loop-plan-execute

Overview

loop-plan-execute implements the Plan-Execute pattern: a sequential pipeline where each step consumes the previous step’s output. It uses TaskCreate/TaskUpdate as a durable state machine, supports replanning on step failure, and returns all step artifacts to the runtime without synthesizing them.

When to use

Triage selects this archetype when steps have explicit dependencies — step N needs step N-1’s output before it can run. Typical examples: research → outline → draft → revise, or extract → transform → load → validate. For independent tasks that can run in parallel, use loop-orchestrator instead.

Inputs

A phase spec block from plan.yml:

phase: pipeline
archetype: plan-execute
steps:
  - id: s1
    name: research
    model: sonnet
    prompt: "..."
    output: /tmp/loop-runs/<ts>/intermediates/research.md
    depends_on: []
  - id: s2
    name: build-matrix
    model: sonnet
    prompt: "Read {{s1.output}}, extract structured comparison matrix."
    output: /tmp/loop-runs/<ts>/intermediates/matrix.csv
    depends_on: [s1]
replan_policy: on_failure
budget:
  max_total_tokens: 80000

Outputs

When a step fails, a small replanner worker decides: retry with an adjusted prompt, skip the step if downstream is still viable, or escalate to the user. Maximum 2 replans per phase.

Example

# Triage selects plan-execute for:
# "Research our top 3 auth libraries and write a recommendation doc"
#
# s1: research.md  →  s2: matrix.csv  →  s3: recommendation.md
# Each step reads the previous step's file before running.