Doctor

Grillmester Doctor for OpenCode v1

Audit the consumer-owned boundary around the portable Grillmester target. Distinguish observed runtime behavior from configuration that merely appears correct, and recommend the smallest useful improvement.

Keep the audit read-only

Never create, edit, delete, rename, install, authenticate, enable, disable, or update anything while this skill is active. Do not print provider credentials, tokens, request headers, or complete configuration values that may contain secrets. If the user also asks for a fix, complete the audit and return its evidence and bounded file set to the calling workflow. That workflow can then implement within the existing authorization without a new approval round.

Invoking this skill proves only that the skill is visible in this session. It does not prove that every Grillmester agent or command is discovered, that a different OpenCode installation uses the same config directory, or that the selected model can complete the workflow reliably.

Discover the effective OpenCode boundary

Establish the Git root, working directory, OpenCode version, selected agent, and the effective config directory when they are observable without mutation. Inspect only applicable sources:

• AGENTS.md files from the repository root through the working directory;
• opencode.json or opencode.jsonc at project and user scope, without exposing provider options or environment-derived secrets;
• OPENCODE_CONFIG and OPENCODE_CONFIG_DIR as path evidence only;
• project-local .opencode/agents, .opencode/commands, and .opencode/skills, plus portable .agents/skills roots;
• the configured Grillmester target's manifest.json, agents/, commands/, and skills/ trees;
• repository build definitions, entry points, tests, and maintained domain, architecture, security, and operations documentation.

OpenCode merges several configuration scopes. A file's existence does not prove it wins. When effective config or the runtime roster cannot be observed, mark precedence and collision coverage UNVERIFIED rather than guessing.

Diagnose the actual skill resolution

Capture the exact failed invocation, requested ID, current runtime skill roster, and offered name. Load the offered identity through the native skill tool; do not invent a prefix, use a heading as an ID, or retry an obsolete grillmester- alias. Command wrappers are user-facing entry points and must reference that same skill ID.

Record the loaded path, target version/revision, nav-pilot agent package when visible, and whether the selected manifest is full or focused. Compare actual loaded content or its hash with that manifest when observable. An omitted focused skill requires the matching full target or an included workflow. A listed skill whose load fails is a resolution failure, not proof of absence; retain the exact call and error. Neither filesystem presence nor an agent's self-description proves the runtime winner.

Inspect same-ID user and repository components, older prefixed copies, legacy Hovmester workflows, and instruction references. Mark unobservable source or precedence links UNVERIFIED. The source checkout's read-only scripts/audit_consumer_setup.py can inventory these without a legacy manifest or caller; user roots must be passed explicitly. Review its exact-copy file hashes before any authorized removal. Preserve customized or unknown components and consumer domain rules; names or manifest membership alone are not deletion evidence.

Check the native target contract

Verify these dimensions separately:

1. Target integrity — generated file counts and hashes agree with manifest.json; no runtime dependency artifact is treated as Grillmester source.
2. Agent discovery — four primary agents and three hidden subagents are present under agents/; agent frontmatter has mode, hidden, and native permissions, and omits model so the session/provider choice is inherited.
3. Delegation — Grillmester can use task only for kokk, grill-inspektor, and researcher; other roles preserve their narrower task boundary.
4. Skills and commands — all manifest skills exist under skills/, each command wrapper resolves the same skill ID, and manual-only source skills require approval through ordered permission.skill rules.
5. Instructions — consumer facts remain in applicable AGENTS.md or other consumer-owned sources; Grillmester does not distribute a competing repository instruction file.
6. Provider and model — the selected provider/model is observable and supports the tool calls and context needed by the chosen workflow. Record identity and capability, never credentials. A configured provider is not a successful model smoke test.
7. Permissions — effective project or user configuration does not silently widen a role's denied task, shell, edit, web, or skill boundaries.
8. Shadowing — no higher-precedence agent, command, or skill with the same ID silently replaces a Grillmester component.

Treat the OpenCode target as generated output. Recommend changing its canonical plugin source, adapter policy, or overlay and regenerating it; never recommend hand-editing a generated file.

Keep consumer instructions minimal

Keep one concise source for shared repository facts and constraints. AGENTS.md may own that content or be a tiny pointer to another consumer-owned instruction file, provided the intended clients demonstrably read the shared content. Verify effective discovery; do not relocate useful rules merely to prefer a filename or duplicate them across clients.

Retain only essential context that models cannot reliably infer from maintained code and tooling. Prefer existing documentation pointers over copied guidance. Missing extra instruction files are not a defect. Do not recreate setup workflows, sync checks, generic task procedures or instruction trees to fill this audit's categories.

Return one compact report

GRILLMESTER_OPENCODE_DOCTOR: READY | READY_WITH_GAPS | NOT_READY | UNVERIFIED

Runtime:
- OpenCode version: <observed version or UNVERIFIED>
- Effective config directory: <path evidence or UNVERIFIED>
- Provider/model: <non-secret identity or UNVERIFIED>
- Requested/offered skill ID: <exact IDs or UNVERIFIED>
- Loaded source and version: <path/target/revision or UNVERIFIED>
- Projection: full | focused | UNVERIFIED
- Repository/user conflicts: <exact paths and evidence or UNVERIFIED>

Coverage:
- Target integrity: OK | BLOCKER | GAP | UNVERIFIED — <evidence>
- Agents/delegation: OK | BLOCKER | GAP | UNVERIFIED — <evidence>
- Skills/commands: OK | BLOCKER | GAP | UNVERIFIED — <evidence>
- Instructions: OK | GAP | DUPLICATION | UNVERIFIED — <evidence>
- Permissions: OK | BLOCKER | GAP | UNVERIFIED — <evidence>
- Shadowing: OK | BLOCKER | UNVERIFIED — <evidence>

Smallest next action:
- <one bounded consumer- or source-owned change, or none>

Use NOT_READY only for an evidenced blocker. Keep facts, inference, and unknowns separate.