Issue Management

Issue Management

OpenCode v1: Skill names below are exact IDs from the active catalog, not slash commands. Load them with the native skill tool. Slash commands are direct user entry points only.

Apply GitHub issue mechanics to work that is already understood. The caller owns grilling, planning, specifications, and ticket decomposition; this skill owns the resulting tracker mutations.

1. Discover tracker context

Read consumer-owned instructions and tracker documentation before using tracker tools. Do not assume a path or that an adapter exists. Establish the target repository and account, authorization boundary, issue templates, language, available labels and types, project mapping, and supported native relationships from repository evidence or read-only tracker queries.

If any fact required for the requested mutation remains ambiguous, ask the user. Never guess a repository, account, label, relationship, project or status, and never create a parallel tracker.

2. Inspect before drafting

Resolve a referenced number or URL in the confirmed repository and read the complete issue and comments. A bare number is ambiguous outside an established repository context. Search for an existing issue before proposing a new one. For an epic, inspect its native children and dependency graph rather than inferring state from a list of links in prose.

3. Draft the smallest useful change

Use the repository's issue template and documented language when they exist. Otherwise ask when language affects the result, then keep the issue self-contained and concise. Give it a functional title that says what changes, not which file or mechanism will be touched. Make the opening useful to a non-technical teammate:

## In short

<one or two sentences: who or what benefits, what changes, and why it matters>

## Acceptance criteria

- [ ] <verifiable result>

## Implementation context and proof

<relevant current state, constraints, risks, test seams, and evidence>

Retain the context, non-goals, dependency rationale, rollback conditions, risk, and evidence needed to make the issue independently actionable for a developer or agent. Do not copy an umbrella specification, decision history, file inventory, or implementation walkthrough into every child issue. Use a user story only when a real actor and value are clearer in that form; it is not a required wrapper for technical work. Keep parent and dependency graph state in native relationships rather than duplicating it in body sections.

Select only issue types, labels, assignees, parents, dependencies, and project fields that consumer evidence or live metadata establishes. Present those proposed metadata values with the draft; metadata is part of the task contract even when it is not body prose.

Before proposing a pickable status, apply the consumer's documented readiness gate. A thin intake issue may remain in backlog; do not compensate for missing evidence or scope by inventing technical detail.

4. Confirm scope and write

The user must have authorized the target and bounded set of external writes. Reuse authorization already supplied by the request or calling brief; moving from a specification or breakdown into tracker mechanics requires no new approval. If authority or a material choice is missing, present the exact issues, metadata and mutations, then ask only for what remains unresolved.

In grillmester local run, the original local run prompt may carry that human authorization when it names the exact, bounded mutation and repository. That mode has no second client tool dialog. Existing authorization does not allow unstated writes or resolve an ambiguous repository, account, issue body or metadata choice; return Status: NEEDS_INPUT when any such choice remains.

For Doctor Who, and for any runtime without explicit command-execution capability, use only an available, approved semantic GitHub integration. Never fall back to gh, shell, raw HTTP, or ask the user to run a command on the agent's behalf. If read evidence is unavailable, request the smallest pasted or exported issue/project excerpt and return Status: NEEDS_INPUT. If a write tool is unavailable, keep the reviewed mutation as a draft and return Status: NEEDS_INPUT, naming the required integration or the exact manual GitHub UI action.

An implementation agent in a terminal session explicitly started through Grillmester and cplt may use cplt-guarded gh issue commands as its approved GitHub adapter. Use only the confirmed current repository and account, keep the command inside cplt, and never bypass cplt's repository scope, GitHub guard or command restrictions. Prefer semantic GitHub tools when they are available. When GH_TOKEN is supplied, never inspect or print the token or enumerate the full environment; invoke only the bounded gh command needed for the approved mutation. Outside that explicit cplt context, do not substitute gh or shell for a missing semantic tracker capability. Never use raw HTTP for a tracker mutation.

Verify every created or changed issue, project item, field value, and relationship after writing. Report partial failure without silently omitting project state or substituting a text-only relationship.

5. Maintain lifecycle without taking over delivery

Use native sub-issues and dependencies as the graph. Recommend the first open, unblocked child when the user asks for the frontier; selection does not start implementation or mutate project state.

Link delivery through the repository's PR convention. Comment on or close an issue only when explicitly requested or included in the authorized delivery. When an epic has no open children, propose its summary and closure; never close it automatically.