A fourth runtime, and a second opinion.
Muse Code joins Claude Code, Codex and opencode with both memory legs proven on a real session: context injected before the turn, the finished session captured after it. It is also the only runtime so far that earned an orchestration role on measured results rather than on the assumption that another model family is useful.
What works today
- Context injection before each turn, proven on a live session rather than inferred from Muse's docs.
- Session capture into the same local store every other runtime writes to, so a decision made in Muse is recalled in Claude Code or Codex.
- Available as an orchestration review body in its own model family, distinct from Claude and Codex.
Earning the review slot
Hydrate's argument for multi-family orchestration is that the model that reviews work should not be the model that produced it. That only helps if the reviewing family actually catches things. Muse was measured against a seeded-defect benchmark before being wired in, not after.
| Reviewer | Tier-1 defects | Tier-2 defects |
|---|---|---|
| Muse | 16 / 16 | 10 / 12 |
| Codex | 15 / 16 | 8 / 12 |
| Claude (control) | 13 / 16 | 11 / 12 |
Muse also caught a non-idempotent-retry defect that Codex missed on three separate runs. It is roughly twice as slow as Codex per review, which is exactly why it is available but never a default on the reviewer path — that path makes two calls per round and the latency compounds.
What's rough
- Muse's plugin system is gated behind an experimental flag; without it, hooks are unavailable in the build.
- Installing the plugin does not arm the hooks — they must also be approved. An installed-but-unapproved plugin looks perfectly healthy and silently injects nothing, so
hydrate doctorchecks Muse's own plugin list rather than trusting Hydrate's manifest. - Capture reads Muse's event log directly, because the hook payload never carries a transcript path.
One memory, four runtimes, one bill each.
Hydrate is the layer between them, not another subscription.