pre-plan

Pre-Planning Context Gathering

Gather everything needed to discuss a solution with the user, so a plan can be written afterward and then implemented. You are NOT writing a plan in this command, and you are NOT implementing anything. The output is a context briefing for discussion.

Scope

$ARGUMENTS

Phase 1 — Decide the source

Inspect the Scope above:

• Looks like a ticket ref — an ID, a tracker URL, a #number; teamwork:workflow-identify holds the project's real pattern: treat as ticket → run Phase 2.
• Looks like a freeform problem description: skip Phase 2, use the description as the brief.
• Empty: use the current conversation context. If it's unclear what we're gathering context for, ask the user before proceeding.

Phase 2 — Read the ticket (only if a ticket ref was given)

Delegate to ONE subagent. Goal: load the full ticket and everything it transitively references — the ticket's own content and the external systems it links to. This subagent never opens the codebase. Reading code is Phase 3 and Phase 4 work. A ticket subagent that greps a repository or reads source files has left its scope, and what it brings back from there is noise the later phases redo properly.

The subagent must:

• Read title, description, and all comments.
• Inspect references, attachments, labels, project, status, links — anything rich.
• Follow links to related tickets, MRs/PRs, design docs, dashboards. Read an MR/PR through the forge CLI (glab, gh) — description, discussions, and the diff stat (which files changed, how much) — never the diff itself, and never a local checkout of the repository. Whether a diff is relevant is decided later, once the Phase 3 code map exists.
• If the ticket links to chat threads (Slack or similar) and they aren't already mirrored into ticket comments, read those threads too.
• Dump all findings into ./.claude/plans/pre-plan-<ticket-slug>-ticket.md (create the directory if missing) as a single markdown file. Include: full ticket content (title, description, all comments verbatim or near-verbatim where they carry real information), summarized linked context per link, a list of every concrete code/file/table/system reference that the ticket or its linked items mention — transcribed as written, not searched for, not verified against code — and a list of any unresolved questions or ambiguities spotted in the ticket.
• Return ONLY the path to that file (plus a 2-3 sentence high-level gist). Do NOT inline the full dump into the subagent's reply — the orchestrator will read the file as needed.