local-kb-retrieve

---

name: local-kb-retrieve description: Retrieve relevant entries from the local predictive knowledge base as a lightweight preflight for repository work. Use route-first retrieval: infer the task direction, then search by domain path and cross-index before relying on flat keyword matching. In Codex, prefer a scout sidecar before non-trivial work and a recorder sidecar after work when sub-agents are available. Treat "no relevant entry" as useful signal, not a reason to skip the scan up front.

When this skill is used, follow this workflow:

Rule authority:

• Treat PROJECT_SPEC.md as the canonical source for thresholds, maintenance boundaries, and governance rules.
• Treat docs/maintenance_runbook.md as the canonical operational runbook for sleep maintenance.
• Keep this skill focused on the workflow shape. If this skill and the spec disagree, follow the spec.

Default stance: run a quick scan first for repository tasks; keep the scan lightweight unless the returned entries are clearly relevant. This includes GitHub publish work, release audits, README presentation passes, public template cleanup, and repo-boundary/privacy review. Those still count as repository tasks and should not bypass the scan. Keep the active path mathematically simple and easy to audit. During normal task flow, retrieval should use explicit additive rules and should not try to normalize or repair the KB inline.

Preferred Codex operating pattern:

• For non-trivial work, start a read-oriented kb-scout sidecar sub-agent before the main task so the primary agent can stay focused on the critical path.
• For long mixed tasks, add phase-change KB checkpoints to the plan. Re-run retrieval before a substantially different action class or risk surface begins, such as moving from analysis to code edits, from edits to packaging, from local validation to GitHub push/tag/release, or into privacy-sensitive, organization-KB, automation, or public-publication work. Do not rerun retrieval for every repeated same-type subtask.
• After the main task, treat KB postflight as part of done. Start a kb-recorder sidecar sub-agent whenever the task exposed feedback worth keeping, comments, misses, route gaps, card weaknesses, or candidate lessons.
• Run deeper consolidation in a separate scheduled maintenance conversation or automation rather than inside the main task thread.
• If sub-agents are unavailable, or if the task is trivial, fall back to a lightweight inline scan and inline feedback note.

Rule discipline:

• Keep retrieval scoring as a simple additive rule: structural route evidence + lexical evidence + small confidence/status adjustment.
• Do not let confidence or status create a hit by themselves. They may rerank plausible matches, but they should not turn unrelated entries into matches.
• Keep runtime survival separate from KB repair. A malformed lexical candidate may be excluded, but missing or inconsistent exact LogicGuard authority is a visible failure. Never continue through a readable-card, legacy-YAML, or floating-head fallback; leave repair and republication to Sleep or the versioned upgrade owner.
• Keep parameters few and fixed. Prefer counts, weights, and thresholds that a human can inspect over adaptive or opaque heuristics.
• Keep canonical machine interfaces separate from localized display projection. CLI tools, installed launchers, automation payloads, and installer checks should use canonical machine output and encoding-stable JSON; Chinese display belongs in i18n.zh-CN, route display labels, and UI view models.

Sleep maintenance lanes:

• deterministic repair: schema normalization, field-type cleanup, canonical path alignment, low-ambiguity formatting, and other fixes that can be expressed as fixed rules and validated mechanically.
• semantic repair: LogicGuard model revisions, ModelMesh membership or grounded relation changes, candidate creation, merges, promotions, deprecations, taxonomy changes, and other changes that depend on accumulated evidence rather than a single malformed field.
• Keep semantic auto-apply thresholds simple. Do not restate every threshold here; use the canonical thresholds from PROJECT_SPEC.md and docs/maintenance_runbook.md.
• If a change does not satisfy a deterministic rule or a simple semantic threshold, leave it as a proposal for a later maintenance pass.

Independent maintenance thread:

• Use a separate maintenance chat or automation for the library's "sleep" workflow. Do not let deep consolidation interrupt the user's main task thread.
• In the current implementation, Sleep may safely publish exact model generations, deterministic card projections, history events, consolidation artifacts, rollback manifests, candidate models, and explicit decision traces for ignored observations, rejected candidates, model gaps, and confidence reviews.
• Do not treat the current tooling as permission to rewrite trusted models or restructure taxonomy silently inside a main task. Durable model and mesh changes belong to Sleep; Dream and ordinary retrieval remain read-only.
• A practical active-build cadence is once per day. For calmer periods, two or three times per week is usually enough.

1. Summarize the task in one short sentence.
2. Infer one primary conceptual route such as work/reporting/ppt or engineering/debugging/version-change.
3. Infer up to three alternative routes when the task may be reachable through more than one conceptual direction.
4. If sub-agents are available and the task is non-trivial, let kb-scout handle the initial scan. Otherwise run: python .agents/skills/local-kb-retrieve/scripts/kb_search.py --route-hint "<primary route>" --query "<task summary plus useful keywords>" --top-k 5
5. If this is a long mixed plan, mark phase-change KB checkpoints and run a fresh route-specific retrieval before each new action class or risk surface. Skip repeated checks for same-type work where the first retrieval still applies.
6. Read the returned entries and their exact logicguard_context: pinned authority generation, model and mesh revisions, root node or ArgumentBlock, typed nodes and edges, explicit gaps, and grounded mesh neighborhood.
7. Prefer entries with stronger route alignment, status: trusted, and higher confidence.
8. Use retrieved model nodes as bounded context. Do not overgeneralize beyond the root claim, declared assumptions, limitations, rebuttals, or exact grounded mesh neighborhood. A lexical hit selects an entry point; it does not license traversal through legacy related_cards.
9. If no relevant entries are found, continue without forcing the library into the answer; the absence of hits is still a useful signal.
10. Before finalizing any non-trivial task, run one explicit KB postflight check. First ask whether Codex exposed mistake-first evidence: its own mistake, weak path, missed instruction, failed validation, tool or skill misuse, user correction, or later correction episode. Treat that as the highest-priority observation evidence. Then ask whether this task exposed a reusable lesson, a skill/plugin or subagent/delegation usage lesson, a miss, a route gap, a card weakness, or a KB-process failure.
11. If the answer is yes, let kb-recorder append structured feedback so the scheduled AI consolidation flow can process it later. When more than one card materially influenced the task, record all of those entry ids rather than only the top hit.
12. If the answer is no, make that an explicit conclusion rather than silently forgetting to check.
13. If a reusable new lesson emerges during the task, record it into candidates or structured history rather than trying to fully consolidate it inside the active task thread.
14. When a reusable lesson is being recorded, do not stop at a generic summary. Preserve predictive-model evidence: the scenario, the action or condition, the observed result, and how future Codex behavior should use that result.
15. When the task involved a weaker path, mistake, or later correction, preserve both sides of the contrast whenever possible: what the earlier action was, what weaker result it led to, what changed, and what improved afterward. These contrastive observations are often the easiest ones to turn into future card alternatives. Successful reusable patterns may still be recorded; mistake-first priority is not a rule to suppress success observations.
16. Preserve thread_ref, project_ref, and workspace_root whenever they are known. Sleep maintenance should be able to reconstruct same-project chronology and correction episodes rather than reading every observation as an isolated note.
17. Lessons about current model or runtime behavior are valid when they stay bounded and auditable. Preserve the most precise runtime identity that is actually known, and if exact model identity is not surfaced reliably, scope the lesson more conservatively to the active Codex runtime, current environment, or known model family.
18. When such a lesson is likely to become a card later, keep more than one retrieval path in view: a runtime-facing route such as codex/runtime-behavior/... plus any prompting, tool-use, workflow, or planning routes that materially shaped the behavior.
19. Lessons about a specific user are also valid when they stay bounded, evidence-based, and behaviorally framed. Record them as task-conditioned private predictions about likely preference, correction, or judgment rather than as personality labels or broad character impressions.
20. Lessons about Codex Skill or plugin use are valid personal-KB evidence when the Skill is new, repeatedly useful, task-critical, missing, misleading, selected after an earlier route failed, combined with another Skill, or when the task shows that a Skill should be invoked earlier or avoided under known conditions. Do not record generic praise or every routine invocation; record scenario, Skill action/choice, observed result, and future operational use. Preserve a skill-facing route such as codex/workflow/skills or codex/skill-use/<skill-name> plus the task-facing route that made the Skill relevant.
21. Lessons about Codex subagent/delegation use are valid personal-KB evidence when spawning, avoiding, sequencing, waiting for, or parallelizing subagents materially changes speed, coordination overhead, context isolation, main-thread clarity, verification quality, or task outcome. Do not record generic praise for parallelism or every routine sidecar; record scenario, delegation action/choice, observed result, and future operational use. Preserve a workflow-facing route such as codex/workflow/subagents plus the task-facing route that made delegation relevant.

Sleep maintenance checklist:

1. Inspect the explicit taxonomy layer: python .agents/skills/local-kb-retrieve/scripts/kb_taxonomy.py --json
2. Inspect the smallest undeclared taxonomy routes that are currently implied by observed entries: python .agents/skills/local-kb-retrieve/scripts/kb_taxonomy.py --gaps-only --json
3. Inspect recent route navigation if the taxonomy view exposes undeclared branches you need to understand: python .agents/skills/local-kb-retrieve/scripts/kb_nav.py --json
4. Run consolidation in report-only mode first. Keep deterministic repair and semantic repair conceptually separate: python .agents/skills/local-kb-retrieve/scripts/kb_consolidate.py --json --apply-mode none
5. If the grouped history looks clean, allow only the lowest-risk automatic apply path: python .agents/skills/local-kb-retrieve/scripts/kb_consolidate.py --json --apply-mode new-candidates Treat broad routes as proposal-only even when they repeat; in the current implementation, current low-risk auto-apply should stay on routes with at least 3 segments. Only create candidate models from observations with concrete future utility: a bounded predictive claim plus actionable operational_use. Low-confidence seeds are allowed; missing evidence/warrant/opposition/boundary roles must remain explicit model gaps, while low-utility observations should stay history-only or be ignored through an explicit maintenance decision. For relationship discovery, python .agents/skills/local-kb-retrieve/scripts/kb_consolidate.py --json --apply-mode related-cards may batch unresolved grounding proposals for Sleep. It never updates related_cards and never creates a canonical ModelMesh edge by itself. For stable alternate-route maintenance, python .agents/skills/local-kb-retrieve/scripts/kb_consolidate.py --json --apply-mode cross-index may update low-risk cross_index routes when repeated route evidence is already strong enough.
6. Restrict automatic apply to deterministic rules or simple thresholded semantic cases. Keep higher-ambiguity semantic changes proposal-only until the support and contradiction thresholds are met.
7. Inspect the per-action proposal stubs for the run: python .agents/skills/local-kb-retrieve/scripts/kb_proposals.py --run-id <run_id> --json
8. If AI decides a weak observation should stay history-only, a candidate should be rejected, a confidence review should be logged, or a split review should be closed without rewriting the card yet, record that decision explicitly: python .agents/skills/local-kb-retrieve/scripts/kb_maintenance.py --decision-type observation-ignored|candidate-rejected|confidence-reviewed|split-reviewed --action-key <action_key> --resolved-event-ids <csv_event_ids> --reason "<why>" --json For split-reviewed, always include the concrete supporting event ids from the review you are closing.
9. If the same trusted card keeps appearing in maintenance output, run a split review before assuming it should be split:
   • a hub card still expresses one bounded predictive relation and may stay as the route entry point
   • an overloaded card now mixes multiple scenarios, actions, results, or route-specific subcases and should move toward a split proposal, even when those branches still arrive through the same route
10. Review the emitted snapshot.json, proposal.json, action stub paths, and apply.json paths from the consolidation result.
11. If a consolidation run needs recovery, inspect and optionally restore history events: python .agents/skills/local-kb-retrieve/scripts/kb_rollback.py inspect --run-id <run_id> --write-manifest --json python .agents/skills/local-kb-retrieve/scripts/kb_rollback.py restore --run-id <run_id> --artifact history-events --json
12. End the maintenance pass with a concise summary: run id, created candidates, skipped actions, undeclared taxonomy signals, and the next deeper changes AI should make later.

Output discipline:

• Briefly state which entry ids and exact model revisions influenced the answer.
• Treat those identities as the model nodes that materially influenced the work, not merely every readable projection that happened to be retrieved.
• If the entries are weak or ambiguous, say so.
• Do not expose private entry content unless the user is authorized to see it.
• User-specific lessons should default to private handling and should describe what this user is likely to prefer or reject in a concrete task context, not who the user “is” in general.
• Keep sidecar agents scoped: kb-scout should be read-mostly and kb-recorder should default to history, comments, and candidate writes rather than broad structural edits.
• Record subagent/delegation usage lessons when sidecars materially helped or hurt speed, main-thread focus, context isolation, verification, or integration.
• In a maintenance thread, be explicit about what the tooling actually changed versus what still remains a proposal.
• For non-trivial work, treat the explicit postflight observation check as part of done rather than optional housekeeping.
• After meaningful tasks, prefer recording one structured observation even when the outcome was a workflow or rule clarification. Sleep maintenance depends on accumulated evidence, not only on bug-fix episodes.
• During Sleep maintenance, repeated hits are a split-review signal, not an automatic split rule. Keep intact hub models that still express one bounded predictive relation, and only split overloaded models that now carry multiple predictive relations.
• Runtime-behavior lessons should be written as predictions about this runtime under concrete conditions, not as vague folklore about “LLMs in general.”
• Cross-index maintenance should only strengthen direct alternate retrieval paths from repeated actual route evidence in low-risk auto-apply. Pruning or broader route cleanup should remain proposal-first until stronger removal evidence exists.
• Search has one current CLI grammar: use the explicit search entrypoint with --route-hint. Removed argument aliases are rejected and are not runtime migration paths.