epic-decompose

name	epic-decompose
description	Decompose batches of RHAISTRAT strategies into implementation epic DAGs. Accepts explicit IDs or a JQL query. Non-interactive.
user-invocable	true
allowed-tools	Glob, Bash, Agent

You are a non-interactive epic decomposition pipeline. Do not ask questions or wait for confirmation. Make all decisions autonomously.

Turn Discipline (critical)

You run headless. A response that contains only text and no tool call ends the session and aborts the pipeline mid-run. Until the dispatch loop returns done, every one of your turns MUST end with a tool call — never with prose.

In particular, after you launch a wave's Agent(s), do not write narration such as "the agent is running, waiting for it to complete." That text-only turn terminates the pipeline before the wave is collected. Your very next action, in the same turn as (or immediately after) the launch, MUST be the wait-for-wave Bash call. Do not describe what you are about to do — do it. Save any commentary for after the pipeline reaches DONE.

Setup

Parse $ARGUMENTS for:

• --jql "<query>", --limit N, --batch-size N (default 25), --data-dir "<path>"
• --headless, --announce-complete, --reprocess, --skip-if-has-epics
• Remaining arguments: explicit RHAISTRAT IDs

1. Init

python3 scripts/pipeline_state.py init [--batch-size N] [--headless] [--announce-complete]

2. IDs

JQL mode (--jql):

python3 scripts/fetch_strategy.py fetch "<query>" --ids-file tmp/pipeline-all-ids.txt [--limit N] [--data-dir "<path>"] [--skip-if-has-epics]

Print [DECOMPOSE] JQL: <jql> from stderr output.

Explicit mode:

python3 scripts/state.py write-ids tmp/pipeline-all-ids.txt <IDs>

If no IDs and no JQL, stop with usage instructions.

3. Bootstrap

bash scripts/fetch-architecture-context.sh
python3 scripts/fetch_components.py

Retry each once on failure. If retry fails, stop: "bootstrap failed."

4. Resume check + batch

Read all IDs: python3 scripts/state.py read-ids tmp/pipeline-all-ids.txt

Split into batches of batch_size. Write each:

python3 scripts/state.py write-ids tmp/pipeline-batch-1-ids.txt <batch_1_IDs>
python3 scripts/state.py write-ids tmp/pipeline-batch-2-ids.txt <batch_2_IDs>

Start the pipeline:

python3 scripts/pipeline_state.py set total_batches=<M>
python3 scripts/pipeline_state.py set-phase BATCH_START

Dispatch Loop

Repeat until action is done:

Step 1: Get next action

python3 scripts/pipeline_state.py next-action

Parse the YAML output for: action, phase, message, agents.

Step 2: Execute

done: Exit loop. Run teardown.

run_script: Run python3 scripts/pipeline_state.py run-phase. Go to step 1.

launch_wave: For each agent in the agents list:

• Build prompt: "<vars>\n\nRead <prompt_file> and follow all instructions exactly."
• vars are pre-rendered KEY=VALUE lines with {ID} already substituted.
• Launch as background Agent (with subagent_type if present).

Immediately wait for completion. Do not end your turn or emit any text after launching — your next tool call MUST be wait-for-wave:

python3 scripts/pipeline_state.py wait-for-wave

On exit 0 (complete): go to step 1. On exit 3 (still pending): re-run python3 scripts/pipeline_state.py wait-for-wave as your next tool call — do not write any text between attempts, as a text-only turn would terminate the pipeline. Any other exit code is an error.

Example launch_wave output

action: launch_wave
phase: DECOMPOSE
message: "DECOMPOSE: wave 1/2 (5 IDs)"
agents:
  - prompt_file: skills/epic-decompose/prompts/decompose-agent.md
    vars: |
      ID=RHAISTRAT-1234

Teardown

After phase reaches DONE:

python3 scripts/batch_summary.py --counts-only $(python3 scripts/state.py read-ids tmp/pipeline-all-ids.txt)

$ARGUMENTS