AGENTS.md

1# Third Brain V7 Skills — Codex CLI
2 
3This repository contains Agent Skills for the Codex CLI environment. Place skills in `~/.agents/skills/`.
4 
5V7 treats Obsidian as durable disk and governance, while every skill uses a profile-aware execution contract. V7.1 adds the 20th skill, `graph-engineering`, for bounded static dependency DAGs. Do not turn wiki signals into rules until they pass the promotion gate: repeated source support or local verification, bounded macro action, preserved provenance/permissions, and a cheap objective check.
6 
7## Skills
8 
9### 📥 Knowledge Pipeline
10- **wiki-ingest** — STOW pipeline with source-risk classification, block refs, clipping lifecycle, understanding gate, promotion controls, and post-ingest lint.
11- **knowledge-ops** — Multi-layer knowledge management with Markdown-first retrieval, optional vector storage, evidence hierarchy, deduplication, Agent/Wiki flywheel, and knowledge debt queues.
12- **wiki-lint** — Wiki health check for graph integrity, provenance, links, clipping lifecycle, understanding, and promotion readiness.
13 
14### 🔄 Daily Loop
15- **daily-okr** — 7 Key Results daily knowledge compound cycle; may consume the scheduled daily knowledge-loop note when present.
16- **cognitive-compile** — 8-section deep learning compile framework.
17 
18### 🎨 Behavior & Creativity
19- **behavior-design** — Behavior change system with HAS framework.
20- **creativity-engine** — Combinatorial ideation + minimum experiments.
21 
22### 🔬 Research & Quality
23- **deep-research** — STOW-compatible research harness with ChatGPT-style preflight, source/claim ledgers, activity trace, citations, privacy checks, and wiki-ingest handoff.
24- **verify-before-claim** — Verification-first quality gate.
25 
26### 🔄 Learning
27- **session-learn** — Knowledge extraction with Closure Protocol.
28- **project-flow-ops** — Project triage and tracking.
29 
30### 📊 Context & Cost
31- **context-manager** — Runtime-derived context budgets, checkpoints, compaction, retrieval, and capability routing.
32- `token-cost-tracker` is a command under `commands/`, not an Agent Skill.
33 
34### 🏗️ Engineering
35- **agentic-engineering** — Agent-as-process workflow refactoring with autonomy defaults, delegated-action boundaries, state checkpoints, write-back, and verification gates.
36- **loop-engineering** — Temporal-depth control through bounded Trigger -> Execute -> Verify -> State loops, durable contracts, hard budgets, and stop/recovery rules.
37- **graph-engineering** — Dependency-width control for bounded static DAGs with explicit dependencies, independent branches, typed joins, and node-local recovery; dynamic and cyclic graphs are out of scope in V7.1.
38- **harness-engineering** — Runtime kernel design: scheduler, permissions, tools as system calls, MCP/Skills/Hooks selection, provenance ledgers, observability, and recovery.
39- **agent-teams-command** — Multi-agent process ownership, IPC, worktree isolation, async budget envelopes, integration, cleanup, and evidence gates.
40 
41### 💼 Strategy & Operations
42- **startup-evaluation** — Startup health diagnosis with entrepreneurship, VC 5T, PMF, runway, team, and next-test frameworks.
43- **anthropic-os** — Self-evolving work method engine.
44- **ai-six-sigma-property-os** — AI + Ontology + DMAIC Black Belt operating model for property work orders, dispatch, quotes, evidence, CTQ dashboards, and MVP quality control.
45 
46## Compatibility
47 
48All skills follow the [Agent Skills](https://agentskills.io) open format. Skills are model-agnostic markdown files compatible with Codex CLI.
49 
50## Skill Contract
51 
52When selecting a skill, read its frontmatter before executing:
53 
54- `assumes` — required operating assumptions.
55- `conflicts_with` — boundaries that must not be silently overridden.
56- `metadata.profile` — `one-shot`, `stateful`, `loop`, or `high-risk` controls.
57- `## Failure Protocol` and `## Output Contract` — standard stop status and receipt.
58- `## Success Metrics` — the minimum observable result for one successful run.
59- `## Quality Gates` — checks that must pass before claiming completion.
60 
61For wiki-writing skills, resolve paths from `system/config.md` when available. Defaults include `SOURCES_DIR=sources/`, `CONCEPTS_DIR=wiki/concepts/`, `ENTITIES_DIR=wiki/entities/`, and `LOG_FILE=system/log.md`.
62 
63## V7 Promotion Gate
64 
65Promote wiki knowledge into a skill, SOP, schema rule, or automation only when:
66 
67- At least two durable wiki/source pages support it, or one high-quality source plus local verification supports it.
68- The rule can be expressed as Trigger -> Execute -> Verify -> State with owner, budget, stop condition, recovery, and write-back.
69- Source immutability, block refs, clipping lifecycle, review queues, permissions, and human approval boundaries are preserved.
70- The change has a cheap check: skill lint, wiki lint, link check, script receipt, test output, dashboard metric, or review receipt.
71 
72## Adoption Ladder
73 
741. Start with `wiki-ingest` + `verify-before-claim`.
752. Add `daily-okr` + `session-learn` after daily ingest and evidence checks are working.
763. Add `cognitive-compile`, `behavior-design`, and `creativity-engine` when the wiki has enough material to synthesize.
774. Add `knowledge-ops` and `loop-engineering` when retrieval or repeated verified execution becomes a bottleneck.
785. Add `graph-engineering` after `loop-engineering` only when explicit dependencies, parallel branches, typed joins, or node-local recovery provide more value than orchestration and review cost.
796. Add `harness-engineering`, `agentic-engineering`, and `agent-teams-command` only when runtime controls, workflow autonomy, or multi-agent process ownership requires them.
80 
81Routing boundary: Loop = temporal depth; Graph = dependency width; Agent Teams = process ownership, IPC, and integration; Harness = runtime scheduler, permissions, and observability.
82 
83## Karpathy LLM OS
84 
85LLM=CPU · Context=RAM · Storage=Disk · Tools=System Calls · Skills=Programs · Harness=Kernel · Agent Teams=Processes
86