Decompose Task

Decompose Task

Steps

1. Read the task description.
2. List the concrete observable effects it requires (file written? API called? row inserted?).
3. Group effects into ordered steps. Prefer fewer, bigger steps over many small ones — atomicity, not granularity.
4. For each step, name the tool and sketch args_template. Mark steps as idempotent or non-idempotent; the last step should be idempotent.
5. Return the plan. If you can't decompose unambiguously, ask the user.

Output shape

{
  "plan": [
    {
      "step": 1, "name": "fetch-input",
      "tool": "fetch", "args_template": {"url": "<from user>"},
      "expected_effect": "200 + JSON body cached", "idempotent": true
    },
    {
      "step": 2, "name": "transform",
      "tool": "shell", "args_template": {"cmd": "jq ... > out.json"},
      "expected_effect": "out.json written", "idempotent": true
    }
  ]
}

Failure modes

• 17 steps because each line is its own step → users can't reason about it.
• A non-idempotent final step (validator catches this).
• Hidden side effects ("step 4 also clears the cache") that aren't in expected_effect.