Internal AI SDLC skill, not client-facing by default. Every rule below is important to follow. None of it can be skipped. The graph is a generated projection; repository artifacts and Git remain authoritative.

• Skill name:

  ai-sdlc-delivery-graph

• Primary audience: Delivery, Dev, QA, Architecture
• Supporting audience: PM, BA, Security, Release
• Audience tags: Delivery, Dev, QA, Architecture, BA
• SDLC stage: Traceability and readiness
• Purpose: Build a deterministic repository-wide lifecycle graph and answer trace, gap, coverage, and orphan questions from stable evidence anchors.
• Output: complete

  _ai_sdlc/delivery-graph.toon

  for agents, plus

  _ai_sdlc/delivery-graph.json

  for schema/interoperability and

  _ai_sdlc/delivery-graph.md

  for human review when

  --write

  is requested

• Repository root containing lifecycle Markdown and, optionally, Git history.
• A stable trace ID or graph node ID for path queries.
• Explicit structured trace lines when a relationship cannot be derived safely.

• Report ambiguous short IDs with all matching scoped node IDs.
• Never invent an edge because two terms look semantically similar.
• Treat missing links as gaps or orphans instead of guessing intent.

• Support

  --quick-flow

  and

  --full-flow

  ; full flow takes precedence.
• Both modes rebuild from current authoritative inputs before answering.
• Full flow requires reviewers to resolve every reported ambiguity and high-value requirement coverage gap before claiming readiness.

• Report the graph fingerprint, node and edge counts, source fingerprint, coverage, gaps, and orphans.
• Return query paths as ordered node and edge evidence, not prose-only claims.
• Return the validation and handoff summary 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.
• Do not create ad hoc summaries outside the canonical graph outputs.

• Write generated graph data only below repository

  _ai_sdlc/

  .
• Keep versioned schemas and interpretation rules in this skill package.
• Preserve repository-relative evidence paths and one-based source line numbers.

• Graph indexing is read-only with respect to feature lifecycle state.
• Read canonical

  _ai_sdlc/state.toon

  before interpreting feature status; do not create a feature-local

  state.toon

  or change its sequencing.
• A graph gap may block a later gate but cannot advance or reopen state itself.

• Node and edge records use

  ai-sdlc-delivery-node/v1

  and

  ai-sdlc-delivery-edge/v1

  .
• The aggregate uses

  ai-sdlc-delivery-graph/v1

  and deterministic SHA-256 fingerprints derived from normalized content.
• Evidence producers use

  ai-sdlc-evidence-source/v1

  ; generated freshness records and ledgers use

  ai-sdlc-evidence-record/v1

  and

  ai-sdlc-evidence-ledger/v1

  .
• Source Markdown participates through canonical

  artifact_metadata

  and

  metatags

  ; graph output does not replace or rewrite that metadata.

• Read canonical

  _ai_sdlc/specs-index.toon

  first and use

  specs-index.md

  for human discovery across implementation and refinement workspaces.
• Scope short trace IDs by feature directory to prevent cross-feature identity collisions.
• Do not mutate either specs index during graph generation.

• Read

  references/delivery-graph-contract.md

  before interpreting nodes, edges, gaps, or query paths.
• Use

  references/delivery-node.schema.json

  ,

  references/delivery-edge.schema.json

  , and

  references/delivery-graph.schema.json

  for machine contracts.
• Use

  scripts/delivery_graph.py

  for deterministic indexing and queries.
• Read

  references/evidence-ledger-contract.md

  before registering or judging evidence freshness. Validate inputs and outputs with

  references/evidence-source.schema.json

  and

  references/evidence-ledger.schema.json

  .
• Use

  scripts/evidence_ledger.py

  to recalculate current file identities, propagate stale state, and query fresh coverage.

• Stable trace identifiers in lifecycle Markdown.

• Refs:

  lines or co-located declaration references.
• Optional explicit

  Component: <path> -> <trace-id>

  and

  Evidence: <path> -> <trace-id>

  links.
• Conventional commit bodies containing

  Spec:

  and

  Task:

  and annotated Git tags for release nodes.

1. Scan canonical lifecycle Markdown and hash every input file.
2. Scope declared and referenced trace IDs to their feature.
3. Add only evidence-backed semantic edges and Git task/release edges.
4. Normalize and fingerprint nodes, edges, sources, gaps, and coverage.
5. Inspect orphan and missing-coverage results.
6. Run trace queries using a scoped node ID when a short ID is ambiguous.
7. Write generated projections only when explicitly requested.
8. Register evidence with captured artifact and dependency hashes, expiry, and upstream evidence identities.
9. Rebuild the ledger for an explicit

   as_of

   date; resolve missing, changed, expired, unknown, ambiguous, or cyclic evidence before claiming coverage.

• Pass when all node and edge evidence resolves, identities are unambiguous, fingerprints recompute, and required coverage gaps are understood.
• Fail when an explicit query is absent or ambiguous, an edge endpoint is missing, or generated output drifts from current authoritative inputs.

• The same

  AC-001

  in two features is two nodes and requires a scoped query.
• A reference without a local declaration still becomes a placeholder node and is reported as an orphan or coverage gap.
• Git-less fixture directories still produce valid lifecycle graphs.
• Generated graph files never become their own inputs.
• Evidence dependency cycles, duplicate IDs, ambiguous subjects, and unknown upstream evidence fail closed.

• Do not infer semantic similarity, ownership, or policy decisions.
• Do not mark evidence fresh without matching current hashes, expiry, and all upstream evidence states.
• Do not edit lifecycle artifacts to make a graph appear complete.
• Do not treat graph generation as approval, validation evidence, or release.
• Do not overwrite producer manifests or evidence artifacts during indexing.