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.
Python agent-loopagent-pluginsagent-skillsagenticai-agentsclaude-codeclosed-loopcursorgarry-tangbraingstackknowledge-managementloop-engineeringobsidianpeople-crmproductivitysecond-brainself-evolvingverification-harnessworker-agents
style: drop em-dashes from v3.8.0 prose
Huy Tieu58c5bd22916d
diff --git a/AGENTS.md b/AGENTS.mdindex 8148633..4aec0d9 100644--- a/AGENTS.md+++ b/AGENTS.md@@ -446,7 +446,7 @@ The following 5 skills implement the V-model closed loop described in `WORKFLOW.### /closed-loop-**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`).+**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>`@@ -459,7 +459,7 @@ The following 5 skills implement the V-model closed loop described in `WORKFLOW.**What it does:**1. Classifies the task into a risk lane (`bash .claude/lib/lane-classify.sh classify "<task>"`)2. Runs the worker at CP-3, traced to `AC-n`-3. Dispatches a fresh-context, read-only `task-verifier` at CP-3v that observes the *artifact* (curl the URL, re-read the file, screenshot the page) — never the worker's summary+3. Dispatches a fresh-context, read-only `task-verifier` at CP-3v that observes the *artifact* (curl the URL, re-read the file, screenshot the page), never the worker's summary4. On `FAIL:fixable`, dispatches `fix-agent` (max 2 retries), then re-verifies; on `FAIL:escalate`, stops5. Runs `integration-verifier` at CP-4 for multi-task / `full`-lane work6. Records checkpoints and evidence rows to the run's `evidence/ledger.md`@@ -470,7 +470,7 @@ The following 5 skills implement the V-model closed loop described in `WORKFLOW.### /ultragoal-**Description:** Run a goal too big to ship in one session — a chain of phases, each its own full closed-loop run, with cross-session state and a final north-star acceptance gate.+**Description:** Run a goal too big to ship in one session: a chain of phases, each its own full closed-loop run, with cross-session state and a final north-star acceptance gate.**Triggers:**- `/ultragoal <goal>`@@ -502,7 +502,7 @@ The following 5 skills implement the V-model closed loop described in `WORKFLOW.- After a correction, a rejected deliverable, or a surprising discovery- Nightly self-enhancement runs-**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.+**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.**What it does:**1. Scans the session for corrections, rejected outputs, non-obvious discoveries, and repeated friction@@ -517,7 +517,7 @@ The following 5 skills implement the V-model closed loop described in `WORKFLOW.### /retro-**Description:** CP-7 retrospective — audit the run's checkpoints, evidence quality, action items, and harvest candidates. Closes the V-model cycle and feeds the next one.+**Description:** CP-7 retrospective: audit the run's checkpoints, evidence quality, action items, and harvest candidates. Closes the V-model cycle and feeds the next one.**Triggers:**- `/retro <run or spec>`@@ -528,7 +528,7 @@ The following 5 skills implement the V-model closed loop described in `WORKFLOW.**What it does:**1. Reads the run's evidence ledger and checkpoint log-2. Audits evidence *quality* — did each row observe the artifact, or restate a tool return value?+2. Audits evidence *quality*. Did each row observe the artifact, or restate a tool return value?3. Identifies which checkpoints caught real problems and which were ceremony4. Extracts action items and harvest candidates5. Writes `04-projects/harness/retro/YYYY-MM-DD-<slug>.md`@@ -539,7 +539,7 @@ The following 5 skills implement the V-model closed loop described in `WORKFLOW.### /review-cockpit-**Description:** One living review document per multi-item session — a cockpit header (Progress checklist, Working folder, Context) plus per-item review cards you approve or request changes on directly in the doc.+**Description:** One living review document per multi-item session: a cockpit header (Progress checklist, Working folder, Context) plus per-item review cards you approve or request changes on directly in the doc.**Triggers:**- Any session with **≥2 deliverables you need to review or approve**@@ -567,7 +567,7 @@ The following 5 skills raise output quality on writing and visual work. They enc### /no-ai-slop-**Description:** Edit drafts into sharper, more human writing while preserving the writer's personal voice — or detect AI-slop patterns without rewriting.+**Description:** Edit drafts into sharper, more human writing while preserving the writer's personal voice, or detect AI-slop patterns without rewriting.**Triggers:**- "make this less AI-sounding"@@ -589,7 +589,7 @@ The following 5 skills raise output quality on writing and visual work. They enc### /editorial-illustrations-**Description:** Generate meaning-carrying editorial data-illustrations in a near-black grayscale + single-accent aesthetic. A **generative guide**, not a template gallery — it teaches the "claim → geometry" method.+**Description:** Generate meaning-carrying editorial data-illustrations in a near-black grayscale + single-accent aesthetic. A **generative guide**, not a template gallery. It teaches the "claim → geometry" method.**Triggers:**- "make a chart/diagram/illustration for this"@@ -599,7 +599,7 @@ The following 5 skills raise output quality on writing and visual work. They enc**Purpose:** Most generated figures decorate; they don't argue. This teaches the session to derive the *right* geometry from what the text actually claims, then render it as a self-contained, theme-aware, reduced-motion-safe HTML/SVG figure.**What it does:**-1. Extracts the claim, then asks "where is the point?" — the one element that gets the accent+1. Extracts the claim, then asks "where is the point?", the one element that gets the accent2. Selects geometry from the claim's shape (a rail, a ramp, a bridge, a lane split), not from a chart-type menu3. Renders self-contained HTML/SVG with a grayscale ramp plus exactly one accent, AA-safe on both grounds4. Runs a pre-flight checklist: one accent on the point, no clipped mono text, no misaligned SVG, theme toggle honored@@ -610,19 +610,19 @@ The following 5 skills raise output quality on writing and visual work. They enc### /data-forms-**Description:** Pick the right way to represent a dataset so a reader gets the finding in three seconds — a catalog of 20+ chart and diagram forms with when-to-use and failure modes, plus the encoding decisions that make any of them readable.+**Description:** Pick the right way to represent a dataset so a reader gets the finding in three seconds: a catalog of 20+ chart and diagram forms with when-to-use and failure modes, plus the encoding decisions that make any of them readable.**Triggers:**- Charting survey results, benchmark data, usage metrics, or research findings- "the default bar chart is burying the point"- Building a post, brief, deck, or report with data in it-**Purpose:** A repertoire, not a style. What makes good data illustration work is not the palette — it's form selection and encoding discipline, both of which are portable to any visual language.+**Purpose:** A repertoire, not a style. What makes good data illustration work is not the palette. It is form selection and encoding discipline, both of which are portable to any visual language.**What it does:**1. Matches the dataset's question to a form from the catalog (with each form's failure modes stated)2. Applies the encoding rules that carry across all of them: takeaway headline, direct labels, kill the axis, highlight-and-mute, show the caveat-3. Stays style-agnostic — hands off to `dataviz` (or your host design system) for palette and accessibility+3. Stays style-agnostic, handing off to `dataviz` (or your host design system) for palette and accessibility**References:** `references/forms.md`.@@ -637,13 +637,13 @@ The following 5 skills raise output quality on writing and visual work. They enc- Deck backgrounds, social cards, essay figures, spec cover art- Any visual that needs aesthetic weight and credibility-**Purpose:** Curated, historically significant art reads as credible and sophisticated; AI-generated imagery reads as slop. Stacks with `/no-ai-slop`. Does **not** replace `/editorial-illustrations`, which owns claim-driven diagrams — museum art is for photographic, hero, decorative, and mood imagery.+**Purpose:** Curated, historically significant art reads as credible and sophisticated; AI-generated imagery reads as slop. Stacks with `/no-ai-slop`. Does **not** replace `/editorial-illustrations`, which owns claim-driven diagrams. Museum art is for photographic, hero, decorative, and mood imagery.**What it does:**1. Queries museum open-access APIs (verified keyless recipes per institution)2. Confirms public-domain status and records the licensing terms3. Returns high-res source URLs with attribution metadata-4. Fetches fresh each time — no reusable image pool, so visuals don't repeat across posts+4. Fetches fresh each time, so there is no reusable image pool, so visuals don't repeat across posts**References:** one file per institution under `references/`.@@ -661,7 +661,7 @@ The following 5 skills raise output quality on writing and visual work. They enc**Purpose:** Distinct from `/weekly-checkin`, where you supply the input. Here **the agent is the author** and the log accrues in the background, so the record exists even on days you'd never sit down to write one.**What it does:**-1. **`log` (default, mostly implicit):** appends one entry after meaningful work — what was done, focus thread, artifacts touched, optional signal+1. **`log` (default, mostly implicit):** appends one entry after meaningful work: what was done, focus thread, artifacts touched, optional signal2. Skips trivia: one-line lookups, scratch work, its own writes, anything you asked to keep out3. **`reflect`:** reads the day's log plus recent days, summarizes the day back to you, asks 2-4 light questions adapted to what the log shows, then writes your answers plus a synthesis into the day's file4. On request, synthesizes the last 7 files into a week-in-review@@ -819,7 +819,7 @@ The following 6 skills form a complete product management lifecycle:COG includes 10 specialized agents (`.claude/agents/`) that handle data-heavy and verification tasks using Sonnet while the lead session (Opus) handles reasoning and synthesis. Inspired by [garrytan/gstack](https://github.com/garrytan/gstack) specialist sessions and [garrytan/gbrain](https://github.com/garrytan/gbrain) knowledge patterns.-**Workers** — do the gathering and the mutating:+**Workers** handle the gathering and the mutating:| Agent | What it does | When it's used ||---|---|---|@@ -830,7 +830,7 @@ COG includes 10 specialized agents (`.claude/agents/`) that handle data-heavy an| **worker-publisher** | Publishing to Slack, Confluence, Notion, webhooks | Brief publishing, wiki sync || **brief-people-updater** | Batch-update people profiles from meetings/briefs | After team briefs, meeting processing |-**Verifiers** — read-only, fresh-context, and they cannot edit files or mutate external state:+**Verifiers** are read-only and fresh-context, and they cannot edit files or mutate external state:| Agent | What it does | Checkpoint ||---|---|---|@@ -841,7 +841,7 @@ COG includes 10 specialized agents (`.claude/agents/`) that handle data-heavy an**Key rules:**- Workers write results to `/tmp/{task-slug}.md` and return only a short status + file path. The lead session reads the file for synthesis.-- Verifiers receive **paths only** — never paste a worker's output into a verifier's prompt. Pasted context induces narrativisation: the verifier classifies the framing instead of independently reading the source.+- Verifiers receive **paths only**. Never paste a worker's output into a verifier's prompt. Pasted context induces narrativisation: the verifier classifies the framing instead of independently reading the source.---diff --git a/CHANGELOG.md b/CHANGELOG.mdindex 89ed1bc..dfe6d8e 100644--- a/CHANGELOG.md+++ b/CHANGELOG.md@@ -8,9 +8,9 @@ All notable changes to COG (Cognition + Obsidian + Git) will be documented in thv3.7.0 made trust a runtime decision for stored facts and mutations. v3.8.0 asks the harder question: **who checks the checker?**-The unit of work in v3.7.0 was a skill run. A skill run reports success. Nothing above it decides whether that success was real, because the only thing that ever looked at the result was the agent that produced it. That is not a verification problem, it is a *structural* one — the worker grades its own homework.+The unit of work in v3.7.0 was a skill run. A skill run reports success. Nothing above it decides whether that success was real, because the only thing that ever looked at the result was the agent that produced it. That is not a verification problem, it is a *structural* one. The worker grades its own homework.-v3.8.0 replaces the flat skill run with a **V-model lifecycle**. Work descends the left arm (goal → falsifiable criteria → tasks), builds at the apex, and ascends the right arm through gates that each demand **evidence traced back to a criterion ID**. A criterion with no evidence row does not ship. An evidence row with no criterion is noise. The verifier is a separate agent, read-only, with fresh context and no access to the worker's narrative — it looks at the artifact.+v3.8.0 replaces the flat skill run with a **V-model lifecycle**. Work descends the left arm (goal → falsifiable criteria → tasks), builds at the apex, and ascends the right arm through gates that each demand **evidence traced back to a criterion ID**. A criterion with no evidence row does not ship. An evidence row with no criterion is noise. The verifier is a separate agent, read-only, with fresh context and no access to the worker's narrative. It looks at the artifact.Two ideas do most of the work:@@ -21,44 +21,44 @@ Closed-loop execution and the risk-lane concept are adapted from [dwarves-kit](h### Added-#### New Skills — Verification Harness (5)-- **`closed-loop`** — the execute pipeline: CP-2 plan → CP-3 build → CP-3v component verify → CP-4 integration verify → CP-5 acceptance. Dispatches a fresh-context, read-only `task-verifier`, routes `FAIL:fixable` to `fix-agent` (max 2 retries), escalates otherwise. Every verify step emits `EVIDENCE <AC-id> | <CP> | PASS|FAIL | <observation> | <artifact>`.-- **`ultragoal`** — for goals too big to ship in one session. One spec with a north-star and `AC-n` criteria, decomposed into phases, each phase a complete closed-loop run with its own evidence bundle. A living `STATUS.md` lets a cold session resume without re-reading history. Two acceptance gates: per-phase, and a final north-star verifier that checks every criterion has a PASS row. Ultragoals never downgrade the lane, because wrongness compounds across sessions.-- **`harvest`** — captures durable session learnings (corrections, rejected outputs, non-obvious discoveries, repeated friction) at the session boundary and **stages** them. Auto-promoting session noise into durable knowledge poisons the well, so nothing reaches `05-knowledge/` without approval.-- **`retro`** — CP-7. Audits not just whether checkpoints passed but whether the **evidence was any good**: did each row observe an artifact, or restate a tool return value? Identifies which gates caught real problems and which were ceremony.-- **`review-cockpit`** — one living review doc per multi-item session. Cockpit header (Progress / Working folder / Context) plus per-item cards with a **🗒 Your call** approval slot you edit in place. The doc is the interaction surface, not a summary of it.+#### New Skills: Verification Harness (5)+- **`closed-loop`**: the execute pipeline is CP-2 plan → CP-3 build → CP-3v component verify → CP-4 integration verify → CP-5 acceptance. Dispatches a fresh-context, read-only `task-verifier`, routes `FAIL:fixable` to `fix-agent` (max 2 retries), escalates otherwise. Every verify step emits `EVIDENCE <AC-id> | <CP> | PASS|FAIL | <observation> | <artifact>`.+- **`ultragoal`**: for goals too big to ship in one session. One spec with a north-star and `AC-n` criteria, decomposed into phases, each phase a complete closed-loop run with its own evidence bundle. A living `STATUS.md` lets a cold session resume without re-reading history. Two acceptance gates: per-phase, and a final north-star verifier that checks every criterion has a PASS row. Ultragoals never downgrade the lane, because wrongness compounds across sessions.+- **`harvest`**: captures durable session learnings (corrections, rejected outputs, non-obvious discoveries, repeated friction) at the session boundary and **stages** them. Auto-promoting session noise into durable knowledge poisons the well, so nothing reaches `05-knowledge/` without approval.+- **`retro`**: CP-7. Audits not just whether checkpoints passed but whether the **evidence was any good**: did each row observe an artifact, or restate a tool return value? Identifies which gates caught real problems and which were ceremony.+- **`review-cockpit`**: one living review doc per multi-item session. Cockpit header (Progress / Working folder / Context) plus per-item cards with a **🗒 Your call** approval slot you edit in place. The doc is the interaction surface, not a summary of it.-#### New Skills — Craft (5)-- **`no-ai-slop`** — edit a draft sharper while preserving voice, or detect slop without rewriting. Guards both failure directions: leaving AI patterns in, and stripping distinctive voice out along with them. Vendored from [petergyang/no-ai-slop](https://github.com/petergyang/no-ai-slop) (MIT, with `LICENSE` and `SOURCE.md`).-- **`editorial-illustrations`** — a generative guide, not a template gallery. Teaches the **claim → geometry** method: extract what the text argues, find where the point is, derive the right geometry from the claim's shape, then render a self-contained, theme-aware, reduced-motion-safe HTML/SVG figure with a grayscale ramp and exactly one accent.-- **`data-forms`** — 20+ chart and diagram forms with when-to-use and failure modes, plus the encoding decisions that carry across all of them (takeaway headline, direct labels, kill the axis, highlight-and-mute, show the caveat). Style-agnostic: a repertoire, not a palette.-- **`museum-art`** — source real public-domain artwork from museum open-access APIs (Met, Cleveland, SMK, Rijksmuseum, NGA, Art Institute of Chicago, Getty, Smithsonian) instead of AI-generated or stock imagery. Keyless recipes per institution, licensing rules included, fetched fresh so visuals don't repeat.-- **`daily-journal`** — a passive work journal the agent keeps *for* you. Appends entries after meaningful work; the record exists even on days you'd never sit down to write one.+#### New Skills: Craft (5)+- **`no-ai-slop`**: edit a draft sharper while preserving voice, or detect slop without rewriting. Guards both failure directions: leaving AI patterns in, and stripping distinctive voice out along with them. Vendored from [petergyang/no-ai-slop](https://github.com/petergyang/no-ai-slop) (MIT, with `LICENSE` and `SOURCE.md`).+- **`editorial-illustrations`**: a generative guide, not a template gallery. Teaches the **claim → geometry** method: extract what the text argues, find where the point is, derive the right geometry from the claim's shape, then render a self-contained, theme-aware, reduced-motion-safe HTML/SVG figure with a grayscale ramp and exactly one accent.+- **`data-forms`**: 20+ chart and diagram forms with when-to-use and failure modes, plus the encoding decisions that carry across all of them (takeaway headline, direct labels, kill the axis, highlight-and-mute, show the caveat). Style-agnostic: a repertoire, not a palette.+- **`museum-art`**: source real public-domain artwork from museum open-access APIs (Met, Cleveland, SMK, Rijksmuseum, NGA, Art Institute of Chicago, Getty, Smithsonian) instead of AI-generated or stock imagery. Keyless recipes per institution, licensing rules included, fetched fresh so visuals don't repeat.+- **`daily-journal`**: a passive work journal the agent keeps *for* you. Appends entries after meaningful work; the record exists even on days you'd never sit down to write one.#### New Agents (4)All read-only except `fix-agent`, all fresh-context, none can mutate external state:-- **`task-verifier`** (CP-3v) — checks worker output against acceptance criteria by observing the artifact.-- **`integration-verifier`** (CP-4) — cross-task wiring and global acceptance for multi-task specs.-- **`fix-agent`** — targeted fixes after a `FAIL:fixable` verdict; implements only what the verifier flagged, max 2 attempts.-- **`harvest-curator`** (CP-7) — shapes session learnings into adoption notes; propose-only.+- **`task-verifier`** (CP-3v): checks worker output against acceptance criteria by observing the artifact.+- **`integration-verifier`** (CP-4): cross-task wiring and global acceptance for multi-task specs.+- **`fix-agent`**: targeted fixes after a `FAIL:fixable` verdict; implements only what the verifier flagged, max 2 attempts.+- **`harvest-curator`** (CP-7): shapes session learnings into adoption notes; propose-only.#### New Framework Files-- **`WORKFLOW.md`** — the V-model lifecycle: checkpoint table, gate classes (blocking vs advisory), the evidence row contract, risk-lane checkpoint depth, the verification pipeline, and file homes for specs, evidence bundles, and retro output.-- **`.claude/lib/checkpoint.sh`** — records checkpoint results to a run's evidence ledger.-- **`.claude/lib/lane-classify.sh`** — classifies a task into a risk lane (`classify` for the verdict, `explain` for the reasoning).+- **`WORKFLOW.md`**: the V-model lifecycle covers checkpoint table, gate classes (blocking vs advisory), the evidence row contract, risk-lane checkpoint depth, the verification pipeline, and file homes for specs, evidence bundles, and retro output.+- **`.claude/lib/checkpoint.sh`**: records checkpoint results to a run's evidence ledger.+- **`.claude/lib/lane-classify.sh`**: classifies a task into a risk lane (`classify` for the verdict, `explain` for the reasoning).#### New Protocols in CLAUDE.md-- **V-Model Checkpoints** — CP-1 through CP-7 with per-lane blocking rules, the two-way verification requirement, and the **cross-model flagship gate**: on high-stakes runs, overlay a flagship model that is *not* the lead's own as advisor at CP-1 and critic at CP-4/5/6. Same-family verifiers share the lead's blind spots; a different family catches a different error class.-- **Closed-Loop Execute** — the pipeline, the five lanes, and when a verifier subagent is actually warranted. Notably: on `normal`-lane read-only work the lead verifies inline, because spawning an agent to re-read a file the lead just wrote buys nothing.-- **Risk Lanes** — `tiny` / `normal` / `full` / `bug` / `backfill`, classified before executing.-- **Ultragoal** — one spec, phases as full closed-loop runs, a living status ledger, two acceptance gates.-- **Visual Verification** — UI/UX work is not verified by a DOM check. The DOM can be present and the pixels still wrong. Capture the render, *read the image*, name the discrepancy, fix it, re-capture.-- **Delegation Cap** — delegation costs context re-establishment on both ends. Fan out only for genuinely independent, sizeable tracks; if one subagent can do it, use one.+- **V-Model Checkpoints**: CP-1 through CP-7 with per-lane blocking rules, the two-way verification requirement, and the **cross-model flagship gate**: on high-stakes runs, overlay a flagship model that is *not* the lead's own as advisor at CP-1 and critic at CP-4/5/6. Same-family verifiers share the lead's blind spots; a different family catches a different error class.+- **Closed-Loop Execute**: the pipeline, the five lanes, and when a verifier subagent is actually warranted. Notably: on `normal`-lane read-only work the lead verifies inline, because spawning an agent to re-read a file the lead just wrote buys nothing.+- **Risk Lanes**: `tiny` / `normal` / `full` / `bug` / `backfill`, classified before executing.+- **Ultragoal**: one spec, phases as full closed-loop runs, a living status ledger, two acceptance gates.+- **Visual Verification**: UI/UX work is not verified by a DOM check. The DOM can be present and the pixels still wrong. Capture the render, *read the image*, name the discrepancy, fix it, re-capture.+- **Delegation Cap**: delegation costs context re-establishment on both ends. Fan out only for genuinely independent, sizeable tracks; if one subagent can do it, use one.### Changed- **Skill count 21 → 31**, **agents 6 → 10** across `plugin.json`, `.cursor-plugin/plugin.json`, `marketplace-entry.json`, `AGENTS.md`, `README.md`, `SETUP.md`, `.cursorrules`, `.github/MARKETPLACE.md`, and `docs/AGENT-SUPPORT.md`.- **Model Routing table** extended with the four new agents (all Sonnet).-- **`cog-update.sh`** — `FRAMEWORK_FILES` now covers the 10 new skills, 4 new agents, `WORKFLOW.md`, and both `.claude/lib/` scripts.+- **`cog-update.sh`**: `FRAMEWORK_FILES` now covers the 10 new skills, 4 new agents, `WORKFLOW.md`, and both `.claude/lib/` scripts.- Version bumped to **3.8.0** across `COG-VERSION` and all manifests.## [3.7.1] - 2026-07-10diff --git a/CLAUDE.md b/CLAUDE.mdindex 69b0e12..2fcf8c7 100644--- a/CLAUDE.md+++ b/CLAUDE.md@@ -2,7 +2,7 @@> Lifecycle: `WORKFLOW.md` (V-model harness) · Universal surface: `AGENTS.md`-## V-Model Checkpoints — ALWAYS APPLY+## V-Model Checkpoints (ALWAYS APPLY)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.@@ -21,7 +21,7 @@ Work walks the **V**: decompose left (spec → plan), build at the apex, verify**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`.-## Closed-Loop Execute — ALWAYS APPLY (build tasks)+## Closed-Loop Execute (ALWAYS APPLY, build tasks)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).@@ -39,13 +39,13 @@ worker → task-verifier (read-only) → fix-agent (max 2 retries) → escalate**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.+- 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+## Risk Lanes (ALWAYS APPLY)Classify before executing: `bash .claude/lib/lane-classify.sh explain "<task>"`.@@ -55,7 +55,7 @@ Classify before executing: `bash .claude/lib/lane-classify.sh explain "<task>"`.- **bug**: root cause before patch.- **backfill**: no behavior change without approval.-## Ultragoal — ALWAYS APPLY (long-running goals)+## 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`.@@ -65,21 +65,21 @@ An **ultragoal** is a large, multi-session goal that can't ship in one run. It g- **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.-## Visual Verification — ALWAYS APPLY (UI/UX tasks)+## Visual Verification (ALWAYS APPLY, UI/UX tasks)Any task that **implements or changes a UI/UX flow** is not verified by a DOM/selector check. The DOM can be present and the pixels still wrong: overflow, misalignment, clipped text, wrong contrast, broken responsive layout, z-index overlap.- **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** — part of the task, not a follow-up. Re-capture to prove it.+- **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.-## Delegation Cap — ALWAYS APPLY+## Delegation Cap (ALWAYS APPLY)-Delegation is not free: each subagent re-establishes context, re-explores, and reports back, and the lead then re-reads the report. Delegate when the payoff clearly exceeds that overhead — not by reflex.+Delegation is not free: each subagent re-establishes context, re-explores, and reports back, and the lead then re-reads the report. Delegate when the payoff clearly exceeds that overhead, not by reflex.- **Don't delegate work the lead can finish in a handful of tool calls.**-- **Fan out only for genuinely independent, sizeable tracks** — ≥3 unrelated items, a wide multi-source sweep, or parallel workers that would conflict on the same file.+- **Fan out only for genuinely independent, sizeable tracks**: ≥3 unrelated items, a wide multi-source sweep, or parallel workers that would conflict on the same file.- **If one subagent can do it, use one.** Keep spawn counts low.- **Brief precisely the first time.** Avoid launch → wait → re-brief; never redo a subagent's work after it reports.- Independent agents launched together go in **one message with multiple tool calls** so they run concurrently.@@ -244,7 +244,7 @@ Role packs live in `.claude/roles/`. New roles can be added by dropping a file f- `COMPETITIVE-WATCHLIST.md` — Companies/people being tracked### Framework files (updated via `cog-update.sh` or `/update-cog`)-- `.claude/skills/` — Claude Code skills (31 skills)+- `.claude/skills/`: Claude Code skills (31 skills)- `.claude/agents/` — Worker agent definitions (6 agents)- `.claude/roles/` — Role packs for personalized recommendations- `.kiro/powers/` — Kiro powersdiff --git a/README.md b/README.mdindex 734a764..c18d983 100644--- a/README.md+++ b/README.md@@ -121,7 +121,7 @@ The enforcement layer that keeps every other skill honest. Work walks a **V**: d| **closed-loop** | Plan → build → component verify → integration verify → acceptance. The worker never grades its own homework; verifiers observe the artifact, not the worker's summary | "Run this through the closed loop" || **ultragoal** | A goal too big for one session, run as a chain of phases with cross-session state and a north-star acceptance gate | "Track this as an ultragoal" || **harvest** | Capture session learnings, stage them for approval, propose skill patches. Never writes durable knowledge unapproved | "Harvest what we learned" |-| **retro** | Audit the run's checkpoints and *evidence quality* — a run can pass every gate on weak observations | "Retro this run" |+| **retro** | Audit the run's checkpoints and *evidence quality*. A run can pass every gate on weak observations | "Retro this run" || **review-cockpit** | One living doc per multi-item session: progress, per-item cards, and an approval slot you edit in place | "Set up a review cockpit" |### Craft Skills (Writing & Visuals)@@ -158,7 +158,7 @@ COG uses a worker agent architecture inspired by [garrytan/gstack](https://githu> Workers write results to `/tmp/` files and return only a status + path. The lead reads the file for synthesis. This eliminates slow token generation in agent output.>-> Verifiers receive **paths only** — never a worker's output. Pasted context induces narrativisation: the verifier classifies the framing instead of independently reading the source.+> Verifiers receive **paths only**, never a worker's output. Pasted context induces narrativisation: the verifier classifies the framing instead of independently reading the source.### People CRM (Knowledge-Based Team Profiles)@@ -213,7 +213,7 @@ graph TD| **Privacy-First** — Local `.md` files, strict domain separation, no external servers | **Multi-Device** — iCloud sync to iPhone/iPad/Mac; Git for version history | **Obsidian Tasks** — `📅 YYYY-MM-DD` emoji format works with Tasks plugin dashboards || **Garry Tan Inspired** — gstack specialist sessions + gbrain knowledge patterns | **Multi-Platform** — Listed on [skills.sh](https://skills.sh), [agentskill.sh](https://agentskill.sh), [cursor.directory](https://cursor.directory) | **Worker Agents** — Sonnet handles I/O, Opus handles thinking || **Runtime Trust** — memory hygiene sweeps re-verify stored facts against the live environment | **Post-Condition Checks** — mutating skills observe the artifact before reporting success | **Skill Distillation** — [explore once, execute cheap](docs/SKILL-DISTILLATION.md) with small models |-| **V-Model Closed Loop** — every criterion has a task on the left and an evidence row on the right before ship | **Risk Lanes** — ceremony scales with blast radius instead of applying uniformly | **Cross-Model Gate** — a different model family catches a different error class |+| **V-Model Closed Loop** — every criterion has a task on the left and an evidence row on the right | **Risk Lanes** — ceremony scales with blast radius | **Cross-Model Gate** — a different model family catches a different error class |## Your Vault