Repository navigation
add support for more harness/model pairs #428
Description
muse spark 1.3 in muse code
grok 4.7 in grok build
mimo-v2.6-pro in mimo code
gemini 3.8 flash in agy
What a first look found (5 October 2026)
Each CLI was run headless on one machine, with one-line prompts in a throwaway repository with a linked worktree. The details, the working commands and how each fact was confirmed are in docs/research/harness-candidates.md, added in the pull request that mentions this issue. How strong each model is: docs/research/model-strength-for-review.md §0.
All four can:
- run one prompt with no approval prompts and exit;
- take a Model and an Effort;
- stream JSON with a session id;
- resume a session by id;
- load skills from
.agents/skills/when the prompt names one as/name(where thirdshift already links them for Codex); - spawn sub-agents.
| Harness · Model | Gaps for thirdshift |
|---|---|
| agy · Gemini 3.8 Flash | None found. Start-up is slow and varies, 8–95 s. It never reads CLAUDE.md or .claude/skills/. |
Grok Build · grok-4.7 |
Stopping it leaves its running shell command alive. Each repository needs a one-time grok --trust. |
Muse Code · muse-spark-1.3 |
Its sandbox can't start on stock Ubuntu, so it needs --yolo. It needs a trust flag on every run. Its JSONL format is its own, carries no usage, and runs replies together in the final text. There's no free model list. Auto-update must be turned off. |
MiMo Code · mimo-v2.6-pro |
Not on cron's PATH. Exits 0 when it fails, and has no final event. Effort did nothing on the provider tested. Stopping it leaves its running shell command alive. |
Order by readiness: agy, Grok Build, Muse Code, MiMo Code.
Once added, each Harness also becomes a possible reviewer for #429.
To decide in grilling:
- one spec with a Ticket per Harness, or a spec each;
- stopping a session: kill its whole process tree, not just the CLI's process group (Grok, MiMo);
- telling failure from the event stream where the exit code doesn't (MiMo);
- checking the Model and Effort before any work, for each Harness. agy, Grok and MiMo print their catalogs for free; Muse has no such command, and MiMo ignores an unknown effort;
- per-repository setup thirdshift does or documents (Grok's trust), and the environment it sets for each session (auto-update off);
- instruction files: a repository with only a
CLAUDE.mdgives agy no project instructions; - whether Setup offers each one, and how the User config and the command line name them, beyond today's "claude or codex" (
src/harness.rs).
Activity
JacobStephens2 commentedon Oct 17, 2026
Delivered through Spec #432 and its completed Tickets, merged in PR #442. thirdshift supports agy, Grok Build, Muse Code and OpenCode alongside Claude Code and Codex. The agreed Spec explicitly replaced MiMo Code with running MiMo-V2.6-Pro through OpenCode. Closing as delivered under that agreed scope.