hindsight

hindsight skill

You orchestrate regression replay across all of the user's projects. You don't write code — you invoke the hindsight.py script and present its results in a useful way.

When this skill applies

• The user just finished a significant change (refactor, dependency bump, AI-led feature) and asks "did anything regress?"
• The user explicitly says /hindsight ...
• The user mentions "regression sweep", "replay tests", or "check the test plans"

When this skill does NOT apply

• One-off questions about a specific test ("why is this test failing?") — that's a debugging task, not a regression sweep
• Creating new regression entries — those come from the iterative-tdd plugin's /tdd workflow, not hindsight

Workflow

1. Default action: run python3 "${CLAUDE_PLUGIN_ROOT}/scripts/hindsight.py" list to enumerate everything that's currently watched.
2. If the user asked for a sweep, run replay-all. If they named a specific slug, run replay <slug>.
3. Present results in priority order: failures first (with their stderr tail), passes summarized.
4. For any failure, surface:
   • The project + slug
   • The exit code
   • The path to the run dir for full output
   • A suggestion: re-run /hindsight replay <slug> after fixing.

Project discovery

Hindsight uses ~/.config/hindsight/projects.yaml if it exists, otherwise auto-scans ~/Developer/*. If the user adds a new project that's outside ~/Developer/, prompt them to register it: /hindsight projects add <path>.

Cost discipline

Replay is local-only — no LLM in the inner loop. The skill itself is mostly a thin wrapper. Don't read run output files unless something failed and the user wants details.

What hindsight does NOT do (yet)

• Watch mode (continuous replay on file save)
• Git post-commit hook integration
• Smart triage of failures (LLM-assisted "this failed because X changed in commit Y")
• Cross-project dependency awareness

These are future work — flag them if the user asks for them rather than improvising.