Internal AI SDLC skill, not client-facing by default. Every rule below is important to follow. None of it can be skipped. The journal is authoritative; complete state is a recoverable TOON-first projection.

• Skill name:

  ai-sdlc-runtime

• Primary audience: Dev, Delivery
• Supporting audience: QA, Release, Architecture
• Audience tags: Dev, Delivery, QA, Release
• SDLC stage: Controlled execution
• Purpose: Persist task selection and outcomes so interrupted delivery can resume without duplicate work or unsupported completion claims.
• Output:

  _ai_sdlc/runs/<run-id>/journal.jsonl

  , exact

  state.json

  , and complete token-efficient

  state.toon

• Versioned run plan, safe run ID, dependency graph, task input fingerprints, retry limits, budgets, and commit-boundary flags.
• Exact result fingerprint, token use, commit evidence, and reason when recording.

• Ask when task outcome, result identity, commit evidence, or stop reason is missing.
• Reject unknown tasks, invalid transitions, dependency cycles, and journal gaps.
• Never mark an unrecorded task complete from conversational confidence.

• Support

  --quick-flow

  and

  --full-flow

  ; full flow takes precedence.
• Both modes use the same journal, idempotency, dependency, budget, and commit rules.
• Full flow reviews every event, retry, stop reason, and commit boundary.

• Report run status, sequence, current task, ready tasks, budgets, stop reason, recovery status, and next command.
• Return validation and handoff summaries directly in the Codex response.
• Emit

  ai-sdlc-handoff/v1

  with

  result

  ,

  blockers

  ,

  next_required

  , and

  next_optional

  ; actions include

  reason

  ,

  command

  , and

  expected_artifact

  .
• Do not create

  summary.txt

  ,

  *-summary.txt

  , or another standalone summary file.

• Route each run only to

  _ai_sdlc/runs/<run-id>/

  .
• Keep journal append-only, JSON recovery state atomically replaceable, and TOON state as the complete agent-facing projection.
• Never store runtime records inside a feature spec folder.

• Read the owning feature

  _ai_sdlc/state.toon

  before starting feature work.
• Runtime task state does not replace feature

  state.toon

  ; the owning workflow performs feature transitions only after runtime and validation evidence agree.

• Runtime-related Markdown uses canonical

  artifact_metadata

  and

  metatags

  .
• Machine records use

  ai-sdlc-run-plan/v1

  ,

  ai-sdlc-run-event/v1

  , and

  ai-sdlc-run-state/v1

  with deterministic fingerprints.

• Read

  _ai_sdlc/specs-index.toon

  first and use

  specs-index.md

  for human feature discovery before constructing a plan.
• Runtime state does not refresh either specs index.

• Read

  references/runtime-contract.md

  before starting, resuming, retrying, or recording a run.
• Validate plans with

  references/run-plan.schema.json

  , events with

  references/run-event.schema.json

  , and state with

  references/run-state.schema.json

  .
• Use

  scripts/runtime.py

  for every state mutation and recovery operation.

• Immutable plan identity and ordered task IDs with dependencies.
• Per-task maximum attempts and commit-boundary requirement.
• Run-level maximum task starts, failures, and recorded tokens.
• Exact outcome evidence for each started task.

1. Validate plan identity, unique tasks, dependency existence, and acyclicity.
2. Start one atomically created run workspace.
3. Call

   --next

   ; repeated calls return the same running task without another event.
4. Execute the task through the owning host and workflow.
5. Record succeeded, failed, or blocked with exact evidence and resource use.
6. Let retryable failures return to pending; inspect terminal retry or budget stops.
7. Use

   --retry

   only for an eligible blocked or failed task after its cause is resolved.
8. Use

   --resume

   to replay and hash-check the journal and repair stale state.

• Pass when journal chain and replay are valid, dependencies are satisfied, budgets remain, outcome evidence is complete, and commit-boundary tasks carry commit identity.
• Fail when a transition duplicates work, skips dependencies, exceeds a hard boundary, omits evidence, or conflicts with journal history.

• Corrupt or missing summarized state is repaired from a valid journal.
• A journal sequence gap, hash mismatch, or invalid transition fails recovery.
• Retry exhaustion and each budget type have distinct stop reasons.
• A commit-boundary task cannot succeed with a missing or malformed commit SHA.

• Do not execute task commands, create commits, satisfy gates, or approve work.
• Do not edit journal lines, plan identity, attempts, budgets, or fingerprints manually.
• Do not run two mutating runtime commands concurrently for one run.
• Declarative workflow steps and host execution belong to later workflow and adapter layers.