/mimo-code

mimo-code

Overview

Delegate a write-capable coding session to mimo, staying the conductor. Heavy work and tokens move to mimo; you orchestrate and review the diff.

When to use

• The user invokes /mimo-code [provider/model] [variant] <task> or asks to run/resume mimo.
• A prior mimo session needs continuation (it asked something or was cut off).

The rule that shapes everything

Never run mimo in the main thread. ALWAYS dispatch the mimo-delegate subagent (model sonnet) so the live NDJSON stream and any blocking stay out of the main context. You only ever receive its distilled summary.

Parse the invocation

Raw slash-command arguments: $ARGUMENTS

Interpret as [provider/model] [variant] <task> (empty when triggered by description match rather than /mimo-code — then read intent from the user's message):

• A leading xxx/yyy token → model. A following minimal|low|medium|high|max → variant.
• The rest is the task; resume/continue → resume the recorded session (see Resume). Natural-language requests map the same way.

Resolve the model — ASK, do not default (only when no model was given, fresh runs only)

1. Gather options via the mimo-resolve subagent (model sonnet). It runs the resolve-models launcher and returns the authenticated provider∩model options and variants — keeping the catalogue output out of the conductor's context. Empty options (no authenticated provider) → STOP and tell the user to run mimo providers login.
2. Using the returned options (the authenticated intersection — count models, not providers): Exactly one usable model → auto-pick (asking is pointless). ≤4 → one AskUserQuestion listing them. >4 (the common real case — a single provider can expose dozens) → narrow, never dump: AskUserQuestion the provider first (the authenticated set is small), then AskUserQuestion ≤4 models for that provider — lead with mimo-resolve's recommended id, and rely on the auto-added Other free-text for the long tail. One authenticated provider offering several models is still "more than one" → ASK; "low-risk"/"mechanical"/cost or proactivity never justify a silent pick, and printing the whole catalogue for the user to retype an id verbatim is the anti-pattern to avoid.
3. Ask effort/variant with a ≤4 menu — AskUserQuestion caps at 4 options, so you cannot list all five variants + a default. Offer default (omit --variant) plus a few variants, each with a one-line description (default is the safe recommendation for a one-shot delegation); minimal/medium reach via Other.

The conductor does the asking/auto-pick itself — mimo-resolve only gathers, never asks.

Pick a handle — unique per delegation

Generate a UNIQUE slug per fresh delegation: a task-derived stem plus a short random suffix, e.g. json-logger-a1b2 ([a-z0-9_-]). The handle is the registry key (<handle>.sessionid/.lock/.ndjson); a unique handle is what keeps parallel or repeated runs from clashing. Record the handle you used — resume reuses THAT exact handle, never a freshly re-derived bare slug.

Dispatch

Dispatch mimo-delegate with: handle, cwd (absolute), model/variant (fresh only), the prompt, and mode: fresh. (The subagent resolves the launcher path itself.) Relay its summary; never echo raw NDJSON.

Resume

Re-dispatch mimo-delegate with the SAME handle, mode: resume, and the continuation prompt. Do NOT re-ask for a model (the session remembers it). The launcher resumes by recorded session id — never --continue.

Parallel runs

Concurrent tasks → distinct handles, one mimo-delegate each (dispatch in one message to run in parallel). Concurrent write tasks → give each its own git worktree under .worktrees/ and pass it as cwd; the registry is safe regardless, but two mimos editing one tree is a workspace hazard.

Common mistakes (counters to real failures)