| name | dart-retro |
| description | DART Retro: improve the AI harness from successful or unsatisfactory task outcomes |
dart-retro
Use this skill in Codex to run the DART dart-retro workflow. The editable
workflow source lives in .claude/commands/; this file is its generated adapter
in the shared .agents/skills/ catalog.
Invocation
- Claude Code:
/dart-retro <arguments>
- Codex:
$dart-retro <arguments>
Treat the text after the skill name as $ARGUMENTS. When the workflow
references $1, $2, etc., map those to the positional values supplied by the
user.
Command Body
Improve future task execution through a retrospective: $ARGUMENTS
Use the original request and session evidence to improve the harness for future
similar tasks: better quality, completeness, or efficiency. Successful,
unsatisfactory, incomplete, and blocked outcomes qualify. PR merge and dev-task
retirement do not alone establish that the user's goal was met.
Required Reading
@AGENTS.md
@docs/AGENTS.md
@docs/ai/principles.md
@docs/ai/components.md
@docs/ai/verification.md
Load task owners as needed. Read docs/onboarding/ai-tools.md for tool/runtime
questions and before controlled agent runs. docs/ai/components.md owns
placement of accepted learnings; this workflow owns the retrospective method.
Skip If
Conclude without harness edits when the evidence supports no reusable
improvement: existing guidance already covers the lesson and is discoverable,
the finding is task-specific, or no change has a supported benefit. State why
and report missing evidence; success, failure, or a merged PR alone is not a
skip condition.
Workflow
- Reconstruct intent and outcome. Default to the current session. Compare
the initial request and later scope decisions with artifacts and results.
Inspect relevant history around decisions, corrections, and failures,
including substantive domain work before CI/review/closeout. State missing
evidence; do not load every log by default.
- Find the harness contribution. Connect useful decisions, rework, missed
requirements, and wasted context/tool cycles to instructions, routing, tools,
or gates. Separate observations from inferred causes, implementation bugs,
and external blockers. Check existing owners and executable coverage before
proposing rules; investigate why existing guidance was missed.
- Choose a testable improvement. State the observed decision, causal gap,
owner, and what an agent starting from the same brief should do differently.
Name the expected benefit, a check that could disprove it, and successful
constraints to preserve.
- Improve existing owners. Prefer removal, consolidation, or rewriting to
appending rules or files; repair discovery when guidance already exists.
Follow
docs/ai/components.md for placement and adapter regeneration. Keep
session identifiers out of durable guidance. Preserve model/effort/action
limits; a retrospective does not reopen the original implementation or
authorize new GitHub mutations.
- Validate proportionately. For consequential instruction changes, replay
the observed decision and a contrasting similar task; use fresh controlled
agents when permitted and useful. Distinguish structural checks, predicted
benefits, and measured outcomes. Run the relevant gates in
docs/ai/verification.md and pixi run lint before committing.
Output
- Original intent, actual outcome, and material session evidence
- Accepted/rejected harness changes, expected future behavior, and owner files
- Validation results, measured benefits or limits, and remaining work
