PR Custody

fix: add learning-pipeline guardrails to builder, fixer, and reviewer prompts

@zen-agentchecks n/achecks…zen/learning-pipeline-prompt-improvements-73wip7 → next3 files · +19 −1updated 2mo ago

▸Description

Why

Analyzed 72 fix PRs from the mercury learning pipeline (learning-pipeline + tool-fixer labels). Found recurring patterns of mistakes that the builder creates and the fixer makes worse. These targeted prompt additions prevent the top issues at source.

What

Builder prompt (action_builder/prompt.py):

raise_for_status() placement: must come AFTER custom error handlers, not before (was creating dead code in ~6 PRs)
URL-encode path params with quote() (4+ PRs had broken URLs from unencoded emails/IDs)
No duplicate params — use AliasChoices instead of exposing both old and new field names (5+ PRs)
Response verbosity — never inline file content, return download URLs instead

Fixer prompt (fix_action_prompt.py):

Don't fix infra issues at action level (wrong connected_account_id, "tool not found", auth problems)
Single action only — no scope creep across files
Never inline file content in responses
No silent default changes without documentation
No over-engineering (500+ line rewrites for simple bugs)
Verify root cause before cosmetic fixes (e.g., check URL/version before adding retries)
Use AliasChoices not duplicate fields

Reviewer prompt (reviewer/prompt.py):

Added checklist items matching all fixer anti-patterns above so reviewers catch them

How to Test

make chk passes (ruff format, ruff lint, pyrefly type check all green)
Changes are prompt-only (no code logic changes), impact is on agent behavior in future pipeline runs

Notes

Analysis source: 72 open PRs in ComposioHQ/mercury with learning-pipeline + tool-fixer labels. Verdict breakdown: 38 good (52%), 26 acceptable (36%), 6 questionable (8%), 3 bad (4%). The prompt changes target the top recurring patterns across all categories.

loading diff…