Internal AI SDLC skill, not client-facing by default. Every rule below is important to follow. None of it can be skipped. Retrospectives propose improvements; they never silently change policy.

• Skill name:

  ai-sdlc-retrospective

• Primary audience: Delivery, Dev, QA, BA
• Supporting audience: PM, platform maintainers
• Audience tags: Delivery, Dev, QA, BA, PM
• SDLC stage: Post-delivery learning
• Purpose: Separate evidence-backed observations from governed improvements.
• Output:

  retrospective.md

  and

  _ai_sdlc/retrospective.toon

• Completed or paused feature root.
• Retrospective JSON containing observations and proposals.
• Exact artifact or validation evidence for every observation.

• Ask only when evidence, proposal owner, or intended policy target is unclear.
• Keep observed facts separate from interpretation and proposed change.
• Do not mark a proposal accepted without a durable decision reference.
• Do not infer approval from implementation, silence, or prior chat.

• Support

  --quick-flow

  and

  --full-flow

  ; full flow takes precedence.
• Quick flow permits draft proposals but still requires observation evidence.
• Full flow additionally requires every proposal to have an owner and every accepted proposal to have a decision reference.
• Neither mode applies proposals or edits target files.

• Return observation and proposal counts, accepted-decision coverage, blockers, and output paths directly in the Codex response.
• Before the final response, emit

  ai-sdlc-handoff/v1

  with

  result

  ,

  blockers

  ,

  next_required

  , and

  next_optional

  ; every action includes

  reason

  ,

  command

  , and

  expected_artifact

  .
• Do not create

  summary.txt

  ,

  *-summary.txt

  , or policy patches.
• Preserve rejected and deferred proposals for learning history.

• Write human output to

  <feature-root>/retrospective.md

  .
• Write machine output to

  <feature-root>/_ai_sdlc/retrospective.toon

  .
• Keep proposal target paths as references only.
• Apply accepted improvements later through the target-owning workflow.

• Retrospective is a utility workflow and does not advance feature lifecycle.
• Read

  _ai_sdlc/state.toon

  to confirm feature identity and delivery status.

• --state-check

  is read-only;

  --begin-state

  and

  --complete-state

  are rejected by the finalizer.
• A proposal decision changes proposal governance status, not lifecycle state.

• Markdown starts with

  artifact_metadata

  using schema

  ai-sdlc-retrospective-metadata/v1

  .
• Include

  metatags

  for

  ai-sdlc

  ,

  retrospective

  ,

  learning

  , and proposal statuses.
• Record feature, workspace, flow mode, evidence paths, and decision refs.

• Read

  specs/_ai_sdlc/specs-index.toon

  or

  specs-refiniment/_ai_sdlc/specs-index.toon

  before targeted evidence reads.
• Do not refresh

  specs/specs-index.md

  or

  specs-refiniment/specs-index.md

  for a read-only draft.
• Refresh indexes only through the owning workflow after durable report writes.

• Read

  references/retrospective-contract.md

  before preparing input.
• Use

  scripts/retrospective.py

  to validate separation, decision safety, and emit or atomically write the canonical report pair.

• Anchor observations to repository-relative files and positive line numbers.
• Reference proposals back to one or more observation IDs.
• Give every proposal a target, owner, review status, and next action.
• Add

  decision_ref

  only when a proposal has actually been accepted.

1. Review delivery evidence, outcomes, friction, escapes, and effective controls.
2. Record observations without embedding recommended changes.
3. Draft proposals that cite observation IDs and identify target plus owner.
4. Validate evidence lines and governance statuses with the finalizer.
5. Review proposals with accountable owners and record durable decisions.
6. Regenerate the report when proposal status changes.
7. Route accepted work to SDD or the owning policy/configuration workflow.

ai-sdlc-retrospective/v1

based_on

target

change

owner

status

decision_ref

next_action

• Pass when observations and proposals are separate, evidence is exact, every proposal traces to an observation, and accepted proposals cite a decision.
• Fail when a proposal is presented as an observation, approval is implied, or target policy content would be mutated by finalization.

• A retrospective with observations and no proposals is valid.
• Rejected proposals retain their evidence links and rationale in

  next_action

  .
• An accepted proposal without

  decision_ref

  always fails, including quick flow.
• A target may not exist yet; the report records it without creating it.

• Do not edit policy, configuration, skill, or source targets.
• Do not accept proposals on behalf of owners.
• Do not turn anecdotes into observations without durable evidence.
• Do not advance feature lifecycle or hide failed proposals.
• Use

  $ai-sdlc-navigator

  to route accepted implementation work.