archon-cli

Archon CLI

Archon runs multi-step AI workflows through the archon CLI. Git projects run in isolated worktrees by default; registered folder projects run in place. You sit on top of the runs: you write the brief, launch, watch one stream, answer the decisions that reach you, and read what each run left behind.

Route by intent:

The user wants to...	Read
Run a workflow on real work	operate/running.md
Inspect, answer, cancel or resume runs	operate/managing-runs.md
Drive the SDLC pack's queues	operate/queues.md
Find out why a run failed or hangs	operate/troubleshooting.md
Set up Archon or change config	setup/setup-and-config.md
Author or change a workflow	authoring/authoring-workflows.md (then node-reference.md and variables.md in the same folder)
Improve a node prompt or a run message	authoring/prompting-mistakes.md

• When the intent is clear ("build me a workflow for X", "run archon-ship on issue #42"), go straight to the reference. Do not ask.
• When it is ambiguous, ask which of these the user wants.
• On a machine where archon doctor fails, start with setup/setup-and-config.md.

The operating loop

1. Discover. archon workflow list --json gives compact descriptions. For each plausible candidate, read the full description with archon workflow list <name> --full. Choose from full descriptions, never from names you remember.

2. Check the brief. The message, issue body or document the run reads is the contract the run is measured against. It must state the problem, why it is worth solving, why now, the desired outcome, the invariants and the acceptance. If any is missing, say which, propose wording, and get the user's agreement before launching. operate/running.md has the full check.

3. Launch detached. Pass the work as a message, the workflow's declared inputs with --input, or both:

   archon workflow run <workflow> --branch <branch-name> "<the work>" --detach
   archon workflow run <workflow> --input <name>=<value> --input <name>=<value> --detach
   

4. Watch once. Run one stream as a background task of your harness, however many runs you launched:

   archon workflow attention --watch --json        # this project; add --all for the install
   

   Each line is JSON. attention means a run needs a decision, resolved means it no longer does, and terminal means a run ended. Act on the line, then keep reading.

5. Answer deliberately. Read the item's question, its recommendation and the evidence it links. Answer with a reason, detached: archon workflow respond <run-id> <decision> "<reason>" --detach. operate/managing-runs.md covers every form.

6. Read the result. A completed run is not a verdict. archon workflow get <run-id> --json gives the normalized outcome and leave_behind.artifactFiles. Read the report before telling the user what the run concluded.

Hard rules

• Never launch on a thin brief. A weak input does not produce a weak result. It produces a confident, well-formed answer to the wrong question, at full price.
• Launch and answer with --detach. A detached run starts its own session and survives your shell. A foreground respond or resume continues the run inside your shell, and closing the shell stops it.
• One watcher, not one per run. One workflow attention --watch stream covers every run in scope, including runs that triggers or other sessions start. A workflow wait <run-id> per run is fine for one or two runs; with dozens it contends for the database.
• Stop work only when you mean to end it. archon workflow cancel stops the run mid-node, and that node's provider session and unsaved work are lost. Let a run reach its next pause or end when you only want to stop launching more. When one turn is stuck on a tool or background command that never finishes, archon workflow end-turn <run-id> ends that turn and keeps the run (operate/managing-runs.md, "A stalled turn").
• Merge only what was reviewed. When you merge a pull request by hand, pin the merge to the reviewed head (gh pr merge <n> --match-head-commit <sha>), read the open review threads first, and sync your checkout afterwards.
• Author agent-first. In a workflow you write, an agent node with typed output judges, scripts read facts and act at boundaries, and YAML only wires them. authoring/authoring-workflows.md ("Agent-first") has the rule and an example.

Gotchas

• The current directory selects the project for discovery, launches and project listings. workflow status, workflow runs and workflow attention default to that project; add --all for the whole install. For a git project, run from the repo root. Register a non-git folder with workflow run --folder. A full run ID works from any directory.
• Attention items are paused runs. A running run stuck on a stalled tool call is not an item: workflow wait <run-id> and workflow status show it.
• workflow status and workflow runs set scopeFallback: true in JSON when an unregistered project produced an install-wide result.
• Prefer --json whenever you will parse output.
• archon --help is authoritative. If a flag this skill names is rejected, trust the CLI's help and tell the user about the drift.