Pastiche

Pastiche

Orchestrates pastiche-implementer-round1, pastiche-reviewer, and pastiche-implementer-round2 in a bounded loop, then emits a single task report.

Preflight

Run ls pastiche/FACT.md pastiche/KNOWLEDGE.md pastiche/WISDOM.md. If any are missing, stop and report:

> This project does not appear to have pastiche set up — expected pastiche/{FACT,KNOWLEDGE,WISDOM}.md. Run /pastiche-init to scaffold them.

Workflow

1. Dispatch pastiche-implementer-round1 with {task} → capture {r1_report}.
2. Dispatch pastiche-reviewer with {task, r1_report} → capture {doubts}. The reply is either one or more lines of # | , or the literal (no doubts). If any line fails to match the record shape and the output isn't (no doubts), stop and report:

> Reviewer returned malformed output; expected # | records or the literal (no doubts).

1. If {doubts} is (no doubts), go to 6 (no {r2_report}).
2. Dispatch pastiche-implementer-round2 with {task, r1_report, doubts} → capture {r2_report}.
3. Failsafe. For each record in {r2_report}'s unresolved: block (skip if unresolved: -): Read the file, then Edit to insert a comment in the file's native comment syntax above the targeted ` (matching surrounding indentation). Comment text: pastiche-unresolved-doubt: `.
4. Build check. Read build_command from pastiche/config.yaml. If null or absent, skip. Otherwise run it. For each error, patch using the build output as source of truth. Bounded to 3 patch attempts per error. If an error persists after 3 attempts, leave it and note it in Follow-ups.
5. Runtime review. Scan the changed files for common sources of runtime and console errors. Patch any issues found. Bounded to one pass.
6. Emit the final response.

Final response

## Summary
- **Files changed:**  (omit if none)
- **Implementation:** 

## Follow-ups
-  — 
- ... (omit section if no follow-ups)

Files changed is the union of files: entries from {r1_report} and {r2_report}; omit the line if both are -.

Follow-up sources (each follow-up bullet must be self-describing; readers shouldn't need to know about rounds, dispositions, or the loop):

• Each scenario in {r1_report}'s gaps: line → KNOWLEDGE.md — .
• Each (knowledge-gap)-tagged record in {r2_report}'s defended: block → KNOWLEDGE.md — .
• Each (wisdom-gap)-tagged record in {r2_report}'s defended: block → WISDOM.md — .
• Each record in {r2_report}'s unresolved: block → # — .
• Each failsafe comment written in step 5 → : — .

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: retz8
• Source: retz8/pastiche
• License: MIT

Install and usage instructions live in the source repository linked above.