef.has(lessonRef) || rejectedProposalsByRef.has(knowledgeRef));
|
|
2436
|
-
if (recentlyRejectedLesson) {
|
|
2437
|
-
const rejectedEntry = rejectedProposalsByRef.get(lessonRef) ?? rejectedProposalsByRef.get(knowledgeRef);
|
|
2438
|
-
const rejectedAgeMs = rejectedEntry ? Date.now() - new Date(rejectedEntry.ts).getTime() : 0;
|
|
2439
|
-
if (rejectedAgeMs < DISTILL_REJECT_COOLDOWN_MS) {
|
|
2440
|
-
actions.push({
|
|
2441
|
-
ref: planned.ref,
|
|
2442
|
-
mode: "distill-skipped",
|
|
2443
|
-
result: { ok: true, reason: "distill reject grace window" },
|
|
2444
|
-
});
|
|
2445
|
-
appendEvent({
|
|
2446
|
-
eventType: "improve_skipped",
|
|
2447
|
-
ref: planned.ref,
|
|
2448
|
-
metadata: {
|
|
2449
|
-
reason: "distill_reject_grace_window",
|
|
2450
|
-
},
|
|
2451
|
-
}, eventsCtx);
|
|
2452
|
-
completedCount++;
|
|
2453
|
-
info(`[improve] ${completedCount}/${loopRefs.length} ${planned.ref}`);
|
|
2454
|
-
continue;
|
|
2455
|
-
}
|
|
2456
|
-
}
|
|
2457
|
-
}
|
|
2458
|
-
const distillResult = await withLlmStage("distill", () => distillFn({
|
|
2459
|
-
ref: planned.ref,
|
|
2460
|
-
...(parsedPlannedRef.type === "memory" ? { proposalKind: "auto" } : {}),
|
|
2461
|
-
...(options.stashDir ? { stashDir: options.stashDir } : {}),
|
|
2462
|
-
}));
|
|
2463
|
-
actions.push({ ref: planned.ref, mode: "distill", result: distillResult });
|
|
2464
|
-
if (distillResult.outcome === "queued" && distillResult.proposal) {
|
|
2465
|
-
const distillGr = await runAutoAcceptGate([{ proposalId: distillResult.proposal.id, confidence: distillResult.proposal.confidence }], distillGateCfg);
|
|
2466
|
-
gateAutoAcceptedCount += distillGr.promoted.length;
|
|
2467
|
-
gateAutoAcceptFailedCount += distillGr.failed.length;
|
|
2468
|
-
}
|
|
2469
|
-
if (parsedPlannedRef.type === "memory") {
|
|
2470
|
-
const promotedToKnowledge = distillResult.outcome === "queued" && distillResult.proposalKind === "knowledge";
|
|
2471
|
-
if (!promotedToKnowledge)
|
|
2472
|
-
memoryRefsForInference.add(planned.ref);
|
|
2473
|
-
}
|
|
2474
|
-
if (distillResult.outcome === "quality_rejected" && primaryStashDir) {
|
|
2475
|
-
const slug = planned.ref
|
|
2476
|
-
.replace(/[^a-z0-9]/gi, "-")
|
|
2477
|
-
.toLowerCase()
|
|
2478
|
-
.slice(0, 60);
|
|
2479
|
-
writeEvalCase(primaryStashDir, {
|
|
2480
|
-
ref: planned.ref,
|
|
2481
|
-
failureReason: distillResult.reason ?? "quality gate rejected",
|
|
2482
|
-
assetType: parseAssetRef(planned.ref).type ?? "unknown",
|
|
2483
|
-
rejectedAt: Date.now(),
|
|
2484
|
-
source: "distill_quality_rejected",
|
|
2485
|
-
slug: `${slug}-${Date.now()}`,
|
|
2486
|
-
});
|
|
2487
|
-
}
|
|
2488
|
-
// D6: use pre-loaded map instead of per-iteration DB query
|
|
2489
|
-
const rejectedProposalEvent = rejectedProposalsByRef.get(planned.ref);
|
|
2490
|
-
if (rejectedProposalEvent && primaryStashDir) {
|
|
2491
|
-
const slug = planned.ref
|
|
2492
|
-
.replace(/[^a-z0-9]/gi, "-")
|
|
2493
|
-
.toLowerCase()
|
|
2494
|
-
.slice(0, 60);
|
|
2495
|
-
writeEvalCase(primaryStashDir, {
|
|
2496
|
-
ref: planned.ref,
|
|
2497
|
-
failureReason: rejectedProposalEvent.metadata?.reason ?? "proposal rejected",
|
|
2498
|
-
assetType: parseAssetRef(planned.ref).type ?? "unknown",
|
|
2499
|
-
rejectedAt: new Date(rejectedProposalEvent.ts).getTime(),
|
|
2500
|
-
source: "proposal_rejected",
|
|
2501
|
-
slug: `${slug}-rejected`,
|
|
2502
|
-
});
|
|
2503
|
-
}
|
|
2504
|
-
}
|
|
2505
|
-
else if (skipMemoryDistillForWeakSignal) {
|
|
2506
|
-
actions.push({
|
|
2507
|
-
ref: planned.ref,
|
|
2508
|
-
mode: "distill-skipped",
|
|
2509
|
-
result: { ok: true, reason: "memory requires recent feedback signal" },
|
|
2510
|
-
});
|
|
2511
|
-
appendEvent({
|
|
2512
|
-
eventType: "improve_skipped",
|
|
2513
|
-
ref: planned.ref,
|
|
2514
|
-
metadata: { reason: "memory_distill_requires_feedback" },
|
|
2515
|
-
}, eventsCtx);
|
|
2516
|
-
}
|
|
2517
|
-
}
|
|
2518
|
-
catch (err) {
|
|
2519
|
-
// B7: UsageError thrown by akmDistill on validation_failed should be recorded
|
|
2520
|
-
// as mode:"distill" with outcome:"validation_failed", NOT as a generic error.
|
|
2521
|
-
// The distill_invoked event was already emitted inside akmDistill before the throw.
|
|
2522
|
-
if (err instanceof UsageError) {
|
|
2523
|
-
actions.push({
|
|
2524
|
-
ref: planned.ref,
|
|
2525
|
-
mode: "distill",
|
|
2526
|
-
result: { ok: false, outcome: "validation_failed", error: err.message },
|
|
2527
|
-
});
|
|
2528
|
-
}
|
|
2529
|
-
else {
|
|
2530
|
-
actions.push({
|
|
2531
|
-
ref: planned.ref,
|
|
2532
|
-
mode: "error",
|
|
2533
|
-
result: { ok: false, error: err instanceof Error ? err.message : String(err) },
|
|
2534
|
-
});
|
|
2535
|
-
}
|
|
2536
|
-
}
|
|
2537
|
-
completedCount++;
|
|
2538
|
-
info(`[improve] ${completedCount}/${loopRefs.length} ${planned.ref}`);
|
|
2539
|
-
}
|
|
2540
|
-
return { reflectsWithErrorContext, memoryRefsForInference, gateAutoAcceptedCount, gateAutoAcceptFailedCount };
|
|
2541
|
-
}
|
|
2542
|
-
async function runImprovePostLoopStage(args) {
|
|
2543
|
-
const { scope, options, primaryStashDir, actionableRefs, appliedCleanup, cleanupWarnings, memoryRefsForInference, reindexFn, eventsCtx, budgetSignal, improveProfile, consolidationRan, } = args;
|
|
2544
|
-
const allWarnings = [...cleanupWarnings, ...(appliedCleanup?.warnings ?? [])];
|
|
2545
|
-
info("[improve] post-loop maintenance starting");
|
|
2546
|
-
const maintenanceResult = await runImproveMaintenancePasses({
|
|
2547
|
-
options,
|
|
2548
|
-
primaryStashDir,
|
|
2549
|
-
actionableRefs,
|
|
2550
|
-
memoryRefsForInference,
|
|
2551
|
-
allWarnings,
|
|
2552
|
-
reindexFn,
|
|
2553
|
-
consolidationRan,
|
|
2554
|
-
// O-1 (#364): forward the budget signal to memory inference + graph extraction.
|
|
2555
|
-
budgetSignal,
|
|
2556
|
-
eventsCtx,
|
|
2557
|
-
improveProfile,
|
|
2558
|
-
});
|
|
2559
|
-
let deadUrls;
|
|
2560
|
-
if (scope.mode === "all" && primaryStashDir && actionableRefs.length > 0) {
|
|
2561
|
-
try {
|
|
2562
|
-
const knowledgeEntries = actionableRefs
|
|
2563
|
-
.filter((r) => {
|
|
2564
|
-
try {
|
|
2565
|
-
return parseAssetRef(r.ref).type === "knowledge";
|
|
2566
|
-
}
|
|
2567
|
-
catch {
|
|
2568
|
-
return false;
|
|
2569
|
-
}
|
|
2570
|
-
})
|
|
2571
|
-
.slice(0, 10)
|
|
2572
|
-
.map((r) => ({ ref: r.ref, body: "" }));
|
|
2573
|
-
if (knowledgeEntries.length > 0) {
|
|
2574
|
-
info(`[improve] checking URLs in ${knowledgeEntries.length} knowledge refs`);
|
|
2575
|
-
deadUrls = await checkDeadUrls(primaryStashDir, knowledgeEntries);
|
|
2576
|
-
info(`[improve] URL check complete (${deadUrls.length} dead/timeout URLs)`);
|
|
2577
|
-
}
|
|
2578
|
-
}
|
|
2579
|
-
catch {
|
|
2580
|
-
// best-effort
|
|
2581
|
-
}
|
|
2582
|
-
}
|
|
2583
|
-
return {
|
|
2584
|
-
allWarnings,
|
|
2585
|
-
deadUrls,
|
|
2586
|
-
...(maintenanceResult.memoryInference ? { memoryInference: maintenanceResult.memoryInference } : {}),
|
|
2587
|
-
...(maintenanceResult.graphExtraction ? { graphExtraction: maintenanceResult.graphExtraction } : {}),
|
|
2588
|
-
...(maintenanceResult.stalenessDetection ? { stalenessDetection: maintenanceResult.stalenessDetection } : {}),
|
|
2589
|
-
...(maintenanceResult.actions && maintenanceResult.actions.length > 0
|
|
2590
|
-
? { maintenanceActions: maintenanceResult.actions }
|
|
2591
|
-
: {}),
|
|
2592
|
-
memoryInferenceDurationMs: maintenanceResult.memoryInferenceDurationMs,
|
|
2593
|
-
graphExtractionDurationMs: maintenanceResult.graphExtractionDurationMs,
|
|
2594
|
-
orphansPurged: maintenanceResult.orphansPurged,
|
|
2595
|
-
proposalsExpired: maintenanceResult.proposalsExpired,
|
|
2596
|
-
// Consolidation's auto-accept gate counts now accrue in the preparation
|
|
2597
|
-
// stage (#551); post-loop no longer runs an auto-accept gate of its own.
|
|
2598
|
-
gateAutoAcceptedCount: 0,
|
|
2599
|
-
gateAutoAcceptFailedCount: 0,
|
|
2600
|
-
};
|
|
2601
|
-
}
|
|
2602
|
-
// TODO(refactor): mutates the passed-in `allWarnings` array as a hidden side channel. Return warnings in ImproveMaintenanceResult and merge in caller — invasive signature change deferred to next refactor pass.
|
|
2603
|
-
// Exported for tests (#584/#585 DB-locking regression coverage); production
|
|
2604
|
-
// callers reach it only through akmImprove → runImprovePostLoopStage.
|
|
2605
|
-
export async function runImproveMaintenancePasses(args) {
|
|
2606
|
-
const { options, primaryStashDir, memoryRefsForInference, allWarnings, reindexFn, consolidationRan, budgetSignal, eventsCtx, improveProfile, } = args;
|
|
2607
|
-
if (!primaryStashDir)
|
|
2608
|
-
return { memoryInferenceDurationMs: 0, graphExtractionDurationMs: 0 };
|
|
2609
|
-
const config = options.config ?? loadConfig();
|
|
2610
|
-
const sources = resolveSourceEntries(options.stashDir, config);
|
|
2611
|
-
const memoryInferenceFn = options.memoryInferenceFn ?? runMemoryInferencePass;
|
|
2612
|
-
const graphExtractionFn = options.graphExtractionFn ?? runGraphExtractionPass;
|
|
2613
|
-
const stalenessDetectionFn = options.stalenessDetectionFn ?? runStalenessDetectionPass;
|
|
2614
|
-
let db;
|
|
2615
|
-
let memoryInference;
|
|
2616
|
-
let graphExtraction;
|
|
2617
|
-
let stalenessDetection;
|
|
2618
|
-
let reindexedAfterInference = false;
|
|
2619
|
-
const actions = [];
|
|
2620
|
-
let memoryInferenceDurationMs = 0;
|
|
2621
|
-
let graphExtractionDurationMs = 0;
|
|
2622
|
-
let orphansPurged = 0;
|
|
2623
|
-
let proposalsExpired = 0;
|
|
2624
|
-
const openIndexDb = () => openDatabase(getDbPath(), config.embedding?.dimension ? { embeddingDim: config.embedding.dimension } : undefined);
|
|
2625
|
-
// #584: reindexFn opens its own write handle on the same index.db WAL file.
|
|
2626
|
-
// Holding our handle across that call produced SQLITE_BUSY / "database is
|
|
2627
|
-
// locked" failures in production, so the handle is closed BEFORE every
|
|
2628
|
-
// reindex and reopened after — the fresh handle also sees the post-reindex
|
|
2629
|
-
// state that graph extraction and staleness detection below rely on. The
|
|
2630
|
-
// reopen runs in `finally` so a failed reindex still leaves a usable handle.
|
|
2631
|
-
const reindexWithIndexDbReleased = async (stashDir) => {
|
|
2632
|
-
if (db) {
|
|
2633
|
-
closeDatabase(db);
|
|
2634
|
-
db = undefined;
|
|
2635
|
-
}
|
|
2636
|
-
try {
|
|
2637
|
-
await reindexFn({ stashDir });
|
|
2638
|
-
}
|
|
2639
|
-
finally {
|
|
2640
|
-
db = openIndexDb();
|
|
2641
|
-
}
|
|
2642
|
-
};
|
|
2643
|
-
try {
|
|
2644
|
-
db = openIndexDb();
|
|
2645
|
-
// Memory inference candidate-discovery (post-Item 9 fix from
|
|
2646
|
-
// memory:akm-improve-critical-review-2026-05-20). Previously this pass
|
|
2647
|
-
// was gated on memoryRefsForInference.size > 0 AND passed those refs as a
|
|
2648
|
-
// candidateRefs filter. But memoryRefsForInference is populated from refs
|
|
2649
|
-
// distilled THIS RUN — by the time that happens, those parents are
|
|
2650
|
-
// already split (`inferenceProcessed: true`) and `isPendingMemory` excludes
|
|
2651
|
-
// them. The genuinely-pending parents in the stash never entered the
|
|
2652
|
-
// filter. Result: 0/0/0 for 25 consecutive runs.
|
|
2653
|
-
//
|
|
2654
|
-
// Fix: always run the pass when the feature is enabled; let the pass's
|
|
2655
|
-
// own `collectPendingMemories` + `isPendingMemory` predicate find
|
|
2656
|
-
// candidates from the filesystem-of-truth. The this-run set is still
|
|
2657
|
-
// logged as a hint but no longer used as a filter.
|
|
2658
|
-
const memoryInferenceDisabledByProfile = improveProfile?.processes?.memoryInference?.enabled === false;
|
|
2659
|
-
const minPendingCount = improveProfile?.processes?.memoryInference?.minPendingCount;
|
|
2660
|
-
const pendingBelowMinCount = (() => {
|
|
2661
|
-
if (!primaryStashDir || minPendingCount === undefined || minPendingCount <= 0)
|
|
2662
|
-
return false;
|
|
2663
|
-
const pending = collectPendingMemories(primaryStashDir).length;
|
|
2664
|
-
if (pending < minPendingCount) {
|
|
2665
|
-
info(`[improve] memory inference skipped (${pending} pending < minPendingCount ${minPendingCount})`);
|
|
2666
|
-
return true;
|
|
2667
|
-
}
|
|
2668
|
-
return false;
|
|
2669
|
-
})();
|
|
2670
|
-
if (memoryInferenceDisabledByProfile) {
|
|
2671
|
-
info("[improve] memory inference skipped (disabled by improve profile)");
|
|
2672
|
-
}
|
|
2673
|
-
else if (pendingBelowMinCount) {
|
|
2674
|
-
// skipped — message already emitted above
|
|
2675
|
-
}
|
|
2676
|
-
else {
|
|
2677
|
-
const hintRefs = memoryRefsForInference.size;
|
|
2678
|
-
info(hintRefs > 0
|
|
2679
|
-
? `[improve] memory inference starting (${hintRefs} hint refs touched this run; pass discovers all pending)`
|
|
2680
|
-
: "[improve] memory inference starting (discovering pending parents)");
|
|
2681
|
-
const inferenceStart = Date.now();
|
|
2682
|
-
try {
|
|
2683
|
-
// O-1 (#364): pass budget signal so a hung inference call is cancelled.
|
|
2684
|
-
memoryInference = await withLlmStage("memory-inference", () => memoryInferenceFn({
|
|
2685
|
-
config,
|
|
2686
|
-
sources,
|
|
2687
|
-
signal: budgetSignal,
|
|
2688
|
-
db,
|
|
2689
|
-
reEnrich: false,
|
|
2690
|
-
onProgress: (event) => {
|
|
2691
|
-
const current = event.currentRef ? ` ${event.currentRef}` : "";
|
|
2692
|
-
info(`[improve] memory inference ${event.processed}/${event.total}${current} (written ${event.writtenFacts}, skipped ${event.skippedNoFacts})`);
|
|
2693
|
-
},
|
|
2694
|
-
}));
|
|
2695
|
-
memoryInferenceDurationMs = Date.now() - inferenceStart;
|
|
2696
|
-
actions.push({ ref: "memory:_inference", mode: "memory-inference", result: memoryInference });
|
|
2697
|
-
info(`[improve] memory inference complete (${memoryInference.writtenFacts} facts written from ${memoryInference.splitParents} parents)`);
|
|
2698
|
-
}
|
|
2699
|
-
catch (err) {
|
|
2700
|
-
memoryInferenceDurationMs = Date.now() - inferenceStart;
|
|
2701
|
-
allWarnings.push(`memory inference failed: ${err instanceof Error ? err.message : String(err)}`);
|
|
2702
|
-
}
|
|
2703
|
-
}
|
|
2704
|
-
if (memoryInference && (memoryInference.splitParents > 0 || memoryInference.writtenFacts > 0)) {
|
|
2705
|
-
info("[improve] reindexing after memory inference writes");
|
|
2706
|
-
try {
|
|
2707
|
-
await reindexWithIndexDbReleased(primaryStashDir);
|
|
2708
|
-
reindexedAfterInference = true;
|
|
2709
|
-
info("[improve] reindex after memory inference complete");
|
|
2710
|
-
}
|
|
2711
|
-
catch (err) {
|
|
2712
|
-
allWarnings.push(`reindex after memory inference failed: ${err instanceof Error ? err.message : String(err)}`);
|
|
2713
|
-
}
|
|
2714
|
-
}
|
|
2715
|
-
const graphEnabled = isProcessEnabled("index", "graph_extraction", config);
|
|
2716
|
-
const graphExtractionDisabledByProfile = improveProfile?.processes?.graphExtraction?.enabled === false;
|
|
2717
|
-
const graphExtractionFullScan = improveProfile?.processes?.graphExtraction?.fullScan === true;
|
|
2718
|
-
// Build the set of refs actually touched this run.
|
|
2719
|
-
const touchedRefs = new Set();
|
|
2720
|
-
for (const r of args.actionableRefs)
|
|
2721
|
-
touchedRefs.add(r.ref);
|
|
2722
|
-
for (const r of memoryRefsForInference)
|
|
2723
|
-
touchedRefs.add(r);
|
|
2724
|
-
// INVARIANT: graph extraction normally runs only on files touched by
|
|
2725
|
-
// actionable refs (candidatePaths). Full-corpus scans are opt-in via
|
|
2726
|
-
// profile.processes.graphExtraction.fullScan = true (used by the
|
|
2727
|
-
// `graph-refresh` built-in profile and its weekly scheduled task).
|
|
2728
|
-
// The empty-Set fallback is intentional when no refs were touched —
|
|
2729
|
-
// the extractor's filter rejects every file and returns empty, keeping
|
|
2730
|
-
// the pass invoked so the action is recorded and tests stay exercised.
|
|
2731
|
-
if (graphExtractionDisabledByProfile) {
|
|
2732
|
-
info("[improve] graph extraction skipped (disabled by improve profile)");
|
|
2733
|
-
}
|
|
2734
|
-
else if (sources.length > 0 && graphEnabled) {
|
|
2735
|
-
info(`[improve] graph extraction starting${graphExtractionFullScan ? " (full-corpus scan)" : ""}`);
|
|
2736
|
-
const extractionStart = Date.now();
|
|
2737
|
-
try {
|
|
2738
|
-
// D9: if consolidation ran but memory inference did not reindex, force a reindex
|
|
2739
|
-
// so graph extraction sees current DB state after consolidation writes.
|
|
2740
|
-
if (consolidationRan && !reindexedAfterInference) {
|
|
2741
|
-
info("[improve] reindexing after consolidation (graph extraction needs current state)");
|
|
2742
|
-
try {
|
|
2743
|
-
await reindexWithIndexDbReleased(primaryStashDir);
|
|
2744
|
-
reindexedAfterInference = true;
|
|
2745
|
-
info("[improve] reindex after consolidation complete");
|
|
2746
|
-
}
|
|
2747
|
-
catch (err) {
|
|
2748
|
-
allWarnings.push(`reindex after consolidation failed: ${err instanceof Error ? err.message : String(err)}`);
|
|
2749
|
-
}
|
|
2750
|
-
}
|
|
2751
|
-
// #584: no close/reopen needed here — reindexWithIndexDbReleased
|
|
2752
|
-
// already swapped in a fresh post-reindex handle.
|
|
2753
|
-
// Resolve touched refs to absolute file paths. Skipped for fullScan
|
|
2754
|
-
// (candidatePaths stays undefined → extractor processes all files).
|
|
2755
|
-
let candidatePaths;
|
|
2756
|
-
if (!graphExtractionFullScan) {
|
|
2757
|
-
candidatePaths = new Set();
|
|
2758
|
-
if (primaryStashDir && touchedRefs.size > 0) {
|
|
2759
|
-
const writableDirSet = new Set(getWritableStashDirs(primaryStashDir).map((d) => path.resolve(d)));
|
|
2760
|
-
const resolved = await Promise.all([...touchedRefs].map((ref) => findAssetFilePath(ref, primaryStashDir, writableDirSet).catch(() => null)));
|
|
2761
|
-
for (const p of resolved) {
|
|
2762
|
-
if (typeof p === "string" && p.length > 0)
|
|
2763
|
-
candidatePaths.add(p);
|
|
2764
|
-
}
|
|
2765
|
-
}
|
|
2766
|
-
}
|
|
2767
|
-
const progressHandler = (event) => {
|
|
2768
|
-
const current = event.currentPath ? ` ${path.basename(event.currentPath)}` : "";
|
|
2769
|
-
info(`[improve] graph extraction ${event.processed}/${event.total}${current} (extracted ${event.extracted}, entities ${event.totalEntities}, relations ${event.totalRelations})`);
|
|
2770
|
-
};
|
|
2771
|
-
// O-1 (#364): pass budget signal so a hung graph extraction call is cancelled.
|
|
2772
|
-
graphExtraction = await withLlmStage("graph-extraction", () => graphExtractionFn({
|
|
2773
|
-
config,
|
|
2774
|
-
sources,
|
|
2775
|
-
signal: budgetSignal,
|
|
2776
|
-
db,
|
|
2777
|
-
reEnrich: false,
|
|
2778
|
-
onProgress: progressHandler,
|
|
2779
|
-
options: { candidatePaths },
|
|
2780
|
-
}));
|
|
2781
|
-
graphExtractionDurationMs = Date.now() - extractionStart;
|
|
2782
|
-
actions.push({ ref: "graph:_artifact", mode: "graph-extraction", result: graphExtraction });
|
|
2783
|
-
info(`[improve] graph extraction complete (${graphExtraction.quality.extractedFiles} files, ${graphExtraction.quality.entityCount} entities, ${graphExtraction.quality.relationCount} relations)`);
|
|
2784
|
-
}
|
|
2785
|
-
catch (err) {
|
|
2786
|
-
graphExtractionDurationMs = Date.now() - extractionStart;
|
|
2787
|
-
allWarnings.push(`graph extraction failed: ${err instanceof Error ? err.message : String(err)}`);
|
|
2788
|
-
}
|
|
2789
|
-
}
|
|
2790
|
-
else if (sources.length > 0 && !graphEnabled) {
|
|
2791
|
-
info("[improve] graph extraction skipped (features.index.graph_extraction is disabled)");
|
|
2792
|
-
}
|
|
2793
|
-
// Orphan proposal purge — reject pending reflect proposals whose target
|
|
2794
|
-
// asset no longer exists on disk. Runs after graph extraction so newly
|
|
2795
|
-
// promoted assets from accept flows during this run are already present.
|
|
2796
|
-
if (primaryStashDir) {
|
|
2797
|
-
try {
|
|
2798
|
-
const purgeResult = purgeOrphanProposals(primaryStashDir, sources.map((s) => s.path));
|
|
2799
|
-
orphansPurged = purgeResult.rejected;
|
|
2800
|
-
if (purgeResult.rejected > 0) {
|
|
2801
|
-
info(`[improve] orphan purge: ${purgeResult.rejected}/${purgeResult.checked} orphaned proposals rejected (${purgeResult.durationMs}ms)`);
|
|
2802
|
-
}
|
|
2803
|
-
appendEvent({
|
|
2804
|
-
eventType: "proposal_orphan_purge",
|
|
2805
|
-
ref: "proposals:_orphan-purge",
|
|
2806
|
-
metadata: {
|
|
2807
|
-
checked: purgeResult.checked,
|
|
2808
|
-
rejected: purgeResult.rejected,
|
|
2809
|
-
durationMs: purgeResult.durationMs,
|
|
2810
|
-
byType: purgeResult.byType,
|
|
2811
|
-
orphans: purgeResult.orphans.map((o) => o.ref),
|
|
2812
|
-
},
|
|
2813
|
-
}, eventsCtx);
|
|
2814
|
-
}
|
|
2815
|
-
catch (err) {
|
|
2816
|
-
allWarnings.push(`orphan purge failed: ${err instanceof Error ? err.message : String(err)}`);
|
|
2817
|
-
}
|
|
2818
|
-
// Phase 6B (Advantage D6b): expire pending proposals that have aged past
|
|
2819
|
-
// the retention window. Runs AFTER orphan purge so we never double-archive
|
|
2820
|
-
// a proposal that orphan-purge already moved. `expireStaleProposals` emits
|
|
2821
|
-
// its own per-proposal `proposal_expired` events; we additionally emit a
|
|
2822
|
-
// single roll-up event here for parity with the orphan-purge surface.
|
|
2823
|
-
try {
|
|
2824
|
-
const expireResult = expireStaleProposals(primaryStashDir, config);
|
|
2825
|
-
proposalsExpired = expireResult.expired;
|
|
2826
|
-
if (expireResult.expired > 0) {
|
|
2827
|
-
info(`[improve] expiration: ${expireResult.expired}/${expireResult.checked} pending proposals expired ` +
|
|
2828
|
-
`(retention=${expireResult.retentionDays}d, ${expireResult.durationMs}ms)`);
|
|
2829
|
-
}
|
|
2830
|
-
appendEvent({
|
|
2831
|
-
eventType: "proposal_expiration_pass",
|
|
2832
|
-
ref: "proposals:_expiration",
|
|
2833
|
-
metadata: {
|
|
2834
|
-
checked: expireResult.checked,
|
|
2835
|
-
expired: expireResult.expired,
|
|
2836
|
-
durationMs: expireResult.durationMs,
|
|
2837
|
-
retentionDays: expireResult.retentionDays,
|
|
2838
|
-
expiredProposals: expireResult.expiredProposals,
|
|
2839
|
-
},
|
|
2840
|
-
}, eventsCtx);
|
|
2841
|
-
}
|
|
2842
|
-
catch (err) {
|
|
2843
|
-
allWarnings.push(`proposal expiration failed: ${err instanceof Error ? err.message : String(err)}`);
|
|
2844
|
-
}
|
|
2845
|
-
}
|
|
2846
|
-
// Fix #2 (observability 0.8.0): trim the events table in state.db so it
|
|
2847
|
-
// doesn't grow unbounded. `akm health` writes a `health_probe` row on every
|
|
2848
|
-
// invocation, and every command surface emits at least one event besides —
|
|
2849
|
-
// without this trim, state.db is a permanent append-only log. Config key
|
|
2850
|
-
// `improve.eventRetentionDays` (default 90, set 0 to disable) controls the
|
|
2851
|
-
// window. The purge runs against state.db (a different SQLite file from
|
|
2852
|
-
// the index `db` above).
|
|
2853
|
-
{
|
|
2854
|
-
const retentionDays = typeof config.improve?.eventRetentionDays === "number" ? config.improve.eventRetentionDays : 90;
|
|
2855
|
-
if (retentionDays > 0) {
|
|
2856
|
-
// #585: reuse the long-lived eventsCtx.db connection when akmImprove
|
|
2857
|
-
// opened one — opening a second state.db write connection while
|
|
2858
|
-
// eventsDb is still live made two simultaneous writers contend on the
|
|
2859
|
-
// same WAL file ("database is locked"). Only the eventsCtx.dbPath
|
|
2860
|
-
// fallback path (state.db failed to open up-front) opens — and then
|
|
2861
|
-
// owns and closes — its own handle. C2 still holds: the fallback uses
|
|
2862
|
-
// the boundary-pinned path, never a live `process.env` re-read.
|
|
2863
|
-
const ownsStateDb = !eventsCtx?.db;
|
|
2864
|
-
let stateDb;
|
|
2865
|
-
try {
|
|
2866
|
-
stateDb = eventsCtx?.db ?? openStateDatabase(eventsCtx?.dbPath);
|
|
2867
|
-
const purgedCount = purgeOldEvents(stateDb, retentionDays);
|
|
2868
|
-
if (purgedCount > 0) {
|
|
2869
|
-
info(`[improve] events purge: ${purgedCount} event(s) older than ${retentionDays}d removed from state.db`);
|
|
2870
|
-
}
|
|
2871
|
-
appendEvent({
|
|
2872
|
-
eventType: "events_purged",
|
|
2873
|
-
ref: "events:_purge",
|
|
2874
|
-
metadata: { purgedCount, retentionDays },
|
|
2875
|
-
}, eventsCtx);
|
|
2876
|
-
// improve_runs uses the same retention window as events — both are
|
|
2877
|
-
// observability/audit data, both grow append-only, both have a
|
|
2878
|
-
// dedicated purge helper. Mirroring the events purge here means a
|
|
2879
|
-
// single retention knob (improve.eventRetentionDays) governs both.
|
|
2880
|
-
const improveRunsPurged = purgeOldImproveRuns(stateDb, retentionDays);
|
|
2881
|
-
if (improveRunsPurged > 0) {
|
|
2882
|
-
info(`[improve] improve_runs purge: ${improveRunsPurged} run(s) older than ${retentionDays}d removed from state.db`);
|
|
2883
|
-
}
|
|
2884
|
-
appendEvent({
|
|
2885
|
-
eventType: "improve_runs_purged",
|
|
2886
|
-
ref: "improve_runs:_purge",
|
|
2887
|
-
metadata: { purgedCount: improveRunsPurged, retentionDays },
|
|
2888
|
-
}, eventsCtx);
|
|
2889
|
-
}
|
|
2890
|
-
catch (err) {
|
|
2891
|
-
allWarnings.push(`events purge failed: ${err instanceof Error ? err.message : String(err)}`);
|
|
2892
|
-
}
|
|
2893
|
-
finally {
|
|
2894
|
-
if (ownsStateDb && stateDb) {
|
|
2895
|
-
try {
|
|
2896
|
-
stateDb.close();
|
|
2897
|
-
}
|
|
2898
|
-
catch {
|
|
2899
|
-
// best-effort
|
|
2900
|
-
}
|
|
2901
|
-
}
|
|
2902
|
-
}
|
|
2903
|
-
// task_logs in logs.db (#579) shares the same retention window as
|
|
2904
|
-
// events/improve_runs — all three are observability data governed by
|
|
2905
|
-
// the single improve.eventRetentionDays knob. Separate try/finally
|
|
2906
|
-
// because logs.db is a different file: a locked/missing logs.db must
|
|
2907
|
-
// not block the state.db purges above.
|
|
2908
|
-
let logsDb;
|
|
2909
|
-
try {
|
|
2910
|
-
logsDb = openLogsDatabase();
|
|
2911
|
-
const taskLogsPurged = purgeOldTaskLogs(logsDb, retentionDays);
|
|
2912
|
-
if (taskLogsPurged > 0) {
|
|
2913
|
-
info(`[improve] task_logs purge: ${taskLogsPurged} log line(s) older than ${retentionDays}d removed from logs.db`);
|
|
2914
|
-
}
|
|
2915
|
-
appendEvent({
|
|
2916
|
-
eventType: "task_logs_purged",
|
|
2917
|
-
ref: "task_logs:_purge",
|
|
2918
|
-
metadata: { purgedCount: taskLogsPurged, retentionDays },
|
|
2919
|
-
}, eventsCtx);
|
|
2920
|
-
}
|
|
2921
|
-
catch (err) {
|
|
2922
|
-
allWarnings.push(`task_logs purge failed: ${err instanceof Error ? err.message : String(err)}`);
|
|
2923
|
-
}
|
|
2924
|
-
finally {
|
|
2925
|
-
if (logsDb) {
|
|
2926
|
-
try {
|
|
2927
|
-
logsDb.close();
|
|
2928
|
-
}
|
|
2929
|
-
catch {
|
|
2930
|
-
// best-effort
|
|
2931
|
-
}
|
|
2932
|
-
}
|
|
2933
|
-
}
|
|
2934
|
-
}
|
|
2935
|
-
}
|
|
2936
|
-
// Phase 4A (staleness detection). Activates the `deprecated` belief-state
|
|
2937
|
-
// machinery shipped in Phase 1A. Default OFF — gated by
|
|
2938
|
-
// `features.index.staleness_detection.enabled`. Runs after orphan purge
|
|
2939
|
-
// and before the URL check (which lives in the outer caller).
|
|
2940
|
-
if (sources.length > 0) {
|
|
2941
|
-
try {
|
|
2942
|
-
stalenessDetection = await withLlmStage("staleness-detection", () => stalenessDetectionFn({ config, sources, signal: budgetSignal, db }));
|
|
2943
|
-
if (stalenessDetection.considered > 0) {
|
|
2944
|
-
info(`[improve] staleness detection complete (considered ${stalenessDetection.considered}, ` +
|
|
2945
|
-
`deprecated ${stalenessDetection.deprecated}, confirmed ${stalenessDetection.confirmed}, ` +
|
|
2946
|
-
`skipped ${stalenessDetection.skipped}, ${stalenessDetection.durationMs}ms)`);
|
|
2947
|
-
}
|
|
2948
|
-
for (const w of stalenessDetection.warnings)
|
|
2949
|
-
allWarnings.push(`[improve] staleness detection: ${w}`);
|
|
2950
|
-
}
|
|
2951
|
-
catch (err) {
|
|
2952
|
-
allWarnings.push(`staleness detection failed: ${err instanceof Error ? err.message : String(err)}`);
|
|
2953
|
-
}
|
|
2954
|
-
}
|
|
2955
|
-
}
|
|
2956
|
-
finally {
|
|
2957
|
-
if (db)
|
|
2958
|
-
closeDatabase(db);
|
|
2959
|
-
}
|
|
2960
|
-
return {
|
|
2961
|
-
...(memoryInference ? { memoryInference } : {}),
|
|
2962
|
-
...(graphExtraction ? { graphExtraction } : {}),
|
|
2963
|
-
...(stalenessDetection ? { stalenessDetection } : {}),
|
|
2964
|
-
...(actions.length > 0 ? { actions } : {}),
|
|
2965
|
-
memoryInferenceDurationMs,
|
|
2966
|
-
graphExtractionDurationMs,
|
|
2967
|
-
orphansPurged,
|
|
2968
|
-
proposalsExpired,
|
|
2969
|
-
};
|
|
2970
|
-
}
|
|
2971
|
-
function shouldAnalyzeMemoryCleanup(scope, eligibleMemories, primaryStashDir) {
|
|
2972
|
-
if (!primaryStashDir || eligibleMemories === 0)
|
|
2973
|
-
return false;
|
|
2974
|
-
if (scope.mode === "all")
|
|
2975
|
-
return true;
|
|
2976
|
-
if (scope.mode === "type")
|
|
2977
|
-
return scope.value === "memory";
|
|
2978
|
-
if (!scope.value)
|
|
2979
|
-
return false;
|
|
2980
|
-
return parseAssetRef(scope.value).type === "memory";
|
|
2981
|
-
}
|
|
2982
1000
|
function shapeMemoryCleanup(plan) {
|
|
2983
1001
|
return {
|
|
2984
1002
|
analyzedDerived: plan.analyzedDerived,
|
|
@@ -2989,48 +1007,3 @@ function shapeMemoryCleanup(plan) {
|
|
|
2989
1007
|
...(plan.relativeDateCandidates.length > 0 ? { relativeDateCandidates: plan.relativeDateCandidates } : {}),
|
|
2990
1008
|
};
|
|
2991
1009
|
}
|
|
2992
|
-
function buildUtilityMap(refs) {
|
|
2993
|
-
const map = new Map();
|
|
2994
|
-
if (refs.length === 0)
|
|
2995
|
-
return map;
|
|
2996
|
-
const refSet = new Set(refs.map((r) => r.ref));
|
|
2997
|
-
let db;
|
|
2998
|
-
try {
|
|
2999
|
-
db = openExistingDatabase();
|
|
3000
|
-
const allDbEntries = getAllEntries(db);
|
|
3001
|
-
const idToRef = new Map();
|
|
3002
|
-
for (const indexed of allDbEntries) {
|
|
3003
|
-
const ref = makeAssetRef(indexed.entry.type, indexed.entry.name);
|
|
3004
|
-
if (refSet.has(ref))
|
|
3005
|
-
idToRef.set(indexed.id, ref);
|
|
3006
|
-
}
|
|
3007
|
-
const ids = [...idToRef.keys()];
|
|
3008
|
-
if (ids.length > 0) {
|
|
3009
|
-
const { global: scores } = getUtilityScoresByIds(db, ids);
|
|
3010
|
-
for (const [id, score] of scores) {
|
|
3011
|
-
const ref = idToRef.get(id);
|
|
3012
|
-
if (ref)
|
|
3013
|
-
map.set(ref, score.utility);
|
|
3014
|
-
}
|
|
3015
|
-
}
|
|
3016
|
-
}
|
|
3017
|
-
catch (err) {
|
|
3018
|
-
rethrowIfTestIsolationError(err);
|
|
3019
|
-
// best-effort: if DB unavailable, all utilities default to 0
|
|
3020
|
-
}
|
|
3021
|
-
finally {
|
|
3022
|
-
if (db)
|
|
3023
|
-
closeDatabase(db);
|
|
3024
|
-
}
|
|
3025
|
-
return map;
|
|
3026
|
-
}
|
|
3027
|
-
async function findAssetFilePath(ref, stashDir, writableDirSet) {
|
|
3028
|
-
return resolveAssetPath(ref, {
|
|
3029
|
-
stashDir,
|
|
3030
|
-
mode: "disk-only",
|
|
3031
|
-
writableDirSet,
|
|
3032
|
-
directoryIndexNames: ["SKILL.md"],
|
|
3033
|
-
preserveDirectNameFallback: true,
|
|
3034
|
-
honorOrigin: false,
|
|
3035
|
-
});
|
|
3036
|
-
}
|