Retrospective / Learning workflow (mandatory)

Goal

Ensure durable, reusable insights from this chat session are captured in

.agent/memory.md

so future sessions can resume without re-discovering them.

Inputs

• The current chat session transcript (user + assistant messages in this session)

• .agent/constitution.md

  ,

  .agent/memory.md

  ,

  .agent/focus.md

  ,

  .agent/issues/

Extraction rules (strict)

Only capture durable items:

• confirmed workflow rules
• stable project conventions (not one-off)
• confirmed commands (but commands/boundaries belong in constitution, not memory)
• user preferences that affect future work on this repo
• corrections to previous misunderstandings
• pitfalls/gotchas discovered during implementation

Do NOT capture:

• transient task state (belongs in focus)
• speculative ideas not adopted
• secrets/tokens
• personal/sensitive data

Placement rules (strict)

• Project-specific commands/boundaries/constraints →

  .agent/constitution.md

• Durable workflow learnings and recurring conventions →

  .agent/memory.md

• Current session status →

  .agent/focus.md

• Follow-ups and approvals needed →

  .agent/issues/

Procedure

1. Read: constitution/memory/focus/tasks.
2. Scan the chat session for:
   • explicit corrections ("No, do X instead of Y")
   • newly confirmed commands or tools
   • newly confirmed constraints ("never refactor", "only write docs in …")
   • repeated misunderstandings (add a "pitfall to avoid")
   • preferences expressed by the user
3. Update

   .agent/memory.md

   :
   • append a dated subsection:

     ## Retrospective YYYY-MM-DD

   • add short, atomic bullets phrased as "Do/Don't/Prefer"
   • avoid duplication with constitution (link to constitution section if needed)
4. Update focus if the retrospective reveals unresolved items.
5. Invoke

   agent-ops-tasks

   discovery if actionable items found.
6. Do not declare completion unless retrospective has been run.

Issue Discovery During Retrospective

After scanning session, invoke

agent-ops-tasks

discovery for actionable items:

1. Collect actionable learnings:

   • "We should add tests for X" →

     TEST

     issue
   • "This pattern is confusing, needs docs" →

     DOCS

     issue
   • "Found a bug but didn't fix it" →

     BUG

     issue
   • "This could be optimized later" →

     PERF

     issue
   • "Technical debt noticed" →

     CHORE

     or

     REFAC

     issue
   • "Security concern noted" →

     SEC

     issue

2. Present to user:

   📋 Retrospective identified {N} actionable items:
   
   Medium:
   - [DOCS] Document the retry mechanism in PaymentService
   - [TEST] Add integration tests for new OAuth flow
   
   Low:
   - [CHORE] Clean up commented-out code in UserController
   - [PERF] Consider caching user preferences (noted during implementation)
   
   Create issues for these? [A]ll / [S]elect / [N]one
   

3. After creating issues:

   • Mark in memory.md: "Created {ISSUE-IDs} for follow-up"
   • These become part of the project backlog