the-own-lab/session-reflection

session-reflection

Purpose

/dev's reflection artifact (ADR-001 D7). Nine questions: eight standard retro, one prediction-accuracy audit per ADR-002 D3. Output is appended to the brief and produces any reconstruction drills needed.

Inputs

• BRIEF.md ## Human-Owned Core with Final implementation already filled by verify.
• Diff for this session.
• TODO.md and CHANGELOG.md (already updated by update-docs).

Outputs

• Appended block in BRIEF.md containing the 9 answers.
• For each diverged core item, Reconstruction required: yes flipped on and the Reconstruction Drill section filled (by user, with Main Agent help).
• No new file; session reflection stays inside the brief before archive.

Procedure

Ask and record answers for all nine questions:

1. What did we change? (one-paragraph factual summary)
2. How was it implemented? (key design decisions, surprises)
3. Are we satisfied with the implementation? (yes/no + reason)
4. Are the tests meaningful? (map Test Plan → tests written)
5. Did we respect the spec? (any silent drift? any ADR-worthy amendment?)
6. TODO / CHANGELOG updated? (confirm update-docs ran)
7. Best practices followed? (TDD cycle, small diffs, commit hygiene)
8. What problems surfaced? (blockers, rework, flaky tests)
9. Prediction accuracy (ADR-002 Q9): For each Human-Owned Core item, compare User prediction with Final implementation. Main Agent fills initial judgment (aligned | diverged | partially-aligned) with a one-line reason; user confirms or corrects. Write result to that item's Prediction result field.

After Q9:

• For each item with Prediction result = diverged OR partially-aligned without a clear reason OR User selected option ≠ Final implementation OR new core behavior that was not predicted: set Reconstruction required: yes and invoke the four-part drill (invariant restated / pseudocode / edge case / regression test).
• Route pseudocode through spec-conformance reconstruction judge for mechanical comparison — don't let the LLM soften diverged.

Failure modes

• Any of Q1–Q9 left blank → FAIL. Nine means nine.
• Drill triggered but skipped → blocks /dev completion; cannot archive brief.
• LLM softens diverged to "mostly aligned" → override and keep diverged. ADR-002 explicitly forbids softening.

Does not do

• Does not run on /review or /debug or /think (ADR-001 D7).
• Does not judge the quality of the user's drill answers as a score — only the invariant match.
• Does not cross-session track prediction-accuracy statistics (ADR-002 Non-Goal).