Author Workflow

This skill owns workflow + team-formation decisions for a task. Use it both when the user has a concrete task ("add OAuth2 auth") and needs a recommendation, and when they want to build a brand-new workflow.

Modes (pass exactly one flag)

Argument autocomplete does not show flag variants, so the mode menu is always in the body — pick one before proceeding.

--for-task "<description>" — start here when the user describes a task, not a pipeline. Recommend which existing workflow fits and which teams should be on which nodes; fall back to --new only if nothing fits. No files written unless you actually author a workflow.
--new [description] — start here when the user wants a pipeline that does not yet exist. Guided authoring: DAG design, workflow YAML, command prompts, validation, commit.

If invoked with no flag and no clear signal, ask the user which mode fits. Defaulting silently to --new has the user author a workflow when an existing one would have done the job.

Context

Load CLAUDE-CONTEXT-workflows.md before either mode proceeds — it contains the workflow YAML schema, team manifest schema, command prompt format, and all validation rules.

Mode: `--for-task "<description>"`

Step 1: Analyse the task

Read the task description. List current teams and workflows:

ls .archon/teams/*.yaml 2>/dev/null
ls .archon/workflows/*.yaml 2>/dev/null

For each existing workflow, extract its purpose (the description: field) and its node list (the nodes: array). For each team, extract its agent roster.

Step 2: Recommend a formation

Recommend a workflow + team-to-node mapping. Worked example:

Task: "Add OAuth2 authentication to the API"

Recommended formation:

  Workflow: sdlc-feature-development

  Node          Team                     Notes
  ──────────── ──────────────────────── ────────────────────
  plan          general-purpose          architecture + planning
  implement     dev-team-python          primary dev team
  validate      dev-team-python          validation pipeline
  security      security-review-team     auth = security-sensitive
  architecture  general-purpose          architecture review
  quality       general-purpose          code quality review
  synthesise    general-purpose          unified summary

  This task involves authentication, so I've included the
  security-review-team on the security review node.

  (a) Accept and run this now
  (b) Modify — change team assignments
  (c) Show alternative workflow templates
  (d) No existing workflow fits — switch to --new mode
  (e) Skip — I'll assign teams myself

Step 3: Offer to run

If the user picks (a), hand off directly:

/sdlc-workflows:workflows-run <workflow-name>

No file is written — this recommendation flow exists to route the user to the right existing workflow, not to persist task plans. The git commit of the run itself (via workflows-run) captures the record.

If (b), loop back to step 2 with the adjusted mapping. If (c), list the available workflows and restart step 2 against the chosen one. If (d), switch to --new mode below with the task description carried through. If (e), end the skill.

Mode: `--new [description]`

Step 0. Load the schema reference

Do not rely on memory. Explicitly read the canonical schema reference before drafting any YAML:

Read CLAUDE-CONTEXT-workflows.md

This file lives at the repo root and defines every field used in the workflow and manifest YAML, plus the command-prompt format. Skipping this step produces workflows that silently drift from the current schema.

1. Understand the Goal

If the user provided a description as an argument, use it. Otherwise ask:

What should this workflow accomplish? Describe the pipeline: what work needs to happen, in what order, and which teams should do it.

2. Discover Existing Teams

List available team manifests:

ls .archon/teams/*.yaml 2>/dev/null

For each manifest, show the team name, description, and available agents:

for f in .archon/teams/*.yaml; do
    python3 -c "
import yaml
m = yaml.safe_load(open('$f'))
print(f\"  {m['name']}: {m.get('description', 'no description').strip()}\")
print(f\"    agents: {', '.join(m.get('agents', []))}\")
"
done

If no teams exist, suggest creating them first:

No team manifests found in .archon/teams/. Create teams first with /sdlc-workflows:manage-teams --create.

3. Design the DAG

Based on the user's description and available teams, design the workflow DAG:

Identify distinct stages (implement, review, validate, etc.)
Assign each stage to a team image
Determine dependencies (what must complete before what)
Identify parallel opportunities (nodes that can run concurrently)

Present the DAG to the user for approval:

Proposed pipeline:
  1. implement (sdlc-worker:dev-team) — no deps
  2. security-review (sdlc-worker:review-team) — after implement
  3. architecture-review (sdlc-worker:review-team) — after implement
  4. synthesise (sdlc-worker:dev-team) — after both reviews

Nodes 2 and 3 run in parallel.

4. Generate Workflow YAML

Create .archon/workflows/<name>.yaml following the schema in CLAUDE-CONTEXT-workflows.md:

name: must be unique across all workflows in the directory
description: should explain purpose, inputs, and outputs
provider: claude always
Each node needs: id, command, image, and depends_on where applicable
Use trigger_rule: all_success (default) unless the user specifies otherwise
timeout: is in milliseconds. Archon's default is 120000ms (2 min) — too short for real work. Set explicitly on every node: review/synthesis 600000 (10 min), validation 300000 (5 min), implementation/planning 1800000 (30 min)
Do not add non-standard fields to nodes (e.g. name:, label:) — Archon's schema is strict and unknown fields break the graph renderer

Starter template (copy, rename, adjust):

name: <workflow-name>
description: <one-line purpose>
provider: claude
nodes:
  - id: implement
    command: implement
    image: sdlc-worker:dev-team
  - id: review
    command: review
    image: sdlc-worker:review-team
    depends_on: [implement]

5. Generate Command Prompts

For each node, create .archon/commands/<command-name>.md following the format in CLAUDE-CONTEXT-workflows.md:

Role framing: "You are performing X on a Y at /workspace."
Context: what files exist, what state to expect
Task: numbered steps of what to do
Output: where to write results
Commit: cd /workspace && git add -A && git commit -m "<message>"

6. Validate

Run validation checks:

# Name-collision check — refuse to overwrite an existing workflow
if [ -f ".archon/workflows/<name>.yaml" ]; then
    echo "ERROR: Workflow '<name>' already exists at .archon/workflows/<name>.yaml"
    echo "Choose a different name or remove the existing file first."
    exit 1
fi

# Validate that referenced team images have manifests
for team in <list of unique images>; do
    team_name="${team#sdlc-worker:}"
    if [ ! -f ".archon/teams/${team_name}.yaml" ]; then
        echo "WARNING: No manifest for $team_name — run build-team.sh after creating one"
    fi
done

# Check command files exist
for cmd in <list of unique commands>; do
    if [ ! -f ".archon/commands/${cmd}.md" ]; then
        echo "ERROR: Missing command file: .archon/commands/${cmd}.md"
    fi
done

# Validate workflow YAML structure
python3 -c "
import yaml
wf = yaml.safe_load(open('.archon/workflows/<name>.yaml'))
assert 'name' in wf, 'Missing name field'
assert 'nodes' in wf, 'Missing nodes field'
for node in wf['nodes']:
    assert 'id' in node, f'Node missing id'
    assert 'command' in node or 'prompt' in node, f'Node {node[\"id\"]} needs command or prompt'
print('Workflow YAML valid')
"

# Repository-wide cross-reference check — only runs if the validator is present
if [ -f "tools/validation/check_workflow_teams.py" ]; then
    python3 tools/validation/check_workflow_teams.py
fi

7. Commit

git add .archon/workflows/<name>.yaml .archon/commands/<command-files>
git commit -m "feat(workflow): add <name> workflow with <N> nodes"

8. Report

Workflow created:
  File: .archon/workflows/<name>.yaml
  Nodes: <count>
  Teams: <list>
  Commands: <list of .md files>

To build team images:    /sdlc-workflows:deploy-team --name <team>
To run the workflow:     /sdlc-workflows:workflows-run <name>

Designing long-running or cyclical workflows

Authors regularly want pipelines like designer → developer → reviewer → designer … running for hours or days. Before choosing a shape, ask the user these questions and design accordingly:

How long will each node actually take? Real implementation or review work is often 20-60 min, not the 2-min default. Set timeout: in milliseconds explicitly on every node — Archon's default is 120000ms (2 min) which kills real work. Use 600000 (10 min) for reviews, 1800000 (30 min) for implementation. Proven by dogfood: security-review on a 170-file repo took 1m44s and would have been killed at the 2-min default.
Is the iteration count known up front? If yes (e.g. "three review rounds"), unroll as distinct nodes (review-v1, review-v2, review-v3). If no (open-ended "until approved"), the workflow alone cannot express it — you will need an outer-loop wrapper that re-invokes archon workflow run until a signal is detected.
Does the cycle need different specialists each pass? If yes, unrolled iterations or an outer loop are the only options — a single loop: node keeps one team running the whole cycle.
Does the user want live feedback during long nodes? Flag the monitoring gap — workflows-run does not stream live output in v1. They will need a second terminal with docker logs -f or tail -f the archon log.
Is the token cost acceptable? A 30-min node × 4 stages × 5 iterations on Opus can run to serious dollars. Warn the user before shipping the workflow. Set budget: on each node to cap cost if the model spirals.
Tiered termination — always set both timeout: and budget:. Three independent kill mechanisms protect against both runaway work and lost output:
- Tier 1: Budget (budget: 2.0) — kills if the model burns >$2 of tokens (spiral detection)
- Tier 2: Inner timeout (computed automatically: timeout_ms/1000 - 60s) — Claude gets SIGTERM, writes partial output
- Tier 3: Outer timeout (timeout: 600000) — hard kill if inner timeout failed
The save window (60s between tier 2 and 3) means partial output survives. Command prompts instruct agents to write findings incrementally to /workspace/reports/<node-id>/findings.md.

Full discussion of each option (per-node loop:, unrolled iterations, outer-loop wrapper) and the current monitoring surface is in Long-Running Workflows, Cycles, and Monitoring in CLAUDE-CONTEXT-workflows.md. Read that section whenever the user's pipeline is non-trivial.