Internal post-work quality and learning phase. Runs kb-review -> resolution gate -> follow-up resolution -> proof/demo evidence -> compound -> learn -> evolve -> memory refresh/compact -> cleanup after kb-work finishes all slices. Normally invoked by kb-work or kb-complete; not the primary user-facing completion command.
Instalação
Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
After kb-work finishes executing all slices, this skill runs the quality review, follow-up resolution, proof/demo evidence, knowledge capture, memory-health, and cleanup steps. Separated from kb-work so the user can choose when to run it — klfg prompts automatically, standalone users invoke it manually.
Terminal Completion Contract
When invoked by kb-work, this skill is the terminal half of the execution
loop. Continue until Step 5 reports Done, or until a real blocker is recorded
with exact resume criteria.
Do not stop at weaker milestones:
deterministic checks passed before review;
reviewers returned findings;
P0/P1 findings were fixed but proof has not rerun;
proof passed but memory, learn/evolve cadence, or cleanup is unfinished;
a summary was written without the final KB <name> complete report.
If a repo has a project-specific done.md contract such as "can't stop til its
done", interpret "done" as this skill's Step 5 terminal report. KB state still
lives in todo.md, todo-done.md, manifests, and handoffs unless the repo
already opted into a done.md workflow.
Input
#$ARGUMENTS
If input is empty: Scan docs/plans/ for the most recent *-kb-*-manifest.md file with status: completed. If none is found but an active KB manifest exists with pending or in-progress slices, stop and invoke kb-work <manifest-path> first. If no manifest exists, do not run completion from a raw diff; route back to kb-plan/kb-work so completion has slice scope and verification evidence.
If input is a path: Read the manifest at that path.
Pre-Flight
Read the manifest — confirm status: completed (all slices done/skipped). If slices are still pending or in_progress, stop: "This manifest has unfinished slices. Run kb-work first."
Validate gate ledger — the manifest must contain gate_ledger with work-to-complete set to passed, and every completed slice must have a passing slice-<id>-to-done gate. New manifests use allowed_next_action: kb-finalize <manifest-path>; accept legacy kb-complete <manifest-path> while old plans remain active. Run the checker against the recorded allowed action. If missing, blocked, or the checker fails, stop and invoke kb-work <manifest-path> or kb-gate to repair the ledger. Do not run finalization from a manifest that merely says status: completed.
Validate objective contract — if the manifest has objective_contract: true, run go run ./cmd/kbcheck manifest-contract --manifest <manifest-path>. Confirm the top-level done_check is present and each completed slice has proof_check evidence or an accepted no_check_reason recorded in slice notes/gates. If missing, stop and return to kb-work for proof repair.
Collect scope context — scan each slice's notes field for scope-check: and scope-discovery: entries. Build the combined list of actually changed, scope-verified files across all slices. This becomes the review scope.
Collect memory impact — scan slice notes for memory-impact: and kb-map-refresh: entries.
Review routing evidence honestly — collect any routing receipts or host evidence linked from the run state, but treat them as telemetry rather than proof of correctness. Missing, unknown, or mismatched provenance does not invalidate already-proven work.
Identify the branch baseline — run git merge-base HEAD main to establish the diff range.
Run final snapshot sweep — invoke kb-regression-snapshot verify for all snapshots under .kb/snapshots/. If any snapshot fails, STOP before review; later work regressed earlier passing behavior.
If the manifest has no scope-check notes (older format), fall back to git diff --name-only $(git merge-base HEAD main)..HEAD for the file list.
Step 1: Code Review
Before code review, run kb-check against the completed manifest scope. If deterministic checks fail, route to kb-repair or kb-fix before kb-review. LLM review does not replace executable verification.
Do not choose, rerun, or require a different model solely to improve routing telemetry. kb-finalize reviews receipts and notes useful observations for future selection, but proof and review gates remain model-independent.
If the manifest contains slices with test_level of integration, functional-api, functional-cli, functional-browser, or full, run kb-functional-test before kb-review to confirm the functional coverage is real and not mock-only. Also run it when the diff shows user-visible, API/CLI, persistence, auth, streaming, or integration changes without an adequate recorded test level.
If the final diff includes .tsx, .jsx, .vue, or .svelte files, or any changed behavior primarily reached through the rendered UI, require functional-browser evidence before completion. Backend/API/unit-only proof is insufficient for those changes.
Invoke the kb-review skill — structured code review on the feature diff.
kb-review is a skill/orchestrator, not an Agent tool type. Do not call the Agent tool with agent_type: kb-review. Load/run the kb-review skill, and let that skill spawn valid reviewer agent types such as code-review, correctness-reviewer, security-reviewer, or adversarial-reviewer.
This is mandatory. Do not skip, defer, or make it optional. Record the review
mode from kb-review: review-mode: multi-agent when reviewer agents actually
ran, or review-mode: local-fallback when the runtime could not or did not
authorize reviewer subagents. Do not claim multi-agent review happened when
fallback was used.
Pass scope from prior gates: use the collected scope-verified actual file list from Pre-Flight. Pass this as the scoped file list so kb-review skips its own scope discovery (Stage 1). The scope gates already explained planned and discovered files — no need to re-derive from git diff unless the manifest lacks this data.
Pass context: the full git diff of the feature branch against baseline, scoped to the verified file list
Capture the output: each finding has a severity (P0/P1/P2/P3) and confidence score
Capture review mode and any fallback residual risk.
Store findings for the resolution gate (Step 2)
Note: if scope-verified files are unavailable (older manifest, standalone run), let kb-review do its own scope discovery.
Step 2: Resolution Gate
Review findings from kb-review determine whether completion is allowed:
Severity
Action
P0 (critical)
STOP. Fix before proceeding. Re-run affected tests after fix.
P1 (important)
STOP. Fix before proceeding.
P2 (suggestion)
Log in manifest notes. Do not block.
P3 (nit)
Log in manifest notes. Do not block.
This gate is mandatory. The agent MUST NOT proceed to Step 3 while unresolved P0/P1 findings exist.
For any P2/P3 findings, invoke kb-gate with the rectify prompt. Do not silently leave fixable P2/P3 issues when the user would prefer a clean finish.
After resolving all P0/P1s, update the manifest notes with a summary:
review: P0=0 P1=2(resolved) P2=3(logged) P3=1(logged)
Feed learnings to the observation log:
For each resolved P0/P1 finding, append one line to .kb/observations.jsonl:
{"ts":"<ISO-8601>","hook":"kb-review","tool":"kb-finalize","args":{"finding_type":"<category>","severity":"P0|P1","resolution":"<what was fixed>"},"cwd":"<repo-root>","result":"resolved"}
This connects the review → learn pipeline. Only P0/P1 findings are worth learning from — P2/P3 are style preferences, not systemic patterns.
Create .kb/observations.jsonl if it doesn't exist. Append, never overwrite.
Step 2.5: Follow-Up Resolution Gate
Mirror the useful part of the original LFG finish pattern: do not leave known, fixable follow-up work unresolved just because the main implementation passed.
Collect unresolved review findings, TODO files, checklist items, and manifest
notes produced by kb-review, kb-gate, kb-work, kb-qa, or
kb-functional-test.
Resolve all safe/actionable P0/P1 findings before continuing.
For P2/P3 and todo-style follow-ups, run the smallest suitable resolver:
kb-gate rectify prompt, todo-triage, todo-create, or a repo-local
todo/PR-comment resolver if one is installed.
Do not parallelize follow-ups that touch the same files or depend on the same
decision. Parallel resolution is allowed only when file scopes are disjoint.
Record: follow-up-resolution: resolved N, logged M, blocked K.
Blocked/human-required items stay visible in todo.md or the manifest with evidence paths. They must not disappear into chat history.
Step 2.6: Proof and Demo Evidence Gate
Run a final evidence pass after review fixes, because review changes can alter
behavior.
Re-run kb-check or the narrow deterministic commands affected by review
fixes.
If user-visible, API/CLI, browser, persistence, auth, streaming, or
integration behavior changed, run kb-functional-test again on the final
diff. For browser/UI work, verify through the rendered UI with Playwright
where viable, or the repo/platform authenticated browser transport when
Playwright cannot access the route.
If the change is visual, workflow-heavy, reviewer-facing, or the user/PR
expects a demo, capture demo evidence. Use a repo-local feature-video or
equivalent demo skill if installed; otherwise capture screenshots, logs, or a
concise manual demo checklist and add an alert with the missing tool.
Store proof in the manifest notes: commands run, routes/screens/workflows
checked, artifacts created, and any skipped proof with reason.
Every slice must have machine-verifiable proof recorded in the manifest before completion can continue.
Acceptable proof formats:
test file path + exit code + timestamp;
Playwright/Cypress/CDP trace path or browser assertion artifact;
API response log path with status/schema assertion result;
CLI output log path with command and exit code;
go run ./cmd/kbcheck accept --check <check.json> --trace .kb/trace.jsonl result for repaired failures with a RED-before-GREEN trace;
regression snapshot result from .kb/snapshots/<slice-id>.json with all previous snapshots passing.
Not acceptable:
"I checked and it works";
"Page loaded successfully";
"Tests pass" without the actual command or test file, exit code, and timestamp;
latest passing check with no recorded prior RED when the work claims to fix a known failure;
screenshots or prose-only notes without an executable assertion/result;
any model-only visual inspection.
If any slice has only prose proof, this gate fails. Return to kb-work,
kb-check, kb-qa, kb-functional-test, or kb-regression-snapshot to produce
executable evidence before proceeding.
For objective-contract manifests, summarize the top-level done_check and each slice's proof_check result in the manifest notes. If the final objective check cannot run yet, record the blocker and do not declare KB <name> complete.
If routing evidence is partial, record the strongest honest attribution such as exact, explained-external, unknown, or unavailable, then continue with ordinary proof. Never rerun already-correct work only to upgrade provenance.
Step 3: Compound & Learn
After the resolution gate passes, document what this feature taught the system:
Classify steering feedback before /learn runs.
Sources: resolved P0/P1 review findings, manifest notes named
steering-feedback:, /iterate or PR feedback summaries, goal-ledger
feedback, and maintainer comments copied into the completion artifact.
Classify each item as exactly one primary route:
current-only, steering-memory, observation, landmine-candidate, or
instinct-evidence.
current-only: record in manifest notes only; do not update durable memory.
steering-memory: update the steering memory path named by the goal or
manifest, usually the goal ledger's Live Steering section or
docs/context/operations/steering/<slug>.md. Keep entries concise:
durable scope constraints, known false positives, reviewer preferences, or
selection guidance. Do not append raw transcripts or one-off PR details.
observation: append one JSONL entry to .kb/observations.jsonl.
Do not duplicate the resolved P0/P1 entries already written by Step 2;
reference those existing entries when they are the evidence source.
landmine-candidate: apply /learn landmine criteria; record only with
owner surface, concrete evidence, fix condition, and verification.
instinct-evidence: leave the evidence visible to /learn; do not
promote it directly to a skill.
If no steering memory path is named, do not create one automatically. Add a
manifest note: steering-feedback: no durable steering memory path.
Record the result in manifest notes:
steering-feedback: current=<N> memory=<N> observations=<N> landmine-candidates=<N> instinct-evidence=<N>.
Invoke ce-compound with context: a one-sentence summary of what was built and any surprising patterns discovered during implementation.
ce-compound writes to docs/solutions/ with YAML frontmatter — let it run without modification.
If the implementation was pure boilerplate (no novel patterns, no gotchas, no decisions worth preserving), skip with a manifest note: compound: skipped — standard implementation, no novel patterns
Per-slice micro-learnings from slice notes feed into the compound context. Reference them when invoking ce-compound.
Invoke /learn — Extract instincts from this session's work.
Run after compound completes (observations from Step 2 are now available)
/learn reads: observations.jsonl, recent git history, docs/solutions/,
existing instincts, and any steering-memory updates classified above
Record result in manifest notes: learn: N new instincts, M updated or learn: no new patterns
This is automatic — do not ask the user whether to run it
If the work exposed a repo-specific landmine, record it only with owner
surface, concrete evidence, severity, fix condition, and verification.
Reject generic "remember to test/read code" lessons.
If the work fixed an active landmine, move it to resolved/archive only
after the verification command passes.
Check evolution cadence:
Read docs/context/kb/kb-completions.txt (create with 0 if missing)
Increment by 1
Write the new value back
If the new value is divisible by 5:
Invoke /evolve to check for promotable instincts
Log result in manifest notes: evolve: promoted N instincts or evolve: no candidates ready
If not divisible by 5: skip silently
Commit the counter file with the manifest update
Step 3.5: Project Memory Refresh Gate
Before cleanup or final "complete", make sure a fresh session can resume without a lesson from the user.
Run kb-map refresh when any of these are true:
The manifest contains memory-impact: durable.
kb-work left any refresh=pending note.
Review fixes changed behavior, architecture, run/test commands, integrations, or known sharp edges.
docs/context/PROJECT.md points to stale subsystem docs after the feature diff.
Skip with a manifest note only when changes are clearly cosmetic, copy-only, formatting-only, lint-only, or isolated tests with no durable behavior change:
docs/context/PROJECT.md for route-map, command, or subsystem index changes.
docs/context/architecture/* for durable subsystem behavior.
docs/context/operations/* for run/test/deploy/QA changes.
docs/context/landmines.md for active/resolved landmine changes.
docs/context/research/* for reusable research outcomes.
docs/context/decisions/* for accepted/rejected approaches.
todo.md and handoff files for current operational state.
Then add:
kb-map-refresh: done - <docs updated>
Step 3.75: Memory Maintenance Signals
Update docs/context/memory-maintenance.md before cleanup. This file is a targeted signal index for future deep memory review. It must contain pointers, not just counters.
Create the file if it does not exist:
# Memory Maintenance
Last deep review: never
## Counters Since Last Review- Completed KB cycles: 0
- Durable memory refreshes: 0
- Closed handoffs: 0
- Contradiction signals: 0
- Overlap signals: 0
- Stale-doc signals: 0
- Bloat signals: 0
- Repeated-rediscovery signals: 0
## Signals Since Last Review
For every kb-finalize run:
Increment Completed KB cycles by 1.
Increment Durable memory refreshes if Step 3.5 ran kb-map refresh.
Increment Closed handoffs by the number of handoffs moved to docs/handoffs/done/ during this completion.
Append exact signals found during review, compound, learn, evolve, memory refresh, and cleanup.
Signal types:
Type
Record When
contradiction
two memory docs, handoffs, plans, requirements, or architecture notes disagree
overlap
two docs cover the same topic and may need consolidation/cross-linking
stale-doc
a memory doc references old paths, old behavior, old commands, or removed systems
bloat
a memory doc, handoff folder, research index, or todo-done.md is growing past useful startup size
repeated-rediscovery
this work re-learned something already discovered in another brainstorm, plan, research note, or solution
Signal entry format:
### YYYY-MM-DD - <type> - <shorttopic>- Source: `<repo-relative path or manifest note>`- Found during: `kb-finalize` / `<step or skill>`- Signal: <oneortwosentenceswiththeactualissue>- Suggested pass: <refresh, compact, consolidate, replace, cross-link, or promote>
Generic examples:
### 2026-05-24 - contradiction - workflow source of truth- Source: `docs/context/architecture/workflows.md`- Found during: `kb-map refresh`- Signal: requirements and architecture describe different canonical workflow definitions.
- Suggested pass: reconcile architecture and decision docs.
### 2026-05-24 - overlap - release deployment research- Source: `docs/context/research/release-packaging.md`- Found during: `kb-finalize`- Signal: overlaps older deployment research on the same update-delivery path.
- Suggested pass: consolidate or cross-link research notes.
Do not invent signals to satisfy a quota. If no signal exists, only increment counters and add a manifest note: memory-maintenance: no new signals.
If any signal counter crosses a conservative threshold, add a recommendation to the final report, but do not run the deep pass automatically:
Completed KB cycles >= 5
Durable memory refreshes >= 10
Closed handoffs >= 10
Any contradiction signals >= 2
Any combined signals >= 8
Recommendation format:
Memory review recommended: <reason>. Run `kb-memory-review` against docs/context/memory-maintenance.md before the next large feature.
Step 3.7: Compact and Alert Gate
Run kb-compact when review, learning, memory refresh, or maintenance signals show that durable memory is getting too large for fresh-session startup.
Use targeted compaction only:
Compact a specific bloated architecture doc, handoff, research note, todo-done.md, or memory-maintenance section.
Preserve exact paths, commands, current truth, known sharp edges, and unresolved decisions.
Do not compact away active work, blockers, HITL items, open handoffs, or evidence needed for review.
Record the result in manifest notes: compact: <path> compacted or compact: skipped - no startup bloat.
Alert the user in the final report when any condition needs deliberate follow-up:
kb-memory-review threshold crossed.
P2/P3 findings remain logged after the rectify gate.
A memory contradiction, stale-doc signal, or repeated-rediscovery signal was recorded.
Compacting was skipped because the doc was too risky to summarize safely.
A required tool, reviewer, compound, learn, evolve, refresh, or compact step failed.
Alerts are concise status lines, not extra ceremony. They should tell the user exactly what needs attention and which file contains the evidence.
Step 4: Cleanup
Prune ephemeral artifacts. Heavy KB usage generates file sprawl — clean it up per-feature, not manually.
QA screenshots — delete .kb/qa-screenshots/ contents for this feature's slices. Screenshots should already be referenced in commits or PR bodies. Safe to remove.
Observations log — trim .kb/observations.jsonl entries older than 90 days. Matches the recency decay half-life in /learn. Use any available local scripting runtime; if no suitable runtime is available or the file does not exist, skip with a note.
Plan files — leave manifests and slice plans in docs/plans/. Lightweight reference material, useful for tracing decisions.
Log cleanup in the manifest notes:
cleanup: screenshots pruned, observations trimmed to 90d
Todo hygiene — verify todo.md contains only active rows, 🔒 blocked rows, 🛑 human-required rows, the 🧊 Parked / Cold Storage section, or handoff-pointer work. If the completed feature or routine slice completion logs remain there, move a compact summary to todo-done.md and remove those entries from todo.md. todo.md must keep its ## Rules section at the top; do not depend on todo_rules.md, todo-rules.md, or any separate rules file.
Step 5: Done
Before updating the manifest to status: reviewed, write complete-to-ship in the manifest gate_ledger.
Required proof:
kb-check final command/result;
objective done_check result, or explicit scoped human exception;
per-slice proof_check result or accepted no_check_reason;
kb-functional-test or explicit skip reason for every functional/API/CLI/UI slice;
kb-review mode and finding counts;
P0/P1 resolved or human/quarantine blocker recorded;
follow-up-resolution summary;
proof/demo evidence paths or skip reason;
compound/learn/evolve result or non-blocking failure note;
project-memory refresh/skip proof;
memory-maintenance update;
cleanup result;
alerts list.
If any required proof is missing, set complete-to-ship to blocked and do not report KB <name> finalized.
Update the manifest status: reviewed only after complete-to-ship is passed or explicitly quarantined for out-of-scope issues, and after kb-gate/scripts/check_gate_ledger.py <manifest-path> --gate complete-to-ship --allow-quarantine passes. Then report:
KB <name> finalized.
- Review: P0=N P1=N(resolved) P2=N P3=N
- Follow-up resolution: <resolved N | logged M | blocked K>
- Proof/demo evidence: <commands/artifacts | skipped with reason>
- Compound: <written | skipped>
- Learn: <N new, M updated | no new patterns>
- Evolve: <promoted N | skipped | no candidates>
- Project memory: <refreshed | skipped with reason>
- Memory maintenance: <N signals recorded | no new signals | review recommended>
- Compact: <ran on paths | skipped with reason>
- Alerts: <none | concise follow-up lines with evidence paths>
- Cleanup: done
Ready for configured delivery. Return control to `kb-complete <manifest>`.
Failure Handling
Situation
Action
kb-review fails to run
Log error, ask user whether to retry or skip review
P0/P1 fix breaks tests
Re-run tests, treat as new failure, fix before proceeding
proof/demo evidence cannot run
Log the missing tool/server/credential and alert user with the closest deterministic proof completed
compound/learn/evolve fails
Log error, continue — these are non-blocking
kb-compact fails
Log error and alert user with the bloated file path; do not block completion
Manifest not found
Ask user for path
Manifest has unfinished slices
Stop, tell user to run kb-work first
Integration
Input from:kb-work or kb-complete (completed manifest)
Review engine:kb-review with scope passthrough
Follow-up resolution:kb-gate, todo-triage, todo-create, or repo-local resolvers
