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.

fix: make the verification harness opt-in and repair its dangling references (v3.12.0) (#44)

Huy Tieu623ed001ca6e
diff --git a/.agents/skills/closed-loop/SKILL.md b/.agents/skills/closed-loop/SKILL.md
index ba2109e..4e3522d 100644
--- a/.agents/skills/closed-loop/SKILL.md
+++ b/.agents/skills/closed-loop/SKILL.md
@@ -1,6 +1,6 @@
---
name: closed-loop
-description: V-model execute: CP-2 plan → CP-3 build → CP-3v component verify → CP-4 integration verify (full) → CP-5 acceptance. Worker never grades its own homework; evidence rows trace back to AC-n. Use via /execute.
+description: Run one task through the V-model verification loop: CP-2 plan → CP-3 build → CP-3v component verify → CP-4 integration verify (full lane) → CP-5 acceptance. The worker never grades its own homework; evidence rows trace back to AC-n. Opt-in: invoke with /closed-loop or by asking for the closed loop, proper verification, or an evidence trail. Ordinary work does not run this.
---
Read `.claude/skills/closed-loop/SKILL.md` and execute it exactly as written — that file is the authoritative
diff --git a/.agents/skills/ultragoal/SKILL.md b/.agents/skills/ultragoal/SKILL.md
index bcf018a..26c5c7f 100644
--- a/.agents/skills/ultragoal/SKILL.md
+++ b/.agents/skills/ultragoal/SKILL.md
@@ -1,6 +1,6 @@
---
name: ultragoal
-description: Run a large, multi-session goal (e.g. shipping a whole side product) through the full V-model closed loop, one phase at a time, with cross-session state and a final north-star acceptance gate. Ultragoals never downgrade the lane: every phase runs CP-1→CP-6 with adversarial verification. Use via /ultragoal.
+description: Run a large, multi-session goal (e.g. shipping a whole side product) through the full V-model closed loop, one phase at a time, with cross-session state and a final north-star acceptance gate. Ultragoals never downgrade the lane: every phase runs CP-1→CP-6 with adversarial verification. Opt-in: invoke with /ultragoal or by calling something a long-running goal. Ordinary work does not run this.
---
Read `.claude/skills/ultragoal/SKILL.md` and execute it exactly as written — that file is the authoritative
diff --git a/.claude-plugin/plugin.json b/.claude-plugin/plugin.json
index 2cd8e89..0e5d730 100644
--- a/.claude-plugin/plugin.json
+++ b/.claude-plugin/plugin.json
@@ -1,6 +1,6 @@
{
"name": "cog-second-brain",
- "version": "3.11.0",
+ "version": "3.12.0",
"description": "Self-evolving second brain with 33 AI agent skills, 10 agents, and people CRM. Closed-loop harness release: a V-model verification lifecycle (spec to evidence, criterion-traced checkpoints), read-only verifier agents that observe artifacts instead of trusting worker summaries, risk lanes, multi-session ultragoals with a north-star acceptance gate, plus craft skills for writing, visual work, and a paired anti-slop design system for marketing and product UI. Inspired by Garry Tan's gstack specialist sessions, gbrain knowledge patterns, and dwarves-kit closed-loop execution. Works with Claude Code, Antigravity, Cursor, Kiro, Gemini CLI, Codex, and any markdown-reading agent.",
"author": {
"name": "Huy Tieu",
diff --git a/.claude/lib/checkpoint.sh b/.claude/lib/checkpoint.sh
index 5fe3585..55b79a0 100755
--- a/.claude/lib/checkpoint.sh
+++ b/.claude/lib/checkpoint.sh
@@ -16,7 +16,21 @@ shift || true
init_run() {
local dir="$1"
mkdir -p "$dir/evidence"
- [[ -f "$dir/evidence/ledger.md" ]] || cp "$ROOT/04-projects/harness/templates/evidence-ledger.md" "$dir/evidence/ledger.md"
+ if [[ ! -f "$dir/evidence/ledger.md" ]]; then
+ cat > "$dir/evidence/ledger.md" <<'LEDGER'
+# Evidence ledger
+
+One row per verify pass. The observation records what was observed in the
+artifact, never what a worker reported.
+
+```text
+EVIDENCE <AC-id> | <checkpoint> | PASS|FAIL | <observation> | <artifact-path-or-command>
+```
+
+## Rows
+
+LEDGER
+ fi
echo "initialized: $dir/evidence/"
}
diff --git a/.claude/skills/closed-loop/SKILL.md b/.claude/skills/closed-loop/SKILL.md
index 2717dad..2ec9639 100644
--- a/.claude/skills/closed-loop/SKILL.md
+++ b/.claude/skills/closed-loop/SKILL.md
@@ -1,9 +1,11 @@
---
name: closed-loop
description: >
- V-model execute: CP-2 plan → CP-3 build → CP-3v component verify →
- CP-4 integration verify (full) → CP-5 acceptance. Worker never grades
- its own homework; evidence rows trace back to AC-n. Use via /execute.
+ Run one task through the V-model verification loop: CP-2 plan → CP-3 build →
+ CP-3v component verify → CP-4 integration verify (full lane) → CP-5 acceptance.
+ The worker never grades its own homework; evidence rows trace back to AC-n.
+ Opt-in: invoke with /closed-loop or by asking for the closed loop, proper
+ verification, or an evidence trail. Ordinary work does not run this.
---
# Closed-loop execute (V-model right arm)
@@ -12,9 +14,14 @@ Mechanical verification pipeline. Every verify step emits **evidence rows** tied
## When to use
-- `/execute <task>` or `/execute <spec-path>`
-- Any skill at `normal`+ lane after implementation
-- Code changes, deploys, multi-step vault deliverables
+The harness is opt-in. Run it when:
+
+- Invoked as `/closed-loop <task>` or `/closed-loop <spec-path>`.
+- The user asks for the closed loop, proper verification, or an evidence trail.
+- `verification_harness: on` in `00-inbox/MY-PROFILE.md` and this is a build task.
+- Another skill that declares a `normal`+ lane reaches its verify step.
+
+Do **not** run it on a request that did not ask for it. Notes, briefs, research, drafts, and ordinary edits are not harness runs, and a checkpoint ledger on those is pure overhead.
## Phase 0 — Lane + run folder
@@ -36,7 +43,7 @@ Record: `checkpoint.sh record <run-dir> CP-0 PASS|SKIP "<lane>"`
If spec exists, use its `## Acceptance criteria` + traceability matrix. Else write:
-`04-projects/harness/runs/<id>/criteria.md` using `04-projects/harness/templates/SPEC-template.md` (criteria + matrix only).
+`04-projects/harness/runs/<id>/criteria.md` using `references/spec-template.md` (criteria + matrix sections only).
Each criterion: **falsifiable** + `AC-n` ID + verify method.
@@ -83,7 +90,7 @@ For each mutation, observe artifact (curl, screenshot, re-fetch). Emit:
`EVIDENCE AC-n | CP-5 | PASS | <observation> | <artifact>`
-**UI/UX flow changes:** the post-condition is *visual*. Capture with browser-harness (`evidence_shot` per state; `FlowRecorder`→`.save_gif()` for a flow; `pixel_diff` against the intended/prior state), then read the image and confirm no overflow/misalignment/clipping/wrong-color/broken-responsive before PASS. The Observation must describe what you saw; the artifact is the screenshot/GIF in `evidence/`. Fix any visual defect and re-capture. See CLAUDE.md → Visual Verification.
+**UI/UX flow changes:** the post-condition is *visual*. Screenshot every meaningful state with whatever browser tooling the environment has, then read the image and confirm no overflow, misalignment, clipping, wrong color, or broken responsive layout before PASS. The Observation must describe what you saw; the artifact is the screenshot/GIF in `evidence/`. Fix any visual defect and re-capture. See CLAUDE.md → Visual Verification.
Write `evidence/CP-5-acceptance.md`. **Traceability closure**: every AC in matrix has ≥1 PASS row in ledger.
@@ -93,18 +100,18 @@ Record: `checkpoint.sh record <run-dir> CP-5 PASS|FAIL`
- Append to `.claude/logs/loop-ledger.tsv`
- Update spec traceability matrix statuses to `verified`
-- **`full` lane / big task:** generate an HTML rollup from `04-projects/harness/templates/report.html` → `04-projects/harness/runs/<id>/report.html`, filled from `criteria.md` + `evidence/ledger.md` (criteria, AC traceability, verifier verdicts, post-condition observations). Self-contained; `SendUserFile` it or publish as an Artifact. Skip for `normal`/`tiny`.
+- **`full` lane / big task:** generate an HTML rollup from `references/report-template.html` → `04-projects/harness/runs/<id>/report.html`, filled from `criteria.md` + `evidence/ledger.md` (criteria, AC traceability, verifier verdicts, post-condition observations). Self-contained; `SendUserFile` it or publish as an Artifact. Skip for `normal`/`tiny`.
- Suggest `/retro <run-dir>` for CP-7
## Integration
| Skill | Lane | CP-4 |
|---|---|---|
-| ultragoal | `full` per phase (never downgraded) | integration-verifier + north-star acceptance |
-| team-brief | full | claim-verifier |
-| dogfood-release | full | Playwright cross-verify |
-| blog-publish | normal | skip |
-| content-factory | normal | skip |
+| `ultragoal` | `full` per phase (never downgraded) | integration-verifier + north-star acceptance |
+| `team-brief` | full | claim-verifier |
+| `comprehensive-analysis`, `auto-research` | full | claim-verifier on cited claims |
+| `content-factory` | normal | skip |
+| `review-cockpit` | normal | skip; CP-6 is the user's approval per card |
## Escalation template
diff --git a/.claude/skills/closed-loop/references/report-template.html b/.claude/skills/closed-loop/references/report-template.html
new file mode 100644
index 0000000..d6356bc
--- /dev/null
+++ b/.claude/skills/closed-loop/references/report-template.html
@@ -0,0 +1,125 @@
+<!doctype html>
+<!--
+ COG harness run report: fill by editing, do not build a parser.
+ Replace every {{token}}. Delete or repeat each <!-- REPEAT --> row block.
+ Self-contained: no external CSS, JS, fonts, or images. Inline any screenshot
+ as a data: URI, or link it relative to this file inside the run folder.
+-->
+<html lang="en">
+<head>
+<meta charset="utf-8">
+<meta name="viewport" content="width=device-width, initial-scale=1">
+<title>{{title}}: run report</title>
+<style>
+ :root {
+ --bg: #fbfbfa; --panel: #ffffff; --ink: #1a1a19; --muted: #6b6b68;
+ --line: #e4e4e1; --pass: #1a7f4b; --fail: #b4232a; --skip: #8a8a86;
+ --accent: #1a1a19;
+ --mono: ui-monospace, SFMono-Regular, "SF Mono", Menlo, monospace;
+ --sans: -apple-system, BlinkMacSystemFont, "Segoe UI", Inter, system-ui, sans-serif;
+ }
+ @media (prefers-color-scheme: dark) {
+ :root:not([data-theme="light"]) {
+ --bg: #131313; --panel: #1b1b1a; --ink: #ececea; --muted: #9a9a96;
+ --line: #2c2c2a; --pass: #4ec281; --fail: #f2777c; --skip: #7d7d79;
+ --accent: #ececea;
+ }
+ }
+ :root[data-theme="dark"] {
+ --bg: #131313; --panel: #1b1b1a; --ink: #ececea; --muted: #9a9a96;
+ --line: #2c2c2a; --pass: #4ec281; --fail: #f2777c; --skip: #7d7d79;
+ --accent: #ececea;
+ }
+ * { box-sizing: border-box; }
+ body { margin: 0; background: var(--bg); color: var(--ink); font: 15px/1.55 var(--sans); }
+ .wrap { max-width: 62rem; margin: 0 auto; padding: 3rem 1.5rem 6rem; }
+ header { border-bottom: 1px solid var(--line); padding-bottom: 1.5rem; margin-bottom: 2.5rem; }
+ h1 { font-size: 1.6rem; letter-spacing: -0.01em; margin: 0 0 .4rem; }
+ .meta { color: var(--muted); font: 13px/1.5 var(--mono); }
+ h2 { font-size: 1rem; text-transform: uppercase; letter-spacing: .08em; color: var(--muted);
+ margin: 2.75rem 0 .9rem; font-weight: 600; }
+ p { margin: 0 0 1rem; }
+ .northstar { background: var(--panel); border: 1px solid var(--line); border-left: 3px solid var(--accent);
+ padding: 1rem 1.15rem; border-radius: 3px; }
+ .scroll { overflow-x: auto; border: 1px solid var(--line); border-radius: 3px; background: var(--panel); }
+ table { width: 100%; border-collapse: collapse; font-size: 14px; }
+ th, td { text-align: left; padding: .6rem .8rem; border-bottom: 1px solid var(--line); vertical-align: top; }
+ th { font: 600 12px/1.4 var(--mono); text-transform: uppercase; letter-spacing: .05em; color: var(--muted); }
+ tr:last-child td { border-bottom: 0; }
+ td.id, td.cp { font-family: var(--mono); white-space: nowrap; }
+ .v { font: 600 12px/1.4 var(--mono); white-space: nowrap; }
+ .v.pass { color: var(--pass); } .v.fail { color: var(--fail); } .v.skip { color: var(--skip); }
+ .obs { font-family: var(--mono); font-size: 12.5px; color: var(--muted); }
+ figure { margin: 1.25rem 0; }
+ figure img { display: block; max-width: 100%; border: 1px solid var(--line); border-radius: 3px; }
+ figcaption { color: var(--muted); font: 12.5px/1.5 var(--mono); margin-top: .5rem; }
+ .next { background: var(--panel); border: 1px solid var(--line); padding: 1rem 1.15rem; border-radius: 3px; }
+ footer { margin-top: 3.5rem; padding-top: 1.25rem; border-top: 1px solid var(--line);
+ color: var(--muted); font: 12.5px/1.5 var(--mono); }
+</style>
+</head>
+<body>
+<div class="wrap">
+
+<header>
+ <h1>{{title}}</h1>
+ <div class="meta">{{run_id}} · lane {{lane}} · {{date}} · outcome {{shipped|escalated}}</div>
+</header>
+
+<h2>North star</h2>
+<p class="northstar">{{one paragraph: what is true now that was not true before}}</p>
+
+<h2>Phases</h2>
+<div class="scroll">
+<table>
+ <thead><tr><th>Phase</th><th>Scope</th><th>Covers</th><th>State</th></tr></thead>
+ <tbody>
+ <!-- REPEAT one row per phase; delete this section entirely for a single-task run -->
+ <tr><td class="id">{{P0}}</td><td>{{scope}}</td><td class="id">{{AC-01, AC-02}}</td>
+ <td class="v pass">{{done}}</td></tr>
+ </tbody>
+</table>
+</div>
+
+<h2>Acceptance criteria</h2>
+<div class="scroll">
+<table>
+ <thead><tr><th>AC</th><th>Criterion</th><th>Task</th><th>Verdict</th></tr></thead>
+ <tbody>
+ <!-- REPEAT one row per AC in the traceability matrix -->
+ <tr><td class="id">{{AC-01}}</td><td>{{criterion}}</td><td class="id">{{T-01}}</td>
+ <td class="v pass">{{PASS}}</td></tr>
+ </tbody>
+</table>
+</div>
+
+<h2>Evidence</h2>
+<div class="scroll">
+<table>
+ <thead><tr><th>AC</th><th>CP</th><th>Verdict</th><th>Observation</th><th>Artifact</th></tr></thead>
+ <tbody>
+ <!-- REPEAT one row per EVIDENCE line in evidence/ledger.md -->
+ <tr><td class="id">{{AC-01}}</td><td class="cp">{{CP-3v}}</td><td class="v pass">{{PASS}}</td>
+ <td class="obs">{{what was observed, not what was claimed}}</td>
+ <td class="obs">{{path or command}}</td></tr>
+ </tbody>
+</table>
+</div>
+
+<h2>Screenshots</h2>
+<!-- REPEAT one figure per captured state; delete the section if the run had no visual surface -->
+<figure>
+ <img src="{{evidence/shot-01.png}}" alt="{{what this shows}}">
+ <figcaption>{{state name}}: {{what to look at, and what was fixed if anything}}</figcaption>
+</figure>
+
+<h2>Next action</h2>
+<p class="next">{{the single next thing, or "none: run closed"}}</p>
+
+<footer>
+ Generated by the COG verification harness · evidence bundle {{04-projects/harness/runs/&lt;id&gt;/evidence/}}
+</footer>
+
+</div>
+</body>
+</html>
diff --git a/.claude/skills/closed-loop/references/spec-template.md b/.claude/skills/closed-loop/references/spec-template.md
new file mode 100644
index 0000000..3ce5684
--- /dev/null
+++ b/.claude/skills/closed-loop/references/spec-template.md
@@ -0,0 +1,47 @@
+# SPEC-NNN: <title>
+
+> Lane: `tiny|normal|full|bug|backfill` · Run: `04-projects/harness/runs/<id>` · Date: YYYY-MM-DD
+
+## Goal
+
+One paragraph. What is true when this is done that is not true now.
+
+## Non-goals
+
+What this deliberately does not cover, so scope creep is visible.
+
+## Acceptance criteria
+
+Every criterion is **falsifiable**: a named observation either happens or it does not.
+"Works well" is not a criterion. "`curl -s <url> | grep -q '<string>'` exits 0" is.
+
+| ID | Criterion | Verify method |
+|---|---|---|
+| AC-01 | <observable statement> | <command, screenshot, re-fetch, diff> |
+| AC-02 | | |
+
+## Traceability matrix
+
+Status: `pending` at CP-2, `built` at CP-3, `verified` once a PASS row exists in the ledger.
+
+| AC | Task | Evidence row | Status |
+|---|---|---|---|
+| AC-01 | T-01 | | pending |
+| AC-02 | T-02 | | pending |
+
+## Tasks
+
+| ID | Task | Covers |
+|---|---|---|
+| T-01 | <what gets built> | AC-01 |
+| T-02 | | AC-02 |
+
+## Phases (ultragoal only)
+
+| Phase | Scope | Covers | State |
+|---|---|---|---|
+| P0 | | AC-01 | not started |
+
+## Risks and open questions
+
+- <risk>: mitigation or the decision needed.
diff --git a/.claude/skills/harvest/SKILL.md b/.claude/skills/harvest/SKILL.md
index 8ab5811..6fa9c75 100644
--- a/.claude/skills/harvest/SKILL.md
+++ b/.claude/skills/harvest/SKILL.md
@@ -64,12 +64,13 @@ Harvest: <n> staged, <m> promoted, <k> folded
- Contradictions with existing knowledge → flag both, ask the user.
- Pair with `/memory-hygiene` monthly for environment-dependent facts in `~/.claude/.../memory/`.
-## Nightly enhance (optional launchd)
-
-```bash
-# 04-projects/harness/scripts/nightly-enhance.sh
-bash .claude/lib/install-harness.sh --check # hooks alive
-# /memory-hygiene (skill)
-# /harvest promote if staging non-empty
-# content-factory if scheduled
+## Nightly enhance (optional)
+
+If you want harvest to run unattended, schedule a job that invokes the skills in
+this order. There is nothing to install first, since COG ships no hooks.
+
+```
+/memory-hygiene
+/harvest promote # only if the staging file is non-empty
+/content-factory # only if scheduled
```
diff --git a/.claude/skills/onboarding/references/profile-templates.md b/.claude/skills/onboarding/references/profile-templates.md
index dee6664..9d3c2cc 100644
--- a/.claude/skills/onboarding/references/profile-templates.md
+++ b/.claude/skills/onboarding/references/profile-templates.md
@@ -11,6 +11,7 @@ created: YYYY-MM-DD
onboarding_completed: true
role_pack: [matched role_id or "custom"]
agent_mode: [solo or team, based on role pack suggestion]
+verification_harness: off
tags: ["#profile", "#config", "#cog"]
---
@@ -24,6 +25,7 @@ tags: ["#profile", "#config", "#cog"]
## Settings
- **Agent Mode**: [solo/team] *(solo = handle everything directly; team = delegate to specialist sub-agents for deeper results)*
+- **Verification Harness**: off *(off = the V-model harness runs only when you ask for it, e.g. `/closed-loop`; on = build tasks default to the `normal`-lane pipeline)*
## Active Projects
[If they mentioned projects:]
diff --git a/.claude/skills/retro/SKILL.md b/.claude/skills/retro/SKILL.md
index 0a1dcfd..18aad21 100644
--- a/.claude/skills/retro/SKILL.md
+++ b/.claude/skills/retro/SKILL.md
@@ -48,7 +48,7 @@ Ask:
## Phase 4 — Write retro
-Copy `04-projects/harness/templates/retro.md` to:
+Copy `references/retro-template.md` to:
`04-projects/harness/retro/YYYY-MM-DD-<slug>.md`
diff --git a/.claude/skills/retro/references/retro-template.md b/.claude/skills/retro/references/retro-template.md
new file mode 100644
index 0000000..831a1b7
--- /dev/null
+++ b/.claude/skills/retro/references/retro-template.md
@@ -0,0 +1,32 @@
+# Retro: <run or spec>
+
+> Date: YYYY-MM-DD · Run: `04-projects/harness/runs/<id>` · Lane: `<lane>` · Outcome: shipped | escalated | abandoned
+
+## What happened
+
+Two or three sentences. The sequence, not the narrative.
+
+## Evidence quality
+
+| AC | Had a PASS row | Observed the artifact | Notes |
+|---|---|---|---|
+| AC-01 | yes/no | yes/no: what was observed | |
+
+Any `PASS` that rested on a worker's summary instead of the artifact is a finding, not a pass.
+
+## What the gates caught
+
+| CP | Verdict | What it caught, or why it was silent |
+|---|---|---|
+
+## Friction
+
+Where the run lost time or nearly went wrong. One line each, concrete.
+
+## Actions
+
+`AI-n` rows go to `04-projects/harness/BACKLOG.md`. Skill or `CLAUDE.md` patches are proposed here, applied only with approval.
+
+| ID | Action | Target file | Status |
+|---|---|---|---|
+| AI-1 | | | proposed |
diff --git a/.claude/skills/review-cockpit/SKILL.md b/.claude/skills/review-cockpit/SKILL.md
index 78d4a48..94544d9 100644
--- a/.claude/skills/review-cockpit/SKILL.md
+++ b/.claude/skills/review-cockpit/SKILL.md
@@ -23,7 +23,7 @@ Pairs with `closed-loop` (verification), `harvest` (learnings), and the V-model
## Structure
-Copy `04-projects/harness/templates/session-review.md`. Save the working copy in the relevant project folder as `YYYY-MM-DD-<slug>-review.md` (or `-plan.md`).
+Copy `references/session-review-template.md`. Save the working copy in the relevant project folder as `YYYY-MM-DD-<slug>-review.md` (or `-plan.md`).
**Cockpit (top):**
1. **Status line** — one line: how many items await review + last-updated timestamp.
diff --git a/.claude/skills/review-cockpit/references/session-review-template.md b/.claude/skills/review-cockpit/references/session-review-template.md
new file mode 100644
index 0000000..e9cd86c
--- /dev/null
+++ b/.claude/skills/review-cockpit/references/session-review-template.md
@@ -0,0 +1,37 @@
+# <session topic>: review
+
+> Working folder: `<path>` · Updated: YYYY-MM-DD HH:MM
+
+## Progress
+
+- [x] <done item>
+- [ ] **Item 1** <what needs your call>
+- [ ] **Item 2** <what needs your call>
+
+## Context
+
+What you need to know to judge the items below: what changed, what was deliberately held back, what the constraint was.
+
+---
+
+## Item 1: <name>
+
+**Status:** needs review | approved | changes requested
+
+<the artifact, or a link to it, plus whatever the reader needs to judge it>
+
+### Your call
+
+- [ ] Approve as-is
+- [ ] Change: <write it here or say it in chat>
+
+---
+
+## Item 2: <name>
+
+**Status:** needs review
+
+### Your call
+
+- [ ] Approve as-is
+- [ ] Change:
diff --git a/.claude/skills/ultragoal/SKILL.md b/.claude/skills/ultragoal/SKILL.md
index 4f4c8ac..89eeb7e 100644
--- a/.claude/skills/ultragoal/SKILL.md
+++ b/.claude/skills/ultragoal/SKILL.md
@@ -4,12 +4,14 @@ description: >
Run a large, multi-session goal (e.g. shipping a whole side product) through the full V-model
closed loop, one phase at a time, with cross-session state and a final
north-star acceptance gate. Ultragoals never downgrade the lane: every
- phase runs CP-1→CP-6 with adversarial verification. Use via /ultragoal.
+ phase runs CP-1→CP-6 with adversarial verification. Opt-in: invoke with
+ /ultragoal or by calling something a long-running goal. Ordinary work does not
+ run this.
---
# Ultragoal — the closed loop for goals too big to ship in one run
-An **ultragoal** is the term for a north-star that spans many sessions: fork the repos, combine the strong parts, ship one product. A normal `/execute` run is one task through the loop. An ultragoal is a *chain of phases*, each of which is its own full closed-loop run, tracked so any cold session can resume.
+An **ultragoal** is the term for a north-star that spans many sessions: fork the repos, combine the strong parts, ship one product. A `/closed-loop` run is one task through the loop. An ultragoal is a *chain of phases*, each of which is its own full closed-loop run, tracked so any cold session can resume.
**Core rule (from dwarves-kit, adopted fully): the worker never grades its own homework, and wrongness compounds across sessions — so verify every phase, not just the end.**
@@ -20,7 +22,7 @@ An **ultragoal** is the term for a north-star that spans many sessions: fork the
- `/ultragoal status` — report all ultragoals from the registry
- Trigger phrases: "make this an ultragoal", "this is a long-running goal", "combine these into one product over time"
-Do **not** use for single-run work — that is `/execute`. Rule of thumb: if it needs a phase decomposition and won't finish today, it is an ultragoal.
+Do **not** use for single-run work (that is `/closed-loop`), and do not start one on a request that never asked for one. Rule of thumb: if it needs a phase decomposition and won't finish today, it is an ultragoal.
## Files (one goal = one folder)
@@ -36,7 +38,7 @@ Code, if any, lives outside the vault (e.g. `~/code/<goal>/`) — the spec/statu
## Phase 0 — Charter (`/ultragoal new`)
1. Interview the user for the **north-star** in one sentence (what "done" looks like).
-2. Write `04-projects/<goal>/spec.md` from `04-projects/harness/templates/SPEC-template.md`:
+2. Write `04-projects/<goal>/spec.md` from `../closed-loop/references/spec-template.md`:
- North-star statement
- Falsifiable `AC-n` acceptance criteria (these define done for the *whole* goal)
- Phase decomposition `P0…Pn` — each phase is a shippable increment mapped to a subset of `AC-n`
@@ -97,7 +99,7 @@ North-star acceptance:
Every ultragoal carries a single self-contained HTML report that **covers everything**: north-star, live status, all phases, the full `AC-n` traceability table with pass/open/fail, evidence rows per phase (with screenshots embedded as `data:` URIs), and the open-items / next-action block. It is the human-readable face of the evidence ledger.
-- **Template:** `04-projects/harness/templates/report.html` (house style, theme-aware, rows-not-cards). Copy it, then fill every `{{token}}` and `<!-- FILL -->` / `<!-- REPEAT -->` block from `spec.md` + `STATUS.md` + `evidence/P*/`. You fill it by editing — do not build a parser.
+- **Template:** `../closed-loop/references/report-template.html` (theme-aware, rows-not-cards). Copy it, then fill every `{{token}}` and `<!-- FILL -->` / `<!-- REPEAT -->` block from `spec.md` + `STATUS.md` + `evidence/P*/`. You fill it by editing; do not build a parser.
- **Deliverable path:** `04-projects/<goal>/report.html`. One file, overwritten each phase (it always reflects current truth).
- **When:** regenerate at each phase gate (CP-6) and again at final north-star acceptance. The final report must show every `AC-n` with a PASS row — if any pill is `open`, the goal is not done.
- **Self-contained only:** inline everything, embed screenshots as `data:` URIs, so it also works when published as an Artifact (external hosts are blocked). Before publishing as an Artifact, load the `artifact-design` skill.
diff --git a/.cursor-plugin/plugin.json b/.cursor-plugin/plugin.json
index 3032db6..5c13d6c 100644
--- a/.cursor-plugin/plugin.json
+++ b/.cursor-plugin/plugin.json
@@ -1,7 +1,7 @@
{
"name": "cog-second-brain",
"displayName": "COG: Agentic Second Brain",
- "version": "3.11.0",
+ "version": "3.12.0",
"description": "Self-evolving second brain powered by AI agents, Obsidian markdown, and Git. 33 skills, 10 agents (6 workers + 4 read-only verifiers), people CRM with tiered enrichment. Closed-loop harness release: V-model verification lifecycle with criterion-traced evidence, risk lanes, multi-session ultragoals with a north-star acceptance gate, craft skills for writing and visuals, and paired anti-slop design skills for marketing and product UI. Works with Claude Code, Cursor, Kiro, Gemini CLI, and Codex.",
"author": {
"name": "Huy Tieu",
diff --git a/AGENTS.md b/AGENTS.md
index 9c00ff5..b507a16 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -446,7 +446,9 @@ This document defines the available commands/skills for AI agents interacting wi
### Verification Harness Skills
-The following 5 skills implement the V-model closed loop described in `WORKFLOW.md`. They are the enforcement layer that keeps every other skill honest: the worker never grades its own homework.
+The following 5 skills implement the V-model closed loop described in `WORKFLOW.md`: the worker never grades its own homework.
+
+**They are opt-in.** None of them run unless you invoke the skill, ask for the closed loop / proper verification / an evidence trail in those words, or set `verification_harness: on` in `00-inbox/MY-PROFILE.md`. Ordinary work (notes, briefs, research, drafts, edits) carries no checkpoints, no lane classification, and no evidence ledger. `WORKFLOW.md` governs harness runs and nothing else.
---
@@ -455,10 +457,12 @@ The following 5 skills implement the V-model closed loop described in `WORKFLOW.
**Description:** V-model execute pipeline: CP-2 plan → CP-3 build → CP-3v component verify → CP-4 integration verify → CP-5 acceptance. Every verify step emits evidence rows traced to acceptance criterion IDs (`AC-n`).
**Triggers:**
-- `/execute <task>`
-- "run this through the closed loop"
-- Any `normal`-lane or higher build task
-- Automatically, whenever a task mutates external state
+- `/closed-loop <task>` or `/closed-loop <spec-path>`
+- "run this through the closed loop", "verify this properly", "give me an evidence trail"
+- `verification_harness: on` in your profile, on a build task
+- Another skill declaring a `normal`+ lane reaching its verify step
+
+Not triggered by an ordinary request. A task that mutates external state still owes the post-condition check (observe the artifact) whether or not the full loop runs.
**Purpose:** Prevent the **confident-but-unchecked** failure mode. A worker reports success, nothing downstream validates it, and a plausible-but-wrong result ships. The loop makes success a *verified observation*, not a claim.
@@ -480,9 +484,10 @@ The following 5 skills implement the V-model closed loop described in `WORKFLOW.
**Triggers:**
- `/ultragoal <goal>`
-- "this is a multi-week thing"
+- "make this an ultragoal", "this is a multi-week thing"
- Resuming a long-running goal in a cold session
-- Any goal that can't finish in one run
+
+Never started unprompted. A big task is not an ultragoal until you call it one.
**Purpose:** Wrongness compounds across sessions. A goal spanning weeks accumulates unverified assumptions that no single run ever revisits. Ultragoals never downgrade the lane: every phase runs CP-1 → CP-6 with adversarial verification, and nothing is "done" until every `AC-n` has a PASS row.
@@ -504,9 +509,10 @@ The following 5 skills implement the V-model closed loop described in `WORKFLOW.
**Triggers:**
- `/harvest`
-- SessionEnd hook (automatic staging)
-- After a correction, a rejected deliverable, or a surprising discovery
-- Nightly self-enhancement runs
+- After a correction, a rejected deliverable, or a surprising discovery, when you ask for it
+- A scheduled self-enhancement job, if you set one up
+
+COG ships no hooks, so nothing stages automatically until you wire it up yourself.
**Purpose:** Tacit knowledge dies in the transcript. Corrections you made, workarounds discovered, and patterns that worked are all lost when the session closes. Harvest catches them at the boundary, but stages rather than commits, because auto-promoting session noise into durable knowledge poisons the well.
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 0cda42a..fc4d41e 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -2,6 +2,30 @@
All notable changes to COG (Cognition + Obsidian + Git) will be documented in this file.
+## [3.12.0] - 2026-08-25
+
+### Changed
+
+#### The verification harness is opt-in, not mandated
+Reported in [#43](https://github.com/huytieu/COG-second-brain/issues/43): `CLAUDE.md` carried four **ALWAYS APPLY** sections mandating the V-model harness on every task, while `WORKFLOW.md` opened by asserting that "every non-`tiny` task walks the V." Nothing in a normal COG session does that, and a mandate no session honors teaches the model to discount every other rule in the file. The harness is good machinery for the runs that want it and pure ceremony on a braindump.
+
+- `CLAUDE.md`: the four sections (V-Model Checkpoints, Closed-Loop Execute, Risk Lanes, Ultragoal) collapse into one **Verification Harness (opt-in, off by default)**. It names the three ways to turn the harness on and states plainly that a session which never invoked it owes no checkpoints, no lane classification, and no evidence ledger. Two rules survive as always-on because they cost nothing: verification observes the artifact, and a worker never grades its own homework.
+- `WORKFLOW.md`: new **Scope: opt-in only** section at the top. The V-model text now reads "inside a harness run, every non-`tiny` task walks the V."
+- Third opt-in path: `verification_harness: on` in `00-inbox/MY-PROFILE.md` makes the `normal`-lane pipeline the default for build tasks. Absent or `off` means per-request. Added to the onboarding profile template.
+- `AGENTS.md`, `README.md`: the harness section leads with opt-in; per-skill trigger lists no longer claim the loop fires "automatically, whenever a task mutates external state".
+
+### Fixed
+
+#### Harness references that pointed at files COG never shipped
+Also from [#43](https://github.com/huytieu/COG-second-brain/issues/43): the skills instructed the model to copy templates from `04-projects/harness/templates/`, a directory that does not exist in a COG checkout, and `WORKFLOW.md` closed with an **Install** block calling `.claude/lib/install-harness.sh`, a script that does not exist either. `checkpoint.sh init` would have failed on the same missing template path.
+
+- Templates now ship with the skills, so `/update-cog` keeps them current: `closed-loop/references/spec-template.md`, `closed-loop/references/report-template.html` (self-contained, theme-aware), `retro/references/retro-template.md`, `review-cockpit/references/session-review-template.md`. All four added to the updater's framework file list.
+- `checkpoint.sh init` writes the evidence-ledger header inline instead of copying a missing file.
+- The Install block is replaced by **No install step**: the two `.claude/lib` scripts ship executable, run directories are created on demand, and COG ships no hooks. The `harvest` nightly block no longer calls the phantom installer, and the `harvest` trigger list no longer claims a SessionEnd hook stages automatically.
+- `/execute` never existed as a command; every reference now points at `/closed-loop`. `WORKFLOW.md`'s domain-routing and self-enhancement tables listed skills COG does not ship (`dogfood-release`, `aut-skill-capture`, gstack, lizard) and now list the ones it does.
+- Vault-specific leakage removed from the harness surface: the `browser-harness` Python helpers in `closed-loop`, and the `Agent(model="fable")` call shape in `WORKFLOW.md`.
+- `CLAUDE.md` said `.claude/agents/` holds 6 agents; it holds 10, and `.claude/lib/` was missing from the framework file list.
+
## [3.11.0] - 2026-08-24
### Added
diff --git a/CLAUDE.md b/CLAUDE.md
index fa81d2d..25b6f02 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -1,6 +1,6 @@
# COG Second Brain — Framework Instructions
-> Lifecycle: `WORKFLOW.md` (V-model harness) · Universal surface: `AGENTS.md`
+> Opt-in verification harness: `WORKFLOW.md` · Universal surface: `AGENTS.md`
## Response Style (ALWAYS APPLY, every agent, every deliverable)
@@ -17,68 +17,22 @@ Optimize for **information gain, not apparent completeness**. The failure mode i
Applies to chat answers, reports, briefs, specs, docs, and all subagent outputs.
-## V-Model Checkpoints (ALWAYS APPLY)
+## Verification Harness (opt-in, off by default)
-Work walks the **V**: decompose left (spec → plan), build at the apex, verify right with **evidence** traced to criterion IDs (`AC-n`). See `WORKFLOW.md` for the full diagram.
+COG ships a V-model verification harness: decompose left (spec then plan), build at the apex, verify right with evidence traced to criterion IDs (`AC-n`). **It does not run unless you ask for it.** Ordinary work (notes, briefs, research, drafts, edits) takes no checkpoints, no lane classification, and no evidence ledger.
-| CP | Phase | Blocking when |
-|---|---|---|
-| CP-1 | Spec + traceability matrix | `normal`+ |
-| CP-2 | Plan (tasks ↔ AC-n) | `normal`+ |
-| CP-3 | Build | always |
-| CP-3v | Component verify (`task-verifier`) | `normal`+ |
-| CP-4 | Integration verify (`integration-verifier`) | `full`+, multi-task |
-| CP-5 | Acceptance (post-condition artifacts) | mutations |
-| CP-6 | Ship (Review Gate / user approval) | external |
-| CP-7 | Retro (`/retro`) + harvest | advisory, expected |
+Three ways to turn it on:
-**Two-way verification:** every `AC-n` has a task on the left and ≥1 `EVIDENCE AC-n | CP-* | PASS | …` row on the right before ship. Record checkpoints with `bash .claude/lib/checkpoint.sh record <run-dir> <CP> PASS|FAIL "<note>"`. Evidence bundles live in `04-projects/harness/runs/<id>/evidence/ledger.md`.
+- **By skill.** `/closed-loop`, `/ultragoal`, `/retro`, `/harvest`, `/review-cockpit`.
+- **By phrasing.** "Run this through the closed loop", "verify this properly", "track this as an ultragoal", "give me an evidence trail".
+- **By profile.** Set `verification_harness: on` in `00-inbox/MY-PROFILE.md` frontmatter to make the `normal`-lane pipeline the default for build tasks. Absent or `off` means opt-in per request.
-**Cross-model flagship gate (`full`+ / ultragoal / irreversible):** overlay a flagship model that is NOT the lead's own as a cross-model advisor at CP-1 and critic at CP-4, CP-5, CP-6. Fresh context, read-only, advisory. Same-family verifiers share the lead's blind spots; a different model catches a different error class. A critical cross-model finding blocks auto-ship. Skip on `tiny`/`normal`.
+Inside a harness run: checkpoints, gate classes, risk lanes, and file homes are in `WORKFLOW.md`; the build-verify-fix pipeline is in `.claude/skills/closed-loop/SKILL.md`; multi-session goals are in `.claude/skills/ultragoal/SKILL.md`. Those documents oblige nothing in a session that never invoked them.
-## Closed-Loop Execute (ALWAYS APPLY, build tasks)
+Two of the harness's rules are worth applying whether or not it is on, because they cost nothing:
-Every **build task** (`normal` lane and above) runs the verification pipeline before reporting success. Skill: `.claude/skills/closed-loop/SKILL.md`. Inspired by [dwarves-kit](https://github.com/dwarvesf/dwarves-kit).
-
-```
-worker → task-verifier (read-only) → fix-agent (max 2 retries) → escalate to the user
-```
-
-| Lane | Ceremony | Classifier |
-|---|---|---|
-| `tiny` | worker + post-condition if mutating | `bash .claude/lib/lane-classify.sh classify "<task>"` |
-| `normal` | full pipeline above | default |
-| `full` | pipeline + claim-verifier for auditable claims | briefs, publishes |
-| `bug` | evidence ledger before fix; 3-fix wall | regressions |
-| `backfill` | audit/plan only until approved | doc retrofits |
-
-**Rules:**
-- `task-verifier` and `fix-agent` are **fresh-context** subagents (paths only, never paste worker output into a verifier).
-- Verifier agents are **read-only**. They cannot edit files or mutate external state.
-- **Dispatch a verifier subagent when there is external state, or when a fresh pair of eyes is the point**: mutations (publishes, pushes, tracker transitions, deploys), auditable claims, `full`+ lanes, every ultragoal phase.
-- **On `normal`-lane read-only work, the lead verifies inline instead.** Spawning a verifier to re-read a file the lead just wrote buys nothing.
-- Verification means **observing the artifact** (curl the URL, screenshot the page, re-fetch the issue, diff the file), never re-reading the worker's own summary of it.
-- Max **2** fix-agent retries per task, then `ESCALATE` with evidence. Log outcomes to `.claude/logs/loop-ledger.tsv`.
-
-## Risk Lanes (ALWAYS APPLY)
-
-Classify before executing: `bash .claude/lib/lane-classify.sh explain "<task>"`.
-
-- **tiny**: skip verifier unless external mutation.
-- **normal**: closed-loop + single-file deliverable. Verifier subagent only when the task mutates external state.
-- **full**: closed-loop + adversarial claim-verifier + Review Gate for external.
-- **bug**: root cause before patch.
-- **backfill**: no behavior change without approval.
-
-## Ultragoal (ALWAYS APPLY, long-running goals)
-
-An **ultragoal** is a large, multi-session goal that can't ship in one run. It gets the closed loop applied in full and is never lane-downgraded. Skill: `.claude/skills/ultragoal/SKILL.md`.
-
-- **One spec, north-star + `AC-n`.** `04-projects/<goal>/spec.md` holds the north-star, falsifiable acceptance criteria, and a phase decomposition (`P0…Pn`).
-- **Every phase is a full closed-loop run** (CP-1 → CP-6). No `tiny` shortcut inside an ultragoal.
-- **Living status ledger.** `04-projects/<goal>/STATUS.md` tracks phase state, current phase, open `AC-n`, and next action so any session resumes cold.
-- **Two acceptance gates.** Per-phase acceptance (CP-5) AND a final **north-star acceptance verifier** that checks every `AC-n` has a PASS row.
-- **Adversarial by default.** Wrongness compounds across sessions, so verify every phase, not only the end.
+- **Verification means observing the artifact**: curl the URL, screenshot the page, re-fetch the issue, diff the file. Never re-read a worker's own summary of it. Mandatory for external mutations; see § Skill Post-Condition Rule.
+- **A worker never grades its own homework.** When a fresh pair of eyes is the point (external mutations, auditable claims), the verifier is a separate read-only subagent that receives paths and criteria, never the worker's output.
## Visual Verification (ALWAYS APPLY, UI/UX tasks)
@@ -87,7 +41,7 @@ Any task that **implements or changes a UI/UX flow** is not verified by a DOM/se
- **Capture visual evidence.** Screenshot every meaningful state; record multi-step flows.
- **Actually read the image, then compare** against the intended design (mock, spec wireframe, prior state, house style). Name the discrepancy; never declare pass on "element exists."
- **Fix the UI error you spot**. This is part of the task, not a follow-up. Re-capture to prove it.
-- Evidence lands in the run evidence dir and feeds the CP-5 post-condition row.
+- Keep the screenshots next to the deliverable. Inside a harness run they are the CP-5 acceptance evidence.
## Delegation Cap (ALWAYS APPLY)
@@ -111,8 +65,8 @@ When spawning subagents, use the correct model for the task:
| File operations (vault reads/writes, metadata, profiles) | **Sonnet** | `worker-file-ops` |
| Pre-approved mutations (Jira transitions, Linear updates, API calls) | **Sonnet** | `worker-executor` |
| People profile updates from brief/meeting data | **Sonnet** | `brief-people-updater` |
-| Read-only verification (acceptance criteria, post-conditions) | **Sonnet** | `task-verifier` |
-| Cross-task integration verify (CP-4) | **Sonnet** | `integration-verifier` |
+| Read-only verification, harness runs only (acceptance criteria, post-conditions) | **Sonnet** | `task-verifier` |
+| Cross-task integration verify, harness runs only (CP-4) | **Sonnet** | `integration-verifier` |
| Targeted fixes after verifier FAIL:fixable | **Sonnet** | `fix-agent` |
| Harvest staging curation (propose-only) | **Sonnet** | `harvest-curator` |
| Reasoning, synthesis, cross-referencing, writing | **Opus** | Lead session (no delegation) |
@@ -272,8 +226,9 @@ Role packs live in `.claude/roles/`. New roles can be added by dropping a file f
### Framework files (updated via `cog-update.sh` or `/update-cog`)
- `.claude/skills/`: Claude Code skills (33 skills)
-- `.claude/agents/` — Worker agent definitions (6 agents)
+- `.claude/agents/` - Worker and verifier agent definitions (10 agents)
- `.claude/roles/` — Role packs for personalized recommendations
+- `.claude/lib/` - Harness helper scripts (`checkpoint.sh`, `lane-classify.sh`)
- `.kiro/powers/` — Kiro powers
- `.gemini/commands/` — Gemini CLI commands
- `AGENTS.md` — Universal agent documentation
diff --git a/COG-VERSION b/COG-VERSION
index afad818..92536a9 100644
--- a/COG-VERSION
+++ b/COG-VERSION
@@ -1 +1 @@
-3.11.0
+3.12.0
diff --git a/README.md b/README.md
index cd9813e..aa02086 100644
--- a/README.md
+++ b/README.md
@@ -115,7 +115,7 @@ Before publishing or updating framework files, run `./scripts/validate-agent-sur
### Verification Harness
-The enforcement layer that keeps every other skill honest. Work walks a **V**: decompose left into falsifiable criteria, build at the apex, verify right with evidence traced back to each criterion. Full lifecycle in [WORKFLOW.md](WORKFLOW.md).
+Opt-in. Ask for one of these and work walks a **V**: decompose left into falsifiable criteria, build at the apex, verify right with evidence traced back to each criterion. Say nothing and none of it runs: ordinary work carries no checkpoints and no evidence ledger. Turn it on per request, or by default for build tasks with `verification_harness: on` in `00-inbox/MY-PROFILE.md`. Full lifecycle in [WORKFLOW.md](WORKFLOW.md).
| Skill | What it does | Try saying... |
|---|---|---|
diff --git a/WORKFLOW.md b/WORKFLOW.md
index b06afb7..7040eb2 100644
--- a/WORKFLOW.md
+++ b/WORKFLOW.md
@@ -3,9 +3,17 @@
> **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](https://github.com/dwarvesf/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)
-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.
+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)
@@ -51,7 +59,7 @@ No criterion ships without a matching evidence row. No evidence row without a cr
| **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>`
+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)
@@ -70,7 +78,7 @@ Verifier and post-condition steps emit these rows. Consolidate in `evidence/ledg
### Cross-model flagship gate (`full`+ / ultragoal / irreversible)
-The read-only verifiers (`task-verifier` CP-3v, `integration-verifier` CP-4) are **same-family mechanical** checks — they share the lead's blind spots. On high-stakes runs, overlay a **flagship model that is NOT the lead's own** (spin off `Agent(model="fable")` when the lead is Opus) as a cross-model second opinion. Fresh context (paths + question, never the lead's draft-reasoning), read-only, advisory.
+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 |
|---|---|---|
@@ -80,7 +88,7 @@ The read-only verifiers (`task-verifier` CP-3v, `integration-verifier` CP-4) are
| **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 flagship hard-refuses (offensive-security / bio-adjacent), rerun the check on Opus 5 rather than dropping it.
+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)
@@ -97,7 +105,7 @@ Classifier: `bash .claude/lib/lane-classify.sh classify "<task>"`
## Verification pipeline (right arm detail)
```
-orchestrator (/execute or skill)
+orchestrator (/closed-loop or a skill that invokes it)
│
▼
CP-3 BUILD: worker implements (traced to AC-n)
@@ -115,7 +123,7 @@ orchestrator (/execute or skill)
CP-5: post-condition (observe artifact: curl, screenshot, re-fetch)
│
▼
- CP-6: ship gate (Review Gate: you approve external / gstack / deploy proof)
+ CP-6: ship gate (you approve anything external / deploy proof)
│
▼
CP-7: /retro + /harvest
@@ -125,53 +133,54 @@ orchestrator (/execute or skill)
| 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 |
-| Code / deploy | `/execute` + gstack | CP-3v + tests + CP-5 curl |
-| Product dogfood | `/dogfood-release` | Playwright cross-verify (CP-4) |
-| Content | content-factory | voice checklist + screenshot (CP-5) |
-| AUT skills | aut-skill-capture | verdict taxonomy (CP-3v) |
-| Knowledge | `/memory-hygiene` | environment re-verify (CP-5) |
+| 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 |
|---|---|---|
-| **V execute** | CP-3 → CP-5 | Evidence ledger + loop-ledger.tsv |
-| **Evolution** | CP-7 → aut-skill-capture | Skill patches from friction |
+| **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 → lizard |
+| **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 |
-| **Retro** | CP-7 | `04-projects/harness/retro/` |
## 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` |
-| Report template | `04-projects/harness/templates/report.html` |
-| Spec template | `04-projects/harness/templates/SPEC-template.md` |
| 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 |
|---|---|
-| `/execute` | CP-2 through CP-5 |
+| `/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 |
-## Install
+## No install step
-```bash
-bash .claude/lib/install-harness.sh
-bash .claude/lib/install-harness.sh --check
-```
+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.
diff --git a/cog-update.sh b/cog-update.sh
index 9f9ce62..0dcfd77 100755
--- a/cog-update.sh
+++ b/cog-update.sh
@@ -75,6 +75,8 @@ FRAMEWORK_FILES=(
".claude/skills/product-ui-taste/SKILL.md"
# Skill reference files (bundled lookup material, loaded on demand)
+ ".claude/skills/closed-loop/references/report-template.html"
+ ".claude/skills/closed-loop/references/spec-template.md"
".claude/skills/data-forms/references/forms.md"
".claude/skills/editorial-illustrations/references/design-system.md"
".claude/skills/editorial-illustrations/references/elements.md"
@@ -91,6 +93,8 @@ FRAMEWORK_FILES=(
".claude/skills/museum-art/references/smk.md"
".claude/skills/onboarding/references/profile-templates.md"
".claude/skills/onboarding/references/welcome-guide.md"
+ ".claude/skills/retro/references/retro-template.md"
+ ".claude/skills/review-cockpit/references/session-review-template.md"
".claude/skills/product-ui-taste/references/block-skeletons.md"
".claude/skills/product-ui-taste/references/canonical-sources.md"
".claude/skills/product-ui-taste/references/install-commands.md"
diff --git a/marketplace-entry.json b/marketplace-entry.json
index 39c4cc1..3b451d2 100644
--- a/marketplace-entry.json
+++ b/marketplace-entry.json
@@ -7,7 +7,7 @@
"url": "https://github.com/huytieu"
},
"description": "Self-evolving second brain with 33 AI skills, 10 agents, and people CRM. Closed-loop harness release: V-model verification lifecycle with criterion-traced evidence, read-only verifier agents, risk lanes, multi-session ultragoals with a north-star acceptance gate, craft skills for writing and visuals, and paired anti-slop design skills for marketing and product UI. Inspired by Garry Tan's gstack and gbrain, and dwarves-kit. Works with Claude Code, Antigravity, Cursor, Kiro, Gemini CLI, Codex.",
- "version": "3.11.0",
+ "version": "3.12.0",
"keywords": [
"second-brain",
"garry-tan",
diff --git a/plugin.json b/plugin.json
index a4fb6fc..33a885e 100644
--- a/plugin.json
+++ b/plugin.json
@@ -1,7 +1,7 @@
{
"$schema": "https://agent-plugins.org/schemas/1.0.0/plugin.schema.json",
"name": "cog-second-brain",
- "version": "3.11.0",
+ "version": "3.12.0",
"description": "Self-evolving second brain with 33 AI agent skills, 10 agents, and people CRM. Closed-loop harness: a V-model verification lifecycle (spec to evidence, criterion-traced checkpoints), read-only verifier agents that observe artifacts instead of trusting worker summaries, risk lanes, multi-session ultragoals with a north-star acceptance gate, plus craft skills for writing, visual work, and a paired anti-slop design system for marketing and product UI. Works with Claude Code, Antigravity, Cursor, Kiro, Gemini CLI, Codex, and any Agent Plugins conformant client.",
"author": {
"name": "Huy Tieu",
diff --git a/skills/closed-loop/SKILL.md b/skills/closed-loop/SKILL.md
index 2717dad..2ec9639 100644
--- a/skills/closed-loop/SKILL.md
+++ b/skills/closed-loop/SKILL.md
@@ -1,9 +1,11 @@
---
name: closed-loop
description: >
- V-model execute: CP-2 plan → CP-3 build → CP-3v component verify →
- CP-4 integration verify (full) → CP-5 acceptance. Worker never grades
- its own homework; evidence rows trace back to AC-n. Use via /execute.
+ Run one task through the V-model verification loop: CP-2 plan → CP-3 build →
+ CP-3v component verify → CP-4 integration verify (full lane) → CP-5 acceptance.
+ The worker never grades its own homework; evidence rows trace back to AC-n.
+ Opt-in: invoke with /closed-loop or by asking for the closed loop, proper
+ verification, or an evidence trail. Ordinary work does not run this.
---
# Closed-loop execute (V-model right arm)
@@ -12,9 +14,14 @@ Mechanical verification pipeline. Every verify step emits **evidence rows** tied
## When to use
-- `/execute <task>` or `/execute <spec-path>`
-- Any skill at `normal`+ lane after implementation
-- Code changes, deploys, multi-step vault deliverables
+The harness is opt-in. Run it when:
+
+- Invoked as `/closed-loop <task>` or `/closed-loop <spec-path>`.
+- The user asks for the closed loop, proper verification, or an evidence trail.
+- `verification_harness: on` in `00-inbox/MY-PROFILE.md` and this is a build task.
+- Another skill that declares a `normal`+ lane reaches its verify step.
+
+Do **not** run it on a request that did not ask for it. Notes, briefs, research, drafts, and ordinary edits are not harness runs, and a checkpoint ledger on those is pure overhead.
## Phase 0 — Lane + run folder
@@ -36,7 +43,7 @@ Record: `checkpoint.sh record <run-dir> CP-0 PASS|SKIP "<lane>"`
If spec exists, use its `## Acceptance criteria` + traceability matrix. Else write:
-`04-projects/harness/runs/<id>/criteria.md` using `04-projects/harness/templates/SPEC-template.md` (criteria + matrix only).
+`04-projects/harness/runs/<id>/criteria.md` using `references/spec-template.md` (criteria + matrix sections only).
Each criterion: **falsifiable** + `AC-n` ID + verify method.
@@ -83,7 +90,7 @@ For each mutation, observe artifact (curl, screenshot, re-fetch). Emit:
`EVIDENCE AC-n | CP-5 | PASS | <observation> | <artifact>`
-**UI/UX flow changes:** the post-condition is *visual*. Capture with browser-harness (`evidence_shot` per state; `FlowRecorder`→`.save_gif()` for a flow; `pixel_diff` against the intended/prior state), then read the image and confirm no overflow/misalignment/clipping/wrong-color/broken-responsive before PASS. The Observation must describe what you saw; the artifact is the screenshot/GIF in `evidence/`. Fix any visual defect and re-capture. See CLAUDE.md → Visual Verification.
+**UI/UX flow changes:** the post-condition is *visual*. Screenshot every meaningful state with whatever browser tooling the environment has, then read the image and confirm no overflow, misalignment, clipping, wrong color, or broken responsive layout before PASS. The Observation must describe what you saw; the artifact is the screenshot/GIF in `evidence/`. Fix any visual defect and re-capture. See CLAUDE.md → Visual Verification.
Write `evidence/CP-5-acceptance.md`. **Traceability closure**: every AC in matrix has ≥1 PASS row in ledger.
@@ -93,18 +100,18 @@ Record: `checkpoint.sh record <run-dir> CP-5 PASS|FAIL`
- Append to `.claude/logs/loop-ledger.tsv`
- Update spec traceability matrix statuses to `verified`
-- **`full` lane / big task:** generate an HTML rollup from `04-projects/harness/templates/report.html` → `04-projects/harness/runs/<id>/report.html`, filled from `criteria.md` + `evidence/ledger.md` (criteria, AC traceability, verifier verdicts, post-condition observations). Self-contained; `SendUserFile` it or publish as an Artifact. Skip for `normal`/`tiny`.
+- **`full` lane / big task:** generate an HTML rollup from `references/report-template.html` → `04-projects/harness/runs/<id>/report.html`, filled from `criteria.md` + `evidence/ledger.md` (criteria, AC traceability, verifier verdicts, post-condition observations). Self-contained; `SendUserFile` it or publish as an Artifact. Skip for `normal`/`tiny`.
- Suggest `/retro <run-dir>` for CP-7
## Integration
| Skill | Lane | CP-4 |
|---|---|---|
-| ultragoal | `full` per phase (never downgraded) | integration-verifier + north-star acceptance |
-| team-brief | full | claim-verifier |
-| dogfood-release | full | Playwright cross-verify |
-| blog-publish | normal | skip |
-| content-factory | normal | skip |
+| `ultragoal` | `full` per phase (never downgraded) | integration-verifier + north-star acceptance |
+| `team-brief` | full | claim-verifier |
+| `comprehensive-analysis`, `auto-research` | full | claim-verifier on cited claims |
+| `content-factory` | normal | skip |
+| `review-cockpit` | normal | skip; CP-6 is the user's approval per card |
## Escalation template
diff --git a/skills/closed-loop/references/report-template.html b/skills/closed-loop/references/report-template.html
new file mode 100644
index 0000000..d6356bc
--- /dev/null
+++ b/skills/closed-loop/references/report-template.html
@@ -0,0 +1,125 @@
+<!doctype html>
+<!--
+ COG harness run report: fill by editing, do not build a parser.
+ Replace every {{token}}. Delete or repeat each <!-- REPEAT --> row block.
+ Self-contained: no external CSS, JS, fonts, or images. Inline any screenshot
+ as a data: URI, or link it relative to this file inside the run folder.
+-->
+<html lang="en">
+<head>
+<meta charset="utf-8">
+<meta name="viewport" content="width=device-width, initial-scale=1">
+<title>{{title}}: run report</title>
+<style>
+ :root {
+ --bg: #fbfbfa; --panel: #ffffff; --ink: #1a1a19; --muted: #6b6b68;
+ --line: #e4e4e1; --pass: #1a7f4b; --fail: #b4232a; --skip: #8a8a86;
+ --accent: #1a1a19;
+ --mono: ui-monospace, SFMono-Regular, "SF Mono", Menlo, monospace;
+ --sans: -apple-system, BlinkMacSystemFont, "Segoe UI", Inter, system-ui, sans-serif;
+ }
+ @media (prefers-color-scheme: dark) {
+ :root:not([data-theme="light"]) {
+ --bg: #131313; --panel: #1b1b1a; --ink: #ececea; --muted: #9a9a96;
+ --line: #2c2c2a; --pass: #4ec281; --fail: #f2777c; --skip: #7d7d79;
+ --accent: #ececea;
+ }
+ }
+ :root[data-theme="dark"] {
+ --bg: #131313; --panel: #1b1b1a; --ink: #ececea; --muted: #9a9a96;
+ --line: #2c2c2a; --pass: #4ec281; --fail: #f2777c; --skip: #7d7d79;
+ --accent: #ececea;
+ }
+ * { box-sizing: border-box; }
+ body { margin: 0; background: var(--bg); color: var(--ink); font: 15px/1.55 var(--sans); }
+ .wrap { max-width: 62rem; margin: 0 auto; padding: 3rem 1.5rem 6rem; }
+ header { border-bottom: 1px solid var(--line); padding-bottom: 1.5rem; margin-bottom: 2.5rem; }
+ h1 { font-size: 1.6rem; letter-spacing: -0.01em; margin: 0 0 .4rem; }
+ .meta { color: var(--muted); font: 13px/1.5 var(--mono); }
+ h2 { font-size: 1rem; text-transform: uppercase; letter-spacing: .08em; color: var(--muted);
+ margin: 2.75rem 0 .9rem; font-weight: 600; }
+ p { margin: 0 0 1rem; }
+ .northstar { background: var(--panel); border: 1px solid var(--line); border-left: 3px solid var(--accent);
+ padding: 1rem 1.15rem; border-radius: 3px; }
+ .scroll { overflow-x: auto; border: 1px solid var(--line); border-radius: 3px; background: var(--panel); }
+ table { width: 100%; border-collapse: collapse; font-size: 14px; }
+ th, td { text-align: left; padding: .6rem .8rem; border-bottom: 1px solid var(--line); vertical-align: top; }
+ th { font: 600 12px/1.4 var(--mono); text-transform: uppercase; letter-spacing: .05em; color: var(--muted); }
+ tr:last-child td { border-bottom: 0; }
+ td.id, td.cp { font-family: var(--mono); white-space: nowrap; }
+ .v { font: 600 12px/1.4 var(--mono); white-space: nowrap; }
+ .v.pass { color: var(--pass); } .v.fail { color: var(--fail); } .v.skip { color: var(--skip); }
+ .obs { font-family: var(--mono); font-size: 12.5px; color: var(--muted); }
+ figure { margin: 1.25rem 0; }
+ figure img { display: block; max-width: 100%; border: 1px solid var(--line); border-radius: 3px; }
+ figcaption { color: var(--muted); font: 12.5px/1.5 var(--mono); margin-top: .5rem; }
+ .next { background: var(--panel); border: 1px solid var(--line); padding: 1rem 1.15rem; border-radius: 3px; }
+ footer { margin-top: 3.5rem; padding-top: 1.25rem; border-top: 1px solid var(--line);
+ color: var(--muted); font: 12.5px/1.5 var(--mono); }
+</style>
+</head>
+<body>
+<div class="wrap">
+
+<header>
+ <h1>{{title}}</h1>
+ <div class="meta">{{run_id}} · lane {{lane}} · {{date}} · outcome {{shipped|escalated}}</div>
+</header>
+
+<h2>North star</h2>
+<p class="northstar">{{one paragraph: what is true now that was not true before}}</p>
+
+<h2>Phases</h2>
+<div class="scroll">
+<table>
+ <thead><tr><th>Phase</th><th>Scope</th><th>Covers</th><th>State</th></tr></thead>
+ <tbody>
+ <!-- REPEAT one row per phase; delete this section entirely for a single-task run -->
+ <tr><td class="id">{{P0}}</td><td>{{scope}}</td><td class="id">{{AC-01, AC-02}}</td>
+ <td class="v pass">{{done}}</td></tr>
+ </tbody>
+</table>
+</div>
+
+<h2>Acceptance criteria</h2>
+<div class="scroll">
+<table>
+ <thead><tr><th>AC</th><th>Criterion</th><th>Task</th><th>Verdict</th></tr></thead>
+ <tbody>
+ <!-- REPEAT one row per AC in the traceability matrix -->
+ <tr><td class="id">{{AC-01}}</td><td>{{criterion}}</td><td class="id">{{T-01}}</td>
+ <td class="v pass">{{PASS}}</td></tr>
+ </tbody>
+</table>
+</div>
+
+<h2>Evidence</h2>
+<div class="scroll">
+<table>
+ <thead><tr><th>AC</th><th>CP</th><th>Verdict</th><th>Observation</th><th>Artifact</th></tr></thead>
+ <tbody>
+ <!-- REPEAT one row per EVIDENCE line in evidence/ledger.md -->
+ <tr><td class="id">{{AC-01}}</td><td class="cp">{{CP-3v}}</td><td class="v pass">{{PASS}}</td>
+ <td class="obs">{{what was observed, not what was claimed}}</td>
+ <td class="obs">{{path or command}}</td></tr>
+ </tbody>
+</table>
+</div>
+
+<h2>Screenshots</h2>
+<!-- REPEAT one figure per captured state; delete the section if the run had no visual surface -->
+<figure>
+ <img src="{{evidence/shot-01.png}}" alt="{{what this shows}}">
+ <figcaption>{{state name}}: {{what to look at, and what was fixed if anything}}</figcaption>
+</figure>
+
+<h2>Next action</h2>
+<p class="next">{{the single next thing, or "none: run closed"}}</p>
+
+<footer>
+ Generated by the COG verification harness · evidence bundle {{04-projects/harness/runs/&lt;id&gt;/evidence/}}
+</footer>
+
+</div>
+</body>
+</html>
diff --git a/skills/closed-loop/references/spec-template.md b/skills/closed-loop/references/spec-template.md
new file mode 100644
index 0000000..3ce5684
--- /dev/null
+++ b/skills/closed-loop/references/spec-template.md
@@ -0,0 +1,47 @@
+# SPEC-NNN: <title>
+
+> Lane: `tiny|normal|full|bug|backfill` · Run: `04-projects/harness/runs/<id>` · Date: YYYY-MM-DD
+
+## Goal
+
+One paragraph. What is true when this is done that is not true now.
+
+## Non-goals
+
+What this deliberately does not cover, so scope creep is visible.
+
+## Acceptance criteria
+
+Every criterion is **falsifiable**: a named observation either happens or it does not.
+"Works well" is not a criterion. "`curl -s <url> | grep -q '<string>'` exits 0" is.
+
+| ID | Criterion | Verify method |
+|---|---|---|
+| AC-01 | <observable statement> | <command, screenshot, re-fetch, diff> |
+| AC-02 | | |
+
+## Traceability matrix
+
+Status: `pending` at CP-2, `built` at CP-3, `verified` once a PASS row exists in the ledger.
+
+| AC | Task | Evidence row | Status |
+|---|---|---|---|
+| AC-01 | T-01 | | pending |
+| AC-02 | T-02 | | pending |
+
+## Tasks
+
+| ID | Task | Covers |
+|---|---|---|
+| T-01 | <what gets built> | AC-01 |
+| T-02 | | AC-02 |
+
+## Phases (ultragoal only)
+
+| Phase | Scope | Covers | State |
+|---|---|---|---|
+| P0 | | AC-01 | not started |
+
+## Risks and open questions
+
+- <risk>: mitigation or the decision needed.
diff --git a/skills/harvest/SKILL.md b/skills/harvest/SKILL.md
index 8ab5811..6fa9c75 100644
--- a/skills/harvest/SKILL.md
+++ b/skills/harvest/SKILL.md
@@ -64,12 +64,13 @@ Harvest: <n> staged, <m> promoted, <k> folded
- Contradictions with existing knowledge → flag both, ask the user.
- Pair with `/memory-hygiene` monthly for environment-dependent facts in `~/.claude/.../memory/`.
-## Nightly enhance (optional launchd)
-
-```bash
-# 04-projects/harness/scripts/nightly-enhance.sh
-bash .claude/lib/install-harness.sh --check # hooks alive
-# /memory-hygiene (skill)
-# /harvest promote if staging non-empty
-# content-factory if scheduled
+## Nightly enhance (optional)
+
+If you want harvest to run unattended, schedule a job that invokes the skills in
+this order. There is nothing to install first, since COG ships no hooks.
+
+```
+/memory-hygiene
+/harvest promote # only if the staging file is non-empty
+/content-factory # only if scheduled
```
diff --git a/skills/onboarding/references/profile-templates.md b/skills/onboarding/references/profile-templates.md
index dee6664..9d3c2cc 100644
--- a/skills/onboarding/references/profile-templates.md
+++ b/skills/onboarding/references/profile-templates.md
@@ -11,6 +11,7 @@ created: YYYY-MM-DD
onboarding_completed: true
role_pack: [matched role_id or "custom"]
agent_mode: [solo or team, based on role pack suggestion]
+verification_harness: off
tags: ["#profile", "#config", "#cog"]
---
@@ -24,6 +25,7 @@ tags: ["#profile", "#config", "#cog"]
## Settings
- **Agent Mode**: [solo/team] *(solo = handle everything directly; team = delegate to specialist sub-agents for deeper results)*
+- **Verification Harness**: off *(off = the V-model harness runs only when you ask for it, e.g. `/closed-loop`; on = build tasks default to the `normal`-lane pipeline)*
## Active Projects
[If they mentioned projects:]
diff --git a/skills/retro/SKILL.md b/skills/retro/SKILL.md
index 0a1dcfd..18aad21 100644
--- a/skills/retro/SKILL.md
+++ b/skills/retro/SKILL.md
@@ -48,7 +48,7 @@ Ask:
## Phase 4 — Write retro
-Copy `04-projects/harness/templates/retro.md` to:
+Copy `references/retro-template.md` to:
`04-projects/harness/retro/YYYY-MM-DD-<slug>.md`
diff --git a/skills/retro/references/retro-template.md b/skills/retro/references/retro-template.md
new file mode 100644
index 0000000..831a1b7
--- /dev/null
+++ b/skills/retro/references/retro-template.md
@@ -0,0 +1,32 @@
+# Retro: <run or spec>
+
+> Date: YYYY-MM-DD · Run: `04-projects/harness/runs/<id>` · Lane: `<lane>` · Outcome: shipped | escalated | abandoned
+
+## What happened
+
+Two or three sentences. The sequence, not the narrative.
+
+## Evidence quality
+
+| AC | Had a PASS row | Observed the artifact | Notes |
+|---|---|---|---|
+| AC-01 | yes/no | yes/no: what was observed | |
+
+Any `PASS` that rested on a worker's summary instead of the artifact is a finding, not a pass.
+
+## What the gates caught
+
+| CP | Verdict | What it caught, or why it was silent |
+|---|---|---|
+
+## Friction
+
+Where the run lost time or nearly went wrong. One line each, concrete.
+
+## Actions
+
+`AI-n` rows go to `04-projects/harness/BACKLOG.md`. Skill or `CLAUDE.md` patches are proposed here, applied only with approval.
+
+| ID | Action | Target file | Status |
+|---|---|---|---|
+| AI-1 | | | proposed |
diff --git a/skills/review-cockpit/SKILL.md b/skills/review-cockpit/SKILL.md
index 78d4a48..94544d9 100644
--- a/skills/review-cockpit/SKILL.md
+++ b/skills/review-cockpit/SKILL.md
@@ -23,7 +23,7 @@ Pairs with `closed-loop` (verification), `harvest` (learnings), and the V-model
## Structure
-Copy `04-projects/harness/templates/session-review.md`. Save the working copy in the relevant project folder as `YYYY-MM-DD-<slug>-review.md` (or `-plan.md`).
+Copy `references/session-review-template.md`. Save the working copy in the relevant project folder as `YYYY-MM-DD-<slug>-review.md` (or `-plan.md`).
**Cockpit (top):**
1. **Status line** — one line: how many items await review + last-updated timestamp.
diff --git a/skills/review-cockpit/references/session-review-template.md b/skills/review-cockpit/references/session-review-template.md
new file mode 100644
index 0000000..e9cd86c
--- /dev/null
+++ b/skills/review-cockpit/references/session-review-template.md
@@ -0,0 +1,37 @@
+# <session topic>: review
+
+> Working folder: `<path>` · Updated: YYYY-MM-DD HH:MM
+
+## Progress
+
+- [x] <done item>
+- [ ] **Item 1** <what needs your call>
+- [ ] **Item 2** <what needs your call>
+
+## Context
+
+What you need to know to judge the items below: what changed, what was deliberately held back, what the constraint was.
+
+---
+
+## Item 1: <name>
+
+**Status:** needs review | approved | changes requested
+
+<the artifact, or a link to it, plus whatever the reader needs to judge it>
+
+### Your call
+
+- [ ] Approve as-is
+- [ ] Change: <write it here or say it in chat>
+
+---
+
+## Item 2: <name>
+
+**Status:** needs review
+
+### Your call
+
+- [ ] Approve as-is
+- [ ] Change:
diff --git a/skills/ultragoal/SKILL.md b/skills/ultragoal/SKILL.md
index 4f4c8ac..89eeb7e 100644
--- a/skills/ultragoal/SKILL.md
+++ b/skills/ultragoal/SKILL.md
@@ -4,12 +4,14 @@ description: >
Run a large, multi-session goal (e.g. shipping a whole side product) through the full V-model
closed loop, one phase at a time, with cross-session state and a final
north-star acceptance gate. Ultragoals never downgrade the lane: every
- phase runs CP-1→CP-6 with adversarial verification. Use via /ultragoal.
+ phase runs CP-1→CP-6 with adversarial verification. Opt-in: invoke with
+ /ultragoal or by calling something a long-running goal. Ordinary work does not
+ run this.
---
# Ultragoal — the closed loop for goals too big to ship in one run
-An **ultragoal** is the term for a north-star that spans many sessions: fork the repos, combine the strong parts, ship one product. A normal `/execute` run is one task through the loop. An ultragoal is a *chain of phases*, each of which is its own full closed-loop run, tracked so any cold session can resume.
+An **ultragoal** is the term for a north-star that spans many sessions: fork the repos, combine the strong parts, ship one product. A `/closed-loop` run is one task through the loop. An ultragoal is a *chain of phases*, each of which is its own full closed-loop run, tracked so any cold session can resume.
**Core rule (from dwarves-kit, adopted fully): the worker never grades its own homework, and wrongness compounds across sessions — so verify every phase, not just the end.**
@@ -20,7 +22,7 @@ An **ultragoal** is the term for a north-star that spans many sessions: fork the
- `/ultragoal status` — report all ultragoals from the registry
- Trigger phrases: "make this an ultragoal", "this is a long-running goal", "combine these into one product over time"
-Do **not** use for single-run work — that is `/execute`. Rule of thumb: if it needs a phase decomposition and won't finish today, it is an ultragoal.
+Do **not** use for single-run work (that is `/closed-loop`), and do not start one on a request that never asked for one. Rule of thumb: if it needs a phase decomposition and won't finish today, it is an ultragoal.
## Files (one goal = one folder)
@@ -36,7 +38,7 @@ Code, if any, lives outside the vault (e.g. `~/code/<goal>/`) — the spec/statu
## Phase 0 — Charter (`/ultragoal new`)
1. Interview the user for the **north-star** in one sentence (what "done" looks like).
-2. Write `04-projects/<goal>/spec.md` from `04-projects/harness/templates/SPEC-template.md`:
+2. Write `04-projects/<goal>/spec.md` from `../closed-loop/references/spec-template.md`:
- North-star statement
- Falsifiable `AC-n` acceptance criteria (these define done for the *whole* goal)
- Phase decomposition `P0…Pn` — each phase is a shippable increment mapped to a subset of `AC-n`
@@ -97,7 +99,7 @@ North-star acceptance:
Every ultragoal carries a single self-contained HTML report that **covers everything**: north-star, live status, all phases, the full `AC-n` traceability table with pass/open/fail, evidence rows per phase (with screenshots embedded as `data:` URIs), and the open-items / next-action block. It is the human-readable face of the evidence ledger.
-- **Template:** `04-projects/harness/templates/report.html` (house style, theme-aware, rows-not-cards). Copy it, then fill every `{{token}}` and `<!-- FILL -->` / `<!-- REPEAT -->` block from `spec.md` + `STATUS.md` + `evidence/P*/`. You fill it by editing — do not build a parser.
+- **Template:** `../closed-loop/references/report-template.html` (theme-aware, rows-not-cards). Copy it, then fill every `{{token}}` and `<!-- FILL -->` / `<!-- REPEAT -->` block from `spec.md` + `STATUS.md` + `evidence/P*/`. You fill it by editing; do not build a parser.
- **Deliverable path:** `04-projects/<goal>/report.html`. One file, overwritten each phase (it always reflects current truth).
- **When:** regenerate at each phase gate (CP-6) and again at final north-star acceptance. The final report must show every `AC-n` with a PASS row — if any pill is `open`, the goal is not done.
- **Self-contained only:** inline everything, embed screenshots as `data:` URIs, so it also works when published as an Artifact (external hosts are blocked). Before publishing as an Artifact, load the `artifact-design` skill.