Skip to content

Self-evolving second brain with 35 AI skills, 10 agents, and people CRM. Closed-loop harness: a V-model verification lifecycle where the worker never grades its own homework. Plus paired anti-slop design skills for marketing and product UI. Works with Claude Code, Cursor, Kiro, Gemini CLI, Codex.

9.9 KB
Raw Blame

Vault Harness Workflow

V-model harness: decompose on the left, verify with evidence on the right, build at the apex, retro closes the loop. Adapted from dwarves-kit + V-model SDLC discipline.

Scope: opt-in only

This document governs harness runs, and nothing else. A session that never invoked the harness owes it no checkpoints, no lane classification, and no evidence ledger. Notes, briefs, research, drafts, and ordinary edits are not harness runs.

You are in a harness run when you invoked /closed-loop, /ultragoal, /retro, /harvest, or /review-cockpit; asked for the closed loop, proper verification, or an evidence trail in those words; or set verification_harness: on in 00-inbox/MY-PROFILE.md. Otherwise you are not, and the two rules in CLAUDE.md § Verification Harness are the whole of what applies.

Nothing here needs installing. The two helper scripts (.claude/lib/checkpoint.sh, .claude/lib/lane-classify.sh) ship executable, and the run directories are created on first use.

The V-model (primary mental model)

Inside a harness run, every non-tiny task walks the V. Each checkpoint (CP) is a gate: you cannot descend the left arm past a failed CP, and you cannot ascend the right arm without evidence tied back to a criterion ID.

                    CP-0 INTAKE (think)
                   ╱  evidence: questions answered
                  ╱
         CP-1 SPEC ──────────────── CP-5 ACCEPTANCE
        ╱  criteria + trace matrix    ╲  post-condition artifacts
       ╱                               ╲
  CP-2 PLAN ───────────────────── CP-4 INTEGRATION
      task ↔ criterion map              cross-task wiring
              ╲                    ╱
               ╲   CP-3 BUILD   ╱
                ╲  (execute)  ╱
                 ╲────────────╱
                  component verify (task-verifier)
                           │
                      CP-6 SHIP
                           │
                      CP-7 RETRO ──► feeds CP-0 next cycle

Two-way verification

Direction What Artifact
Down (left arm) Decompose goal → falsifiable criteria → tasks Spec + traceability matrix (AC-01 → task T-01)
Up (right arm) Prove each task → prove wiring → prove acceptance Evidence ledger rows (AC-01 ← verifier observation)
Bidirectional Every AC-n has ≥1 task AND ≥1 evidence row before ship Matrix status = traced

No criterion ships without a matching evidence row. No evidence row without a criterion ID.

Checkpoints

CP Phase Gate class Pass requires Evidence file
CP-0 Intake / think advisory Forcing questions answered or lane=tiny skip evidence/CP-0-intake.md
CP-1 Spec blocking (normal+) ## Acceptance criteria with IDs (AC-01…) spec itself + matrix
CP-2 Plan blocking (normal+) Tasks reference AC-n; criteria falsifiable evidence/CP-2-plan.md
CP-3 Build — Worker deliverable exists deliverable path
CP-3v Component verify blocking task-verifier PASS per task evidence/CP-3v-component.md
CP-4 Integration verify blocking (full+, multi-task) integration-verifier PASS evidence/CP-4-integration.md
CP-5 Acceptance blocking (mutations) Post-condition observed (not tool return) evidence/CP-5-acceptance.md
CP-6 Ship blocking (external) Review Gate / your approval / deploy proof evidence/CP-6-ship.md
CP-7 Retro advisory Retro doc + harvest staged 04-projects/harness/retro/YYYY-MM-DD-<slug>.md

Record checkpoints: bash .claude/lib/checkpoint.sh record <run-dir> <CP-id> PASS|FAIL|SKIP <note> (init <run-dir> first to lay down the evidence dir and ledger header)

Evidence row contract (every verify pass)

EVIDENCE <AC-id> | <checkpoint> | PASS|FAIL | <observation> | <artifact-path-or-command>

Verifier and post-condition steps emit these rows. Consolidate in evidence/ledger.md per run.

Gate classes

Class Examples Behavior
Blocking CP-1 spec, CP-3v verifier, CP-5 post-condition, CP-6 ship, safety-gate Stops a bad outcome
Advisory CP-0 think, roundtable, slop-cleaner, CP-7 retro, cross-model flagship gate Surfaces findings; retro strongly expected

Cross-model flagship gate (full+ / ultragoal / irreversible)

The read-only verifiers (task-verifier CP-3v, integration-verifier CP-4) are same-family mechanical checks that share the lead's blind spots. On high-stakes runs, overlay a flagship model that is NOT the lead's own (spawn the verifier with an explicit model override from a different family) as a cross-model second opinion. Fresh context (paths + question, never the lead's draft-reasoning), read-only, advisory.

CP Flagship role What it checks
CP-1 Spec advisor Are the AC-n falsifiable, complete, the right criteria? (highest leverage — a wrong spec poisons the left arm)
CP-4 Integration critic Cross-task wiring / global acceptance holds up under a different reasoner
CP-5 Acceptance critic Artifact truly satisfies the north-star, not just a passing post-condition
CP-6 Ship critic Break it before the user sees it — last gate before external/irreversible
Ultragoal critic Per-phase gate AND final north-star acceptance (adversarial by default; wrongness compounds across sessions)

Advisory, not hard-blocking, but a critical cross-model finding means do not auto-ship; escalate the disagreement to the user to adjudicate. Skip entirely on tiny/normal runs (pure overhead there). If the cross-model reviewer hard-refuses (offensive-security / bio-adjacent), rerun the check on the lead's own family rather than dropping it.

Risk lanes (checkpoint depth)

Lane Checkpoints required
tiny CP-3 → CP-5 (if mutating) only
normal CP-1 → CP-2 → CP-3 → CP-3v → CP-5
full all through CP-4 + claim-verifier + CP-6 Review Gate
bug CP-0 root-cause ledger → CP-3 → CP-3v (3-fix wall)
backfill CP-1 audit only; no CP-3 until approved

Classifier: bash .claude/lib/lane-classify.sh classify "<task>"

Verification pipeline (right arm detail)

orchestrator (/closed-loop or a skill that invokes it)
        │
        ▼
   CP-3 BUILD: worker implements (traced to AC-n)
        │
        ▼
   CP-3v: task-verifier (read-only) ──► evidence rows per AC-n
        │
        ├── FAIL:fixable ─► fix-agent (max 2) ─► re-verify
        └── FAIL:escalate ─► stop
        │
        ▼
   CP-4: integration-verifier (multi-task / full lane only)
        │
        ▼
   CP-5: post-condition (observe artifact: curl, screenshot, re-fetch)
        │
        ▼
   CP-6: ship gate (you approve anything external / deploy proof)
        │
        ▼
   CP-7: /retro + /harvest

Domain routing

Work type Primary skill Right-arm verifiers
Single task through the loop /closed-loop CP-3v task-verifier + CP-5 post-condition
Long-running goal /ultragoal full closed-loop per phase + north-star acceptance verifier
Team intelligence /team-brief claim-verifier (CP-3v) + CP-6
Research / analysis /auto-research, /comprehensive-analysis citation verbatim check (CP-3v)
Content /content-factory voice checklist + screenshot (CP-5)
Memory store /memory-hygiene environment re-verify (CP-5)
Session learnings /harvest human promotes (CP-7 input)
Multi-item review /review-cockpit your approval per card (CP-6)

Self-enhancement loops

Loop CP Output
Closed loop CP-3 → CP-5 Evidence ledger + loop-ledger.tsv
Ultragoal CP-1 → CP-6 per phase 04-projects/<goal>/STATUS.md + evidence per phase + report.html
Harvest CP-7 Staging file the human promotes into 05-knowledge/
Retro CP-7 04-projects/harness/retro/ + skill patches from friction
Memory hygiene CP-5 on memory store last_verified stamps

File homes

Run output lives under 04-projects/harness/, which is created on the first harness run. Templates ship with the skills, so /update-cog keeps them current.

Artifact Path
Spec + traceability matrix 04-projects/<project>/specs/SPEC-NNN-<slug>.md
Run evidence bundle 04-projects/harness/runs/<id>/evidence/
HTML report (ultragoal / big run) 04-projects/<goal>/report.html · 04-projects/harness/runs/<id>/report.html
Retro outputs 04-projects/harness/retro/YYYY-MM-DD-<slug>.md
Harness backlog 04-projects/harness/BACKLOG.md
Ultragoal registry 04-projects/harness/ultragoals.md
Harvest staging 04-projects/harness/harvest/staging-<date>.md
Checkpoint + loop logs .claude/logs/checkpoint-ledger.tsv, loop-ledger.tsv
Spec template .claude/skills/closed-loop/references/spec-template.md
Report template .claude/skills/closed-loop/references/report-template.html
Retro template .claude/skills/retro/references/retro-template.md
Review cockpit template .claude/skills/review-cockpit/references/session-review-template.md

Commands

Command V-model phase
/closed-loop CP-2 through CP-5
/retro <run or spec> CP-7
/harvest CP-7 input
/ultragoal full V per phase, across sessions
/memory-hygiene CP-5 on memory

No install step

The harness has no installer. .claude/lib/*.sh ship executable, run directories are created on demand, and COG ships no hooks. If a skill body ever tells you to run install-harness.sh, that reference is stale; report it.