English | 한국어 | 日本語 | 中文 | Deutsch | Français

Looking for Codex CLI? → my-codex — same orchestration in native TOML format

---

my-claude

All-in-one agent harness for Claude Code. One plugin, 32 curated agents ready.

Boss auto-discovers every agent, skill, and MCP tool at runtime, then routes your task to the right specialist. No config files. No boilerplate.

---

Installation

For Humans

git clone --depth 1 https://github.com/sehoon787/my-claude.git /tmp/my-claude
bash /tmp/my-claude/install.sh
rm -rf /tmp/my-claude

The default install is lean on purpose: every installed skill's description is re-sent as context on every request. Frontend work needs 18 more skills, so they are a separate lane:

bash install.sh --skills=web    # add the React/Vue/Nuxt/Nest, motion, a11y and browser-e2e skills
bash install.sh --full-skills   # every optional lane
bash install.sh --skills=       # back to the default

The choice is saved to ~/.claude/.my-claude-skills, so a later plain bash install.sh keeps it. MY_CLAUDE_SKILLS=web does the same thing.

Or install as a Claude Code plugin first, then run the companion installer:

# Inside a Claude Code session:
/plugin marketplace add sehoon787/my-claude
/plugin install my-claude@my-claude

If you use both routes, install.sh also refreshes the my-claude plugin (marketplace update + plugin update) so its hooks stay in sync with this script's install. The installer also ensures one shared codeburn dashboard and one shared Headroom proxy for both my-claude and my-codex. A later installer reuses healthy services instead of starting duplicates. State and startup logs live under ${XDG_STATE_HOME:-$HOME/.local/state}/agent-harness-services; neither installer changes ANTHROPIC_BASE_URL or OPENAI_BASE_URL.

For AI Agents

Read https://raw.githubusercontent.com/sehoon787/my-claude/main/AI-INSTALL.md and follow every step.

Before running the installer, the agent will ask which companion tools (Serena, Headroom, codeburn) to install, because the interactive checkbox menu only appears in an interactive terminal.

---

Open-Source Tools Used

Every project this stack builds on is described once, here. Five MIT-licensed upstreams are linked as git submodules pinned to an explicit SHA; the rest arrive as version-pinned CLIs, hosted MCP servers, or files vendored with attribution.

#	Project	What my-claude takes from it	How it arrives
1	oh-my-claudecode (OMC) — Yeachan Heo	19 specialist agents (architect, debugger, code reviewer, security reviewer, …) that divide work by role, plus 13 orchestration skills — autopilot, ralph, team. Magic keywords like autopilot: activate automatic parallel execution.	Submodule upstream/omc, SHA-pinned (see Bundled Upstream Versions); install.sh runs npm i -g oh-my-claude-sisyphus@latest and enables the oh-my-claudecode@omc plugin.
2	oh-my-openagent (omo) — code-yeongyu	A multi-platform harness that routes across 8 providers (Claude, GPT, Gemini, …) by category and bridges to Claude Code via claude-code-agent-loader and claude-code-plugin-loader. Its 9 agents (Sisyphus, Atlas, Oracle, …) are adapted here as standalone .md files.	Not a submodule: the 9 agents live in agents/omo/; install.sh runs npm i -g oh-my-opencode@latest for the omo CLI.
3	andrej-karpathy-skills — forrestchang	The 4 AI-coding behavioral guidelines — Think Before Coding, Simplicity First, Surgical Changes, Goal-Driven Execution — always active.	install.sh curls CLAUDE.md at the pinned SHA aa4467f, checksum-verifies it, and appends it to ~/.claude/CLAUDE.md. No code vendored.
4	everything-claude-code (ECC) — affaan-m	278 skills + 67 agents + 94 commands + language rules upstream. my-claude installs a curated 61-skill subset (79 with the opt-in lane) — stack patterns, AI/agent engineering, codebase tooling — plus 9 rule sets and slash commands like , , , .

---

How Boss Works

Boss is the meta-orchestrator at the core of my-claude. It never writes code — it discovers, classifies, matches, delegates, and verifies.

Phase	What Happens
0 · Discovery	Scans agents, skills, MCP tools, and hooks at runtime into a live capability registry
1 · Intent gate	Classifies the request (trivial, build, refactor, mid-sized, architecture, research, …) and counter-proposes a skill when one fits better
2 · Capability matching	Cascades the priority chain below (P0 gstack skill → P1 exact skill → P2 specialist agent → P3 multi-agent orchestration → P4 general-purpose fallback)
3 · Delegation	Sends a 6-section structured prompt to the specialist: TASK / OUTCOME / TOOLS / DO / DON'T / CTX
4 · Verification	Reads the changed files independently, runs tests, lint, and build, cross-references the original intent, retries up to 3× on failure

Priority Routing

Priority	Match Type	When	Example
P0	gstack skill	Ship / QA / deploy / security workflow	"ship this" → gstack /ship
P1	Skill match	Task maps to a self-contained skill	"merge PDFs" → pdf skill
P2	Specialist agent	Domain-specific agent exists	"security audit" → security-reviewer
P3a	Boss direct	2-4 independent agents	"fix 3 bugs" → parallel spawn
P3b	Sub-orchestrator	Complex multi-step workflow	"refactor + test" → Sisyphus
P3c	Agent Teams	Peer-to-peer communication needed	"implement + review" → Review Chain
P4	Fallback	No specialist matches	"explain this" → general agent

Model Routing

Complexity	Model	Used For
Top-level orchestration	claude-opus-5-5	Boss
Advisor group (read-only second opinions)	claude-fable-5-1	Oracle, Metis, Momus
Deep analysis, architecture	claude-opus-5-5	Sisyphus, Atlas, Hephaestus, Prometheus
Standard implementation	claude-sonnet-5	Librarian, Multimodal-Looker, OMC specialists
Quick lookup, exploration	claude-haiku-4-5	Lightweight OMC agents, simple advisory

Effort Tiers

Model choice sets which brain runs a task; the effort: frontmatter field sets how hard it thinks. Every self-owned agent declares one:

Effort	Agents
xhigh	Boss, Oracle, Prometheus, Multi-Agent Systems Architect
high	Sisyphus, Hephaestus, Atlas, Metis, Momus
medium	Librarian, Multimodal-Looker, AI Engineer, DevOps Automator

Skills declare effort too — boss-briefing runs at medium, briefing-vault at low. boss-advanced and gstack-sprint deliberately declare none: a skill's effort overrides the session level while it runs, which would quietly downgrade Boss mid-flow.

Precedence is CLAUDE_CODE_EFFORT_LEVEL (env) > frontmatter > session effort level. xhigh is Fable's supported ceiling; max is Opus-class only and silently falls back elsewhere.

3-Phase Sprint Workflow

For end-to-end feature implementation, Boss orchestrates a structured sprint:

Phase	Mode	What Happens
1 · Design	interactive	User decides scope · engineering review · confirm "design done"
2 · Execute	autonomous	ralph runs execution · auto code review · architect verification
3 · Review	interactive	Compare against the design doc · present comparison table · user approves or asks for improvement

Structured Final Report

Boss closes every working turn — one that edited files, made commits/PRs, changed configuration, or ran verification — with a structured report the reader can scan without opening a diff. Each table is emitted only when its situation actually occurred (never an empty table):

Situation	Table	Columns
Files/settings changed	Changes	Target / Before / After / Rationale
Multiple tasks completed	Work summary	Item / Result / Evidence
Verification was run	Verification	Item / Expected / Actual / Verdict
Commits/PRs produced	Deliverables	PR / Repo / Content / Status
Anything unresolved	Remaining	Item / Status / Next step

It fires only at the very end of the request — never on a turn that launches or relays background work, never as a mid-task progress update — and pure Q&A turns end normally without it. The spec lives in boss.md § FINAL REPORT; the stop-final-report.js Stop hook enforces it, judging from the harness-provided last_assistant_message plus the turn's origin in the transcript, blocking at most once per turn and failing open on any error.

Named Workflows

Deterministic multi-agent workflows. install.sh copies them to ~/.claude/workflows/, so they run from any project through the Workflow tool — not just from this repo.

Workflow	What It Does	Invocation
code-review-fanout	Four dimension reviewers (correctness, security, performance, tests) fan out in parallel, then every finding is adversarially verified before it is reported	Workflow({name: "code-review-fanout"}) — args: the review target (branch, commit range, paths). Defaults to the working-tree diff
upstream-audit	One analyst per upstream — pin delta vs origin, allowlist fit, new overlaps, security signals, health — followed by a synthesized action list	Workflow({name: "upstream-audit"}) — for quarterly or pre-sync audits

How Boss discovers agents and skills

Boss does not have to remember to read anything: routing information is pushed into its context by two hooks.

Piece	What it does
Registry v2 (hooks/build-registry.js)	At session start, scans every agent and skill Claude Code can load — ~/.claude/agents/**, ~/.claude/skills/*, the project's .claude/agents/** and .claude/skills/*, and the agents/skills of every enabled plugin (addressed as <plugin>:<name>, e.g. oh-my-claudecode:architect). Writes names, descriptions, models, and scopes (project / global / plugin:<name>) to ~/.omc/state/capability-registry.json. It rebuilds only when a source file or directory changed, the project changed, or the file predates version 2
Routing summary (SessionStart)	Injects one line per intent with the top 3 candidates, e.g. Architecture → oracle[advisor], oh-my-claudecode:architect, /architecture-decision-records. Bare names are agents (Agent tool subagent_type), /name is a skill, [advisor] marks the Advisor Group. Capped at ~1,500 tokens
Route hint (hooks/route-hint.js, UserPromptSubmit)	Classifies each prompt by keyword (English and Korean) and adds [RouteHint] intent=<X> → <top 3>. Silent when nothing matches, for slash commands, and for teammate / task-notification messages. It only reads the cached registry, so it adds no scan time to a prompt
Advisor gate (hooks/advisor-gate.js, Stop)	Backstop for boss.md § Advisor Gate. When the prompt's route-hint intent is Architecture, Ambiguity, or PlanReview, the turn made no Agent/Task call to oracle / metis / momus, and the final message has no Advisor skipped: <reason> line, it blocks once and asks Boss to consult the advisor (or state why not) and repeat its answer. Never blocks subagents, task-notification / teammate turns, a continuation after a Stop block, or the same turn twice; fails open on any error

Adding your own agents or skills: drop them in ~/.claude/agents/, ~/.claude/skills/<name>/SKILL.md, or the project's .claude/ equivalents, or enable a plugin that ships them. The next session picks them up automatically. Items not named in the routing map still become candidates when their name or description contains one of an intent's description_keywords, so a clear description is what gets them routed.

Adjusting routing: edit hooks/routing-map.json (installed to ~/.claude/hooks/). Each intent has members (preferred agents/skills in order; a bare name also matches a plugin item of that name, and missing ones are skipped), advisor: true on Advisor Group members, description_keywords (classify registry items), and prompt_keywords (classify prompts; the highest keyword count wins and ties go to the intent listed first). Set REGISTRY_OUT to write or read the registry somewhere else, e.g. for a dry run.

Adoption: routing learns which results you keep

Boss records whether you adopt what an agent or skill produced, per intent, and uses it — gated — to reorder routing candidates. Call counts are never used: running an agent often says nothing about whether its output was any good.

Piece	What it does
Offer (hooks/adoption-tracker.js offer, PostToolUse Agent/Task/Skill)	Remembers the agents/skills the main session ran in the latest turn (one per item, 5 max) with the intent route-hint.js classified for that prompt. Subagents' own tool calls are ignored. A background agent or teammate is judged only after its notification has arrived. Process skills listed in adoption_ignore in hooks/routing-map.json (using-superpowers, boss-briefing, briefing-vault, cancel, artifact-design, …; a plugin-qualified id matches by its short name) are never offers: they run regardless of whether the advice was good
Verdict (hooks/adoption-tracker.js verdict, UserPromptSubmit)	Classifies your next reply: accept (진행해, 반영해, 좋아, 그렇게 해, 승인, 머지해, go ahead, yes, lgtm, apply it, ship it, …) or reject (아니, 틀렸, 다시 해, 그만, 하지 마, 되돌려, no,, wrong, redo, revert, stop, …); reject wins when both match, a question (…?) is never an accept, and anything else is neutral (no event). Slash commands and task-notification / teammate messages are not replies. Prints nothing
Ranking (adoptionWeight in hooks/build-registry.js)	Events from the last 180 days, each weighted 0.5^(age_days/90). A (id, intent) pair counts only once its weighted sample reaches 5; its accept rate then moves it up or down by at most two slots within its band — explicit routing-map members always stay ahead of keyword matches. Security and Ship keep the map order and ignore adoption and pins. The SessionStart summary shows (adopted 7/9) once a pair is counted, and [pinned] for pins

The store is shared by harnesses (my-codex can write to the same files) under ~/.config/agent-harness/:

• adoption-ledger.jsonl — append-only, one event per line: {"ts": ISO-8601, "harness": "claude", "session": id, "kind": "agent"|"skill", "id": exact subagent_type or skill name, "intent": route-hint intent or "unknown", "verdict": "accept"|"reject", "signal": "reply"|"choice"|"revert", "evidence": first 120 chars of your reply}. signal is revert when the reply asked for a revert/rollback, reply otherwise; choice is reserved for answers to an explicit choice prompt.
• adoption-pins.json — [{"id", "intent", "ts"}]; a pin forces that item to the top of that intent.
• adoption-archive.jsonl / adoption-audit.jsonl — where reset/undo move events (nothing is hard-deleted), and a log of every CLI mutation.

Inspect or correct it with node ~/.claude/hooks/persona-rule.js adoption <cmd>: list [--intent X] (id × intent table of weighted accept/reject/n), list --events (raw events with their ts), pin <id> <intent>, unpin <id> <intent>, reset [<id>], undo <ts> [<id>]. The ledger and pins are registry sources, so the next session's routing summary reflects any change.

Learning loop (approved by you)

Over time, Boss turns what you correct and what you keep adopting into rules and skills, but only the ones you approve. Detection runs without an LLM call; nothing is written to your rules or skills until you run learn approve.

Piece	What it does
Review (hooks/learning-review.js, SessionEnd)	Reads the session transcript and the adoption ledger, prints nothing, and queues suggestions. Rule: a message of yours, after Boss has replied at least once, that states a preference or correction: English don't …, never …, always … at the start of a sentence, or stop doing, from now on, instead of anywhere; Korean ~지 마, ~지 말고, 다음부터 anywhere, or 항상 / 앞으로 / 대신 / 말고 with an imperative ending (…해줘, …써). It skips questions, code blocks, inline code, quoted lines, slash commands, notifications, and messages over 1,500 characters. It queues at most 2 per session, always/앞으로-style phrasing first, and records the agent or skill that ran just before. Skill: an ordered chain of 2–4 agents/skills you accepted back to back in 3 or more sessions within 30 days (a reject in between breaks the chain)
Queue	Suggestions are deduped by kind plus normalized text or steps. A key you approved or dismissed is never suggested again. At most 5 are pending: new ones beyond that are refused, and each refusal is audited as cap_refused
Surfacing (hooks/session-start.sh)	Adds at most two [Learn] <id>: <kind> — <text>. Approve: … learn approve <id> / dismiss: … learn dismiss <id> lines. Boss asks you about them, at most 2 per session (boss.md Phase 0 Step 4b)
Approval	A rule goes to ~/.claude/rules/user/learned-<slug>.md, which has no paths frontmatter, so it loads in every session, with your original sentence quoted (--as "<English rule>" adds an imperative English line). A skill goes to ~/.claude/skills/learned-<slug>/SKILL.md: an ordered procedure whose description names its intent, so the registry routes it. Both paths are user-owned: install.sh never lists them in its manifest and never deletes them. At most 20 learned items can be live; approving beyond that is refused until you curate or roll back
Curator (learn curate, also silently once every 7 days at SessionStart)

State lives next to the adoption store in ~/.config/agent-harness/: learning-suggestions.jsonl (id, kind, status pending|approved|dismissed, key, created_at, and the text or steps), learning-items.json (approved items, their status, pin, and last use), learning-audit.jsonl (one row per mutation, ids A1, A2, …, with before/after state and paths), and learned-archive/.

Commands: node ~/.claude/hooks/persona-rule.js learn <cmd>: list, show <id|slug>, approve <id> [--as "<rule>"], dismiss <id>, pin <slug>, unpin <slug>, curate [--dry-run], rollback <audit-id>. Rollback undoes one mutation: an approval moves the file to the archive and puts the suggestion back to pending; an archive moves the file back; a dismiss, pin, or status change is reversed. It refuses when a later mutation changed the same item.

How this differs from Hermes Agent: the design follows NousResearch/hermes-agent. A background review looks for your corrections, a curator ages agent-created skills by last use and not by popularity, suggestions go pending → accepted/dismissed with at most 5 pending, every mutation is ledgered and can be rolled back, and size caps refuse writes instead of evicting. The difference is that Hermes rewrites its skills and memory on its own, while here nothing reaches your rules or skills without your explicit learn approve.

---

What's Inside

Category	Count	Source
Agents (always loaded)	32	Boss 1 + OMO 9 + OMC 19 + Vendored 3
Skills	107	ECC 61 · gstack 27 · Superpowers 14 · Core 4 · Archify 1. ECC's 18-skill web lane is opt-in (--skills=web); OMC's 13 come from the OMC plugin and are never copied
Rules	48 files / 9 sets	ECC 46 (3 common files + 8 language dirs) + Core 2
MCP Servers	3	Context7, Exa, grep.app
Hooks	10 files / 6 events	Delegation guard, telemetry, verification, vault, context budget
LSP Servers	2	typescript (typescript-language-server), python (pyright-langserver)
Named Workflows	2	code-review-fanout, upstream-audit
Upstream submodules	5	ecc, omc, gstack, superpowers, archify
CLI Tools	7	omc, omo, ast-grep, comment-checker, codeburn, serena, headroom

Every agent, skill, and rule above is allowlisted in scripts/skill-allowlists.sh and tracked in the install manifest. Anthropic's official document skills (pdf, docx, …) are added separately via claude plugin add anthropics/skills and are deliberately not manifest-tracked.

Per-agent models are in Model Routing; where each source comes from is in Open-Source Tools Used.

Agent	Tier	Role	Source
Boss	core	Dynamic runtime discovery → capability matching → optimal routing. Never writes code.	my-claude
Sisyphus	omo	Intent classification → specialist delegation → verification	oh-my-openagent
Hephaestus	omo	Autonomous explore → plan → execute → verify	oh-my-openagent
Atlas	omo	Task decomposition + 4-stage QA verification	oh-my-openagent
Oracle	omo	Strategic technical consulting (read-only)	oh-my-openagent
Metis	omo	Intent analysis, ambiguity detection	oh-my-openagent
Momus	omo	Plan feasibility review	oh-my-openagent
Prometheus	omo	Interview-based detailed planning	oh-my-openagent
Librarian	omo	Open-source documentation search via MCP	oh-my-openagent
Multimodal-Looker	omo	Image/screenshot/diagram analysis	oh-my-openagent
analyst	omc	Pre-analysis before planning	oh-my-claudecode
architect	omc	System design and architecture	oh-my-claudecode
code-reviewer	omc	Focused code review	oh-my-claudecode
code-simplifier	omc	Code simplification and cleanup	oh-my-claudecode
critic	omc	Critical analysis, alternative proposals	oh-my-claudecode
debugger	omc	Focused debugging	oh-my-claudecode
designer	omc	UI/UX design guidance	oh-my-claudecode

Source	Count	Key Skills
everything-claude-code	61	coding-standards, fastapi-patterns, agent-architecture-audit, springboot-patterns, kubernetes-patterns
gstack	27	/qa, /review, /ship, /cso, /investigate, /office-hours
superpowers	14	brainstorming, systematic-debugging, test-driven-development, writing-plans
my-claude Core	4	boss-advanced, boss-briefing, briefing-vault, gstack-sprint
archify	1	archify

Opt-in web lane (+18) — accessibility, bun-runtime, e2e-testing, frontend-a11y, frontend-patterns, motion-* (3), nestjs-patterns, nextjs-turbopack, nuxt4-patterns, react-patterns, react-performance, react-testing, ui-to-vue, vite-patterns, vue-patterns, windows-desktop-e2e. Browser-UI work only; see Installation for the flags.

OMC skills are not copied. install.sh enables the oh-my-claudecode@omc plugin, which already exposes all 13 (autopilot, ralph, team, execute, ralplan, verify, …) as oh-my-claudecode:<name>. A local copy would put a second description of the same skill in every session. Routing surfaces use the prefixed name; a bare ralph does not resolve.

MCP Servers — what each one does and how it is registered is in Open-Source Tools Used.

Server	Cost
Context7	Free
Exa	Free 1k req/month
grep.app	Free

Behavioral Hooks

Hook	Event	Behavior
Session Setup	SessionStart	Auto-detects missing tools + injects Briefing Vault context
Delegation Guard	PreToolUse	Blocks Boss from directly modifying files
Agent Telemetry	PostToolUse	Logs agent usage to agent-usage.jsonl
Subagent Logger	SubagentStop	Logs agent execution to Briefing Vault
Completion Check	Stop	Runs profile fallback + guards /boss-briefing execution
Vault Reminder	UserPromptSubmit	Suggests /boss-briefing after 5+ messages
Calibrated Response	UserPromptSubmit	Re-injects rules/common/calibrated-response.md every turn
Context Budget	UserPromptSubmit	Every 40 prompts since the last compaction (MY_CLAUDE_COMPACT_EVERY), suggests /compact at the next task boundary
Context Budget reset	SessionStart (compact)	Zeroes that counter after a compaction

Quiet plugin hooks. merge-settings.js sets env.OMC_QUIET=2 unless you already set it. OMC's own plugin hooks otherwise append an advisory line ("Use parallel execution…", "Background operation detected…") to nearly every Bash, Edit, and Read call, and every one of those lines is re-sent as context on every later request. Level 2 drops the advisories and the "Completed: N" agent summaries; real failures are still reported.

Earlier auto-compaction. merge-settings.js also sets env.CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=75 unless you already set it, so a long session compacts before the tail of its transcript gets expensive. The variable can only lower the trigger. See the environment variable reference.

The plugin declares two language servers in .lsp.json. Claude Code starts them on demand, so agents get diagnostics and code navigation instantly instead of paying for a build round-trip. install.sh installs both binaries non-fatally — if one is unavailable, only that server is disabled and the rest of the install continues.

Server	Command	Extensions
typescript	typescript-language-server --stdio	.ts .tsx .js .jsx .mjs .cjs
python	pyright-langserver --stdio	.py

---

Briefing Vault

Obsidian-compatible persistent memory. Every project maintains a .briefing/ directory that auto-populates across sessions. Every path below is created and maintained by hooks:

Path	Description
INDEX.md	Project overview with links to recent decisions and learnings. Auto-created on first session, refreshed by /boss-briefing or Stop hook fallback.
state.json	Session metadata: counters (workCounter, messageCount), lastVaultSync timestamp, sessionStartHead. Auto-managed by hooks.
sessions/	Session summaries, named YYYY-MM-DD-<topic>.md. YYYY-MM-DD-auto.md — scaffold with git diff stats and agent counts. <topic>.md — AI-written summary enforced by Stop hook guard.
decisions/	Architecture and design decisions with rationale, as YYYY-MM-DD-<decision>.md. AI-written, enforced during active work.
learnings/	Patterns, gotchas, non-obvious solutions, as YYYY-MM-DD-<pattern>.md. *-auto-session.md — scaffold with file lists. <topic>.md — AI-written.
references/	Web research URLs. auto-links.md — auto-collected from WebSearch/WebFetch calls.
agents/	Agent telemetry. agent-log.jsonl — per-call log with enriched fields {ts, agent_type, phase, seq, task_hint}. YYYY-MM-DD-summary.md — daily usage breakdown.
persona/	User work style profile. profile.md — tool affinity stats. suggestions.jsonl — routing recommendations. Workflow sequence patterns in rules/workflow-*.md. Run /boss-briefing to analyze.
archives/	PARA Archives (flat) — completed sessions (30+ days), superseded decisions, inactive learnings
wiki/	LLM-wiki concept pages (_schema.md holds the page schema) — distilled knowledge from multiple sessions

Knowledge Management (v2)

BriefingVault v2 integrates three knowledge management methodologies:

Methodology	Applied As
PARA (Tiago Forte)	Directory structure: sessions=Projects, decisions=Areas, references=Resources, archives=Archives
Zettelkasten (Luhmann)	Atomic notes in learnings/, unique IDs (YYYYMMDDHHMMSS), enforced [[wiki-links]]
LLM-wiki (Karpathy)	Concept pages in wiki/ — auto-suggested when keywords appear 3+ times

Claude Code session-end hooks automatically:

• Suggest archiving notes older than 30 days
• Propose wiki pages for frequently mentioned concepts
• Generate unique Zettelkasten IDs for new notes

Session-Specific Diffs

At session start, the current git HEAD is saved to state.json (sessionStartHead field). For non-git projects, a YYYY-MM-DD:cwd identifier is used instead. At session end, diffs are calculated relative to this saved point — showing only changes from the current session, not accumulated uncommitted changes from previous sessions.

Using with Obsidian

1. Open Obsidian → Open folder as vault → select .briefing/
2. Notes appear in graph view, linked by [[wiki-links]]
3. YAML frontmatter (date, type, tags) enables structured search
4. Timeline of decisions and learnings builds automatically over sessions

/boss-briefing

Run /boss-briefing during or at the end of a session to:

• Sync vault: Update profile.md, INDEX.md, and agent summaries
• Detect workflow patterns: Analyze temporal agent call sequences across sessions
• Recover from gaps: Generate recovery summaries if days have passed since the last session
• Propose persona rules: Suggest workflow-based routing preferences (not just frequency)
• Validate session notes: Check that today's session has a proper summary

The Stop hook checks whether /boss-briefing has run today. If not, it blocks session end with a reminder. The existing stop-profile-update.js continues to run as a fallback. A second Stop hook, stop-final-report.js, enforces the Structured Final Report.

---

Where to See Results

Every tool in this stack writes its output somewhere. This is where. What each tool is, is in Open-Source Tools Used.

Tool	Open	How to Run	Where to Look
codeburn	http://127.0.0.1:4747/	install.sh starts or reuses codeburn web --provider all --port 4747 --no-open · codeburn opens the TUI · codeburn report --format json --period week emits a non-interactive dump (also --day, --from/--to, --provider claude)	The shared dashboard; startup log ${XDG_STATE_HOME:-$HOME/.local/state}/agent-harness-services/logs/codeburn.log. Session files are read-only and dollar figures are estimates at API list rates — on a subscription plan a usage proxy, not a bill.
codeburn guard	codeburn guard status	Installed by install.sh by default (--no-codeburn-guard to skip) via codeburn guard install --global	The soft cap ($5) and checkpoint nudge ($3) stay on; the hard cap is written as null right after install, since its cost estimate is list-price and would otherwise stop a normal subscription session in minutes. Set your own hard cap by editing hardUSD in ~/.config/codeburn/guard.json.
Serena	http://localhost:24282/dashboard/index.html	Starts automatically as an MCP server; call get_symbols_overview / find_symbol from any session	The dashboard, while a server is running (logs + per-tool call counts). Per-project memories land in .serena/ inside the repository you are working on; the global config is ~/.serena/serena_config.yml.
Headroom	http://127.0.0.1:8787/stats	MCP tools headroom_compress / headroom_retrieve / headroom_stats; install.sh starts or reuses the shared agent-harness-shared proxy profile	Compression stats, empty until a client explicitly routes through the proxy; startup log .

---

GitHub Actions

Workflow	Trigger	Purpose
CI	push, PR	Validates JSON configs, agent frontmatter, skill existence, upstream file counts
Smoke	push, PR	4 jobs — hooks (hook execution), shell (install script shape), drift (model drift), routing-refs (dangling agent/skill references)
Update Upstream	every 3 days / manual	git submodule update --remote → refresh SHA pins in SOURCES.json → security scan of the upstream diff → auto-merge only when the scan is clean, otherwise the PR is flagged for human review
Auto Tag	push to main	Reads plugin.json version and creates git tag if new
Pages	push to main	Deploys docs/index.html to GitHub Pages
CLA	PR	Contributor License Agreement check
Lint Workflows	push, PR	Validates GitHub Actions workflow YAML syntax

---

my-claude Originals

Features built specifically for this project, beyond what upstream sources provide:

Feature	Description
Boss Meta-Orchestrator	Dynamic capability discovery → intent classification → 5-priority routing → delegation → verification
3-Phase Sprint	Design (interactive) → Execute (autonomous via ralph) → Review (interactive vs design doc)
Agent Tier Priority	core > omo > omc > vendored deduplication. Most specialized agent wins.
Lane Ownership	Orchestration → OMC, dev process → superpowers, ship/QA/deploy/security → gstack (Boss P0), language and stack knowledge → ECC, AI and domain work → vendored agents
Curated Allowlists	scripts/skill-allowlists.sh is the single source of truth — 107 skills, 3 always-on common rules and 8 path-scoped language rule sets survive from thousands upstream, so nothing unlisted ever reaches a session's context
Briefing Vault	Obsidian-compatible .briefing/ directory with sessions, decisions, learnings, references
Agent Telemetry	PostToolUse hook logs agent usage to agent-usage.jsonl
No-op Sync Skip	Upstream sync stages the submodule bumps and SOURCES.json pins, then opens a PR only when that staged diff is non-empty
Agent Dedup Detection	tests/validate-sync.sh compares agent filenames across agents/ and the omc/superpowers submodules and reports collisions

---

Bundled Upstream Versions

Linked via git submodules. Pinned commits are tracked natively by .gitmodules and mirrored as an AI-BOM in upstream/SOURCES.json; install.sh checks out these exact SHAs rather than tracking main.

Source	SHA	Date	Diff
everything-claude-code	4eb71d9	2026-10-10	compare
oh-my-claudecode	454bae0	2026-10-10	compare
gstack	20eb620	2026-10-10	compare
superpowers	bb92a77	2026-10-10	compare
archify	2ab3cae (v3.0.1)	2026-09-28	compare

---

Contributing

Issues and PRs are welcome. When adding a new agent, add a .md file to agents/core/ or agents/omo/ and update SETUP.md.

Credits

Built on the projects listed in Open-Source Tools Used; thank you to every author.

License

MIT License. See the LICENSE file for details.