AI-DLC Portfolio

Maintain a thin control plane around existing AI-DLC workflows. Never replace or modify the child workflow engine.

Resolve every relative resource path against the directory containing this SKILL.md. In the commands below, set SKILL_DIR to that absolute directory:

SKILL_DIR=/absolute/path/to/skills/aidlc-portfolio

Operating Rules

1. Run "$SKILL_DIR/scripts/portfolio.ts" for every portfolio mutation.
2. Keep portfolio state at <root>/portfolio and project checkouts at <root>/repositories.
3. Give each active child intent its own path under <root>/worktrees.
4. Dispatch one aidlc_runner per child intent with CAO assign.
5. Treat child AI-DLC state as opaque. Never edit aidlc-state.md or derive its next stage.
6. Require evidence and confidence for discovered catalog relationships.
7. Ask the human about missing or uncertain business facts before verification.
8. Run dispatch validation before starting or resuming a child runner.
9. Treat the portfolio supervisor as the sole shared-memory approval authority.
10. Never merge worktree-local project.md or team.md changes.
11. The supervisor owns bootstrap; the operator should not have to prepare the workspace, repository checkout, intent files, worktrees, or child harness.
12. Never dispatch until portfolio discovery has explicit human confirmation.
13. Never answer a child AI-DLC question without human input.
14. Stage, project, and verify child harnesses only through harness commands.
15. Advance parent work only through the persisted portfolio lifecycle.
16. Require one structured result per child and pass convergence before Learn.

Bootstrap From Work Items

The normal entry point is an operator naming a repository and one or more work items. Perform setup on their behalf:

1. Derive a stable portfolio ID from the repository and sorted work-item IDs.
2. Use the terminal's launch directory as the portfolio root unless the operator explicitly supplies another root. Never create a nested root.
3. Run init against that exact root; never hand-create portfolio state.
4. Create a fresh canonical checkout under <root>/repositories. Do not reuse, clean, reset, or remove an existing user checkout or worktree.
5. Inspect the work items and register the project, dependencies, and one portfolio intent per independent work item using templates and utility commands.
6. Create each worktree through worktree create.
7. Locate the Claude AI-DLC distribution at $HOME/Project/aidlc-workflows/dist/claude. Run harness stage, harness sync, and harness verify to project one manifest-pinned Opus 5 runtime into every worktree.
8. Run doctor and dispatch check, then assign all dependency-ready aidlc_runner sessions concurrently with CAO working_directory set to each absolute worktree.
9. Report the generated root, intent IDs, worktree paths, and terminal IDs.

The operator owns creation of the empty parent/root directory and launches the supervisor from it. Ask them only when a required repository, work item, credential, or AI-DLC distribution cannot be discovered. Do not turn routine bootstrap inside the root into operator instructions.

Initialize

bun "$SKILL_DIR/scripts/portfolio.ts" init \
  --root /absolute/workspace \
  --id portfolio-id \
  --name "Portfolio Name"

Then run:

bun "$SKILL_DIR/scripts/portfolio.ts" doctor --root /absolute/workspace

Initialization is explicit and idempotent. Never use a hook to create a workspace.

For a workspace created before lifecycle state schemaVersion 2, migrate once:

bun "$SKILL_DIR/scripts/portfolio.ts" lifecycle migrate --root <root>

Read portfolio-lifecycle.md. After every startup or callback, inspect the durable phase and exact next action:

bun "$SKILL_DIR/scripts/portfolio.ts" lifecycle status --root <root>

Discover And Register

Read catalog-model.md before discovering or changing projects, components, capabilities, or dependencies. Read coordination-rules.md before creating child intents, worktrees, or dispatches.

Inspect technical evidence before asking questions. Use the templates under assets/templates, then register validated files:

bun "$SKILL_DIR/scripts/portfolio.ts" project register --root <root> --file <project.yaml>
bun "$SKILL_DIR/scripts/portfolio.ts" dependency add --root <root> --file <dependency.yaml>
bun "$SKILL_DIR/scripts/portfolio.ts" intent create --root <root> --file <intent.yaml>

Record unresolved questions under <root>/portfolio/questions. Support Unknown, Defer, and referral to another stakeholder. Do not upgrade a fact from discovered or proposed to verified without human approval.

After catalog registration, render assets/templates/discovery-decision.yaml. Record the human's decision for organization, business outcomes, business capabilities, and the complete dependency list:

bun "$SKILL_DIR/scripts/portfolio.ts" discovery confirm \
  --root <root> --file <discovery-decision.yaml>

For every unknown or deferred disposition, the corresponding fact name must appear in acceptance.unknowns or acceptance.deferrals. The utility stores a catalog revision; any later catalog change makes confirmation stale and blocks dispatch until the human reviews it again.

Project registration records a canonical checkout; it does not clone one. Before worktree creation, ensure the registered path exists under <root>/repositories and is a Git repository.

Harness Lifecycle

Read harness-lifecycle.md before staging or updating child runtimes. Stage the source distribution after registering intents and creating their worktrees:

bun "$SKILL_DIR/scripts/portfolio.ts" harness stage \
  --root <root> \
  --source "$HOME/Project/aidlc-workflows/dist/claude"

bun "$SKILL_DIR/scripts/portfolio.ts" harness sync --root <root>
bun "$SKILL_DIR/scripts/portfolio.ts" harness verify --root <root>

The staged manifest records source revision, source and staged hashes, overlay revision, and staging time. The default overlay pins the Claude opus[1m] alias to global.anthropic.claude-opus-5[1m]. Use --project and --intent to sync or verify a subset.

Never bypass a tracked-path refusal. Reconcile project-owned .claude/ or aidlc/ content explicitly before retrying.

Dispatch

Create or validate a worktree:

bun "$SKILL_DIR/scripts/portfolio.ts" worktree create \
  --root <root> --project <project> --intent <intent> \
  --branch <branch> --base <base>

bun "$SKILL_DIR/scripts/portfolio.ts" harness verify \
  --root <root> --project <project> --intent <intent>

bun "$SKILL_DIR/scripts/portfolio.ts" dispatch check \
  --root <root> --project <project> --intent <intent>

The worktree must pass harness verification and dispatch validation. Render assets/templates/runner-message.md, then use CAO assign with agent_profile: aidlc_runner and working_directory set to the validated absolute worktree. The supervisor profile enables this CAO parameter through CAO_ENABLE_WORKING_DIRECTORY=true. After assignment returns a terminal ID, record it:

bun "$SKILL_DIR/scripts/portfolio.ts" session update \
  --root <root> --project <project> --intent <intent> \
  --status active --terminal <terminal-id>

If assignment fails, leave the session pending. Never mark a session active before CAO returns a terminal ID.

Parent Lifecycle

Advance only one phase at a time:

bun "$SKILL_DIR/scripts/portfolio.ts" lifecycle advance \
  --root <root> --to discover

Repeat for confirm and plan. Entering Dispatch requires explicit human plan acceptance:

bun "$SKILL_DIR/scripts/portfolio.ts" lifecycle advance \
  --root <root> --to dispatch --accepted-by "<human decision maker>"

After all child sessions complete, advance to integrate, resolve every reported blocker, then advance to learn. Finish only after learnings and integration state are clean:

bun "$SKILL_DIR/scripts/portfolio.ts" lifecycle complete \
  --root <root> --actor "<portfolio supervisor>"

Cross-Project Convergence

Read convergence.md. Before a child reports completion, it renders assets/templates/child-result.yaml and submits it:

bun "$SKILL_DIR/scripts/portfolio.ts" result submit \
  --root <root> --file <child-result.yaml>

After all expected results arrive, compute graph impact and integration status:

bun "$SKILL_DIR/scripts/portfolio.ts" convergence check \
  --root <root> [--intent <intent>]

Missing results and dependency-order risks require remediation. For an eligible breaking, deferred, or unknown risk, obtain an explicit human decision and bind it to the current risk revision:

bun "$SKILL_DIR/scripts/portfolio.ts" convergence decide \
  --root <root> --id <risk-id> --decision accepted \
  --accepted-by "<human decision maker>" --note "<decision rationale>"

Use resolved only when the human confirms the reported risk is resolved. Re-run convergence check; changed evidence invalidates stale decisions.

Human Question Relay

When a child AI-DLC stage asks for human judgment, the runner writes the generated questions to Markdown using [Answer]: markers and submits the unanswered file:

bun "$SKILL_DIR/scripts/portfolio.ts" question submit \
  --root <root> --id <packet-id> --project <project> --intent <intent> \
  --stage <stage> --file <questions.md>

Submission rejects any pre-filled answer and moves the session to waiting. The runner reports the packet and stops. It must never invoke question answer.

The supervisor offers the user Guide, Edit Markdown, or Chat. Write each human answer verbatim on its [Answer]: line without changing generated question text, then resolve the packet:

bun "$SKILL_DIR/scripts/portfolio.ts" question answer \
  --root <root> --id <packet-id> --file <answered.md> \
  --mode guided --answered-by "<human decision maker>"

The command rejects changed question text, missing answers, invalid modes, and repeat answers. It preserves the answered Markdown byte-for-byte, records the human provenance, and returns the session to active. Use question list to verify status before asking the runner to resume.

Shared Memory

Read shared-memory.md before processing child AI-DLC learnings. dispatch check returns canonical baseline revisions for project.md and team.md; include them in the runner assignment.

Child runners submit structured proposals from assets/templates/learning-proposal.yaml:

bun "$SKILL_DIR/scripts/portfolio.ts" learning propose \
  --root <root> --file <proposal.yaml>

The supervisor lists and reviews proposals. Approval performs the canonical write under the portfolio lock:

bun "$SKILL_DIR/scripts/portfolio.ts" learning list \
  --root <root> --project <project> --status pending
bun "$SKILL_DIR/scripts/portfolio.ts" learning approve \
  --root <root> --id <proposal>

If approval reports a stale base, inspect the latest canonical rule set and resolve duplicates or contradictions. Record that human judgment before retrying:

bun "$SKILL_DIR/scripts/portfolio.ts" learning reconcile \
  --root <root> --id <proposal> --note "<review outcome>"

Use memory refresh at safe stage boundaries when the child needs current canonical context. Before completing a child session, inspect and clean both memory destinations. Both mutations require the exact worktree revision returned by inspect:

bun "$SKILL_DIR/scripts/portfolio.ts" memory inspect \
  --root <root> --project <project> --intent <intent> --destination project
bun "$SKILL_DIR/scripts/portfolio.ts" memory clean \
  --root <root> --project <project> --intent <intent> --destination project \
  --expected-worktree-revision <sha256>

Repeat for team. session update --status completed refuses shared-memory changes remaining in the feature worktree.

Recover And Report

Run status after startup or callback delivery:

bun "$SKILL_DIR/scripts/portfolio.ts" status --root <root>

Reconcile terminal loss against durable child AI-DLC state. Relaunch a runner against the same registered worktree and child intent; never create replacement workflow state.