Skill DAG

A minimal SQLite-backed DAG runner (pip install skill-dag). Define nodes (shell commands) and dependency edges, then run the graph in topological order. Every run and per-node result is recorded in a plain SQLite file — no servers, no background services, no external dependencies beyond the Python standard library.

It doubles as a skills runner: the repo ships panel skills (debate, brainstorm, review, reframe) that deliberate in rounds on the DAG, with gates that fail closed. The skillflow command, module, and MCP tools keep their names.

Install

Requires Python 3.10+.

pip install skill-dag
# with the MCP server:
pip install "skill-dag[mcp]"

Or from a checkout (needed for the panel skills and panel/run.sh):

pip install .
pip install ".[mcp]"

Use

skillflow init
skillflow add-node fetch --cmd "curl -s https://example.com -o page.html"
skillflow add-node parse --cmd "python parse.py page.html"
skillflow add-edge fetch parse
skillflow run
skillflow status

Options:

• --db PATH (before the subcommand, e.g. skillflow --db demo.db run) or SKILLFLOW_DB selects the SQLite file. Defaults to ./skillflow.db.
• show prints the DAG as JSON. remove-node NAME [--force] and remove-edge FROM TO edit it (recorded history needs --force).
• status [--run ID] [--json] shows the latest run (or a given one). runs lists every run; diff A B compares two runs by id.
• run exits nonzero if any node fails; downstream nodes do not execute after a failure. --jobs N runs independent nodes concurrently, --from NAME resumes from one node downstream, --retry re-runs from the latest run's first failure, and --dry-run prints the plan only.
• add-node takes --timeout (seconds), --env (JSON object), and --cwd per node. Stored output per node is capped; files carry bulk.
• demo runs a self-contained three-node graph. See examples/ for runnable scripts (chain, parallel, retry).

How it works

Four tables: nodes, edges, runs, node_results. Execution order is computed with Kahn's algorithm; cycles are rejected when an edge is added (and re-checked at run time). Node output (stdout + stderr), exit codes, and timestamps are stored per run.

Panel skills

Debate, brainstorm, review, reframe, and add-skill run in the active conversation. Their prose methods do the intellectual work; a local skillflow DAG enforces the order and makes the session pause before moving on. Nothing launches another model.

• Debate: ground → activate → crossfire → continue/finish → final.
• Brainstorm: ground → activate → divergent field → activate → develop field → final.
• Review: ground → activate → intent → activate → evidence/deltas → final.
• Reframe: ground → activate → stronger-shape field → lineup → continue/finish → final. Generation and lineup have separate pauses.
• Add-skill: ground → draft → record → final. Authors a new skill.

The session selects and activates lenses from the roster, shows the prose in the conversation, and authors every result. The DAG does not select the room, extract semantic tensions, decide whether an argument is good, or synthesize the answer. When prior decisions, corrections, rejections, or failures matter, the active session uses bounded semantic work-history recall during grounding. The current instruction and live sources govern. rewind archives affected artifacts and invalidates their dependent checkpoints when later evidence changes an earlier phase. A debate/reframe refusal still requires an honest final answer.

bash panel/run.sh debate "ship it friday" 3 /tmp/my-debate
# PAUSE ground: do the grounding in the current conversation; write ground.md.
bash panel/run.sh resume /tmp/my-debate
# PAUSE activate-1: perform that phase, save it, then resume again.

Each new checkpoint returns control before accepting its artifact, even if a file was prefilled. Exit 1 with PAUSE means the current session should do the named work and resume, not ask a human to fill a file or approve the next phase. Do not batch-author future phases. final.md is session-authored, never a machine concatenation. The receipts prove sequence, not quality. The final output is the complete useful room, not a checklist of its conclusions.

Brainstorm/review have two mandatory phases. Debate/reframe accept a 1–8 round ceiling (default 3); the session decides whether further rounds are worthwhile. status <dir> reports the current gate; chain <dir> <skill> starts a new skill carrying the prior final as evidence. Full method and recovery instructions are in the shared protocol.

Install the five skills with their shared prose and launcher:

python3 scripts/install_panel_skills.py --skills-dir ~/.codex/skills
# Another local skill root can be passed instead, or with another --skills-dir.

The installed launcher resolves the checkout from a local location file, so it works from any repo. Reinstall if the checkout moves. No harness settings, other skills, or authentication are modified. Copying a lone SKILL.md is insufficient.

Existing session DBs retain their graphs and can still use skillflow run from that session directory. All five skills run the same hybrid loop. The selector and seed tools remain available for explicit standalone use and old sessions; the conversational skills no longer depend on them.

MCP server

Everything above is also an MCP server (stdio). Fourteen tools: skillflow_init, skillflow_add_node, skillflow_add_edge, skillflow_remove_node, skillflow_remove_edge, skillflow_show, skillflow_run, skillflow_status, skillflow_runs, skillflow_session_status, skillflow_session_artifact, panel_seed, panel_select_room, panel_record_seating.

pip install "skill-dag[mcp]"
python -m skillflow.mcp_server

Example client config (stdio):

{
  "mcpServers": {
    "skillflow": {
      "command": "python3",
      "args": ["-m", "skillflow.mcp_server"],
      "cwd": "/path/to/skillflow"
    }
  }
}

The roster ships in the wheel, so the panel tools work from an installed copy. Only the run.py skills need a repo checkout (or an installed skills directory pointing at one).

Note: panel round nodes never prompt. A round with no response stops the run at that stage boundary — a run containing an unwritten round stops there, by design, and resumes when the response is written.

Trust boundary: this server executes arbitrary shell commands from the DAGs you define (skillflow_run is annotated destructive for exactly that reason). Run it locally, for your own agents only — do not expose it to untrusted clients or networks.

Client wiring: Codex + Muse

Codex (~/.codex/config.toml, TOML):

[mcp_servers.skillflow]
command = "python3"
args = ["-m", "skillflow.mcp_server"]
cwd = "/path/to/skillflow"

Muse (~/.config/muse/settings.json, JSON under mcpServers):

{
  "mcpServers": {
    "skillflow": {
      "command": "python3",
      "args": ["-m", "skillflow.mcp_server"],
      "cwd": "/path/to/skillflow"
    }
  }
}

cwd is optional: without it the server runs engine and panel tools from the installed package. Point it at a repo checkout only when sessions must resolve repo-relative paths. Muse documents streamable-HTTP entries; the stdio entry above is confirmed working.

Develop

python -m unittest discover -t . -s tests -v
cd panel && python -m unittest test_select -v

License

MIT. See LICENSE.