git:20260827.09528f5 to git:20260906.3d137c0

- description: Select the minimum planning mode for a codebase change. Emit OpenSpec only for graph mode. Use for $spec, architecture requests, or proposal.md/design.md/tasks.md. Never implement code.

- Act as the architect. Select the minimum process and produce only its required artifact. Never implement code.

-

- ## Workflow

-

- 1. Read repository instructions, relevant history, touched code, ADRs, and prior decisions.

- 2. Run adaptive intake with the request, one direct acceptance check, and only observed task signals. Intake inspects the repository before returning questions.

- 3. Ask only returned material questions. Never invoke `$grill-me` unless the user explicitly asks for that stress test.

- 4. Follow the returned `ProcessDecision v1` mode:

-

- - `direct`: return the objective, scope, check, and stop condition. Write no durable planning file.

- - `verified_single`: return the decision capsule and one-writer hypothesis/check loop. Write no OpenSpec graph.

- - `light_spec`: write `decisions/<slug>.md`. Include revision, invariants, decisions, check, and escalation or reduction triggers. OpenSpec is optional.

- - `graph`: write `openspec/changes/<slug>/proposal.md`, `design.md`, `tasks.md`, and persist the returned transition as `process-decision.json`.

- 5. Give every graph task these fields:

- ```md

- - [ ] API-01 Implement the endpoint

- Depends: [DOMAIN-01]

- Paths: [src/api/, tests/test_api.py]

- Mode: write

- Isolation: auto

- Context: Keep the endpoint within the validated domain boundary.

- Acceptance: The endpoint returns the validated result.

- Check: python3 -m unittest tests.test_api

- ```

- `Depends` may be empty. `Paths` uses normalized repository-relative files or directory prefixes, never globs. `Mode` is `read` or `write`. `Isolation` is `auto` or `worktree`. Use one direct executable for `Check:` and move shell composition into a reviewed script. Do not add test or Markdown paths merely to satisfy a template; name them only when acceptance needs a focused regression, a contract check, or a requested documentation change. Workers follow the `minimal-by-default-v1` artifact budget.

- 6. Load `frontend-visual-validation` for rendered graph work. Add one reasoned `Visual-Scope:` and one `Visual:` entry per supported platform and state.

- 7. In graph mode, run `python3 skills/agent-graph/scripts/agent_graph.py validate --change <slug> --json`. Use `py` on Windows. Validation must pass before handoff and never starts workers. Bootstrap separately verifies that the saved decision revision, packets, checks, permission, budget, integration owner, and cleanup plan still match the graph.

- 8. Run `$spec-council --phase verdict <slug>` only when `--council` is present. Record accepted and rejected findings.

- 9. Before handoff, inspect standing-rule changes directly. Run `rule-curator` only when the approved scope is a whole-corpus rule audit; do not invoke it for an individual rule edit.

- 10. Tell the user to use the `$impl` skill with the decision or approved graph slug.

- Checks and external evidence decide acceptance. Council advice does not.

- Mode selection uses repository facts and task signals. It never depends on a

- provider, model, Canvas, or available worker count. Graph execution hints may

- describe role, risk, tools, context budgets, and a bounded `ContextRef`. Spec

- validates them without dispatching workers. Research collection uses `Mode:

- read`; the main researcher owns source adjudication. A checkbox or process exit

- is never an evidence grade.