adminandClaude Opus 5.5 fd75c183ba Fix #4, #6: resumes keep their task; later instructions reach the verifier
Resuming an agent (SendMessage to a finished agent) raises agent.spawn
again, with resume.agentId set. The router treated every resume as a
new task: new Tn, fresh routing, the brief replaced by the resume
message, the failure count and round base reset. Hence T322's
"routed to Sonnet, ran on Opus" for an Opus agent that was only
resumed (#4), and verifiers that judged later rounds against the wrong
text (#6).

- A resume passes through untouched; its message is kept as a later
  instruction to that agent. Task, routing, failures and rounds stay.
- A SendMessage to a tracked agent (by id or by the name its Agent call
  gave it) that was delivered is kept as a later instruction too.
- The verifier prompt lists later instructions after the brief and
  judges scope added there as in scope.
- [router-rewrite Tn]: the stored rewrite is used up only once its
  agent has started (a failed start kept nothing before: T332 was then
  routed afresh, to Haiku). With nothing stored the call is refused
  instead of starting an agent with no brief. The task keeps Tn.
- The routing label compares with the model the engine resolved at
  start, not the one the router asked for.
- 11 new tests; each new rule was also broken on purpose and caught.

Version 0.4.8.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MdqPWvuDpbpRvDEdURoAwo
2026-11-13 23:41:48 +00:00
2026-11-06 23:46:30 +00:00
2026-11-06 23:46:30 +00:00

model-router

A Claude Code plugin that picks a model for every subagent, checks dev work on a second model before it reaches a PR, and estimates what that saved compared with running everything on Opus.

What it does

Step Behaviour
Route Dev agents run on Sonnet when the brief decides the approach (Expected / acceptance criteria), on Opus when the design is still open. Review and test agents keep the model they are given (Opus when none).
Later instructions Every message sent to a dev agent after its brief (a SendMessage, or the message that resumes it) reaches the verifier as part of the task, so scope added later is judged as in scope. A resumed agent keeps its task number, routing and failure count.
Verify When a dev agent below Opus finishes a round, the router diffs its clone against the merge base, runs the test command in that clone, and has Opus judge the change.
Hand-back line Every dev hand-back gets one line at once: [model-router] Tn verifying on Opus... (a verdict follows), or [model-router] Tn no verification: dev ran on Opus. Go straight to review. Only the real hand-back counts: an interim stop while the dev waits on its own background work (e.g. a long test suite) posts nothing and starts no verifier. Decided by the model the run actually used, not only the routing choice; a difference is shown as (routed to X, ran on Y).
On FAIL 1st: resume the same agent with the issues. 2nd: a new agent one model up, in the same clone. After that: handed back to the manager.
Not verified A test timeout, a test command that cannot run, a missing clone or merge base, or a verdict that cannot be read even after asking once more: reported as NOT VERIFIED, never as a FAIL.
Savings /router savings prices every run on the model it ran on and compares it with the same tokens on Opus 5.5.

Briefs

Every dev, review and test brief carries one line:

Clone: /home/homelab/projects/goodwe/clones/<TICKET>-dev/

Add (Role: dev|review|test) when the path does not end in the role. Tickets with extra parts (OPT-11-P4) are fine. Name the base branch (Base branch: release/1.0).

Install

/plugin marketplace add <this repo's git URL>
/plugin install model-router@model-router

Or for development: claude --plugin-dir /path/to/model-router.

Settings (/config)

Setting Default Notes
testCommand empty Run in the agent's clone; output goes to the verifier
testTimeoutSec 1800 Wrapped in GNU timeout --verbose; a timeout is NOT VERIFIED
defaultBaseBranch release/1.0 After the PR lookup and the brief's own base branch
prBaseCommand empty Optional command that prints the PR's base branch
maxRewrites 2 Failed verdicts before handing back
safetyPatterns control loop, rs485, ... Marks tasks as critical in the log
respectExplicitModel false Keep a named model on agents without a role
allowFable false Rewrite on Fable after Opus fails
prices empty JSON price overrides, USD per MTok

Commands

  • /router: recent decisions, verdicts and savings
  • /router savings [Tn]: the full savings table, or one task
  • /router clear: reset log, records and costs

Tests

claude plugin validate .
claude plugin test .

80 tests. Every routing, verification, cost and timeout rule was also checked by breaking it on purpose (66 mutations in v0.4.4, plus 13 for the verdict reader in v0.4.5 12 for the hand-back line in v0.4.6 10 for interim stops in v0.4.7 and 14 for resumes, later instructions and rewrites in v0.4.8; all caught by a test).

Known limits

  • Savings are an upper bound: Opus may need fewer tokens than the cheaper model did.
  • timeout stops the docker client, not the container: name the container and remove it in a trap in the test command.
  • An Opus rewrite is not verified; the PR review covers it.
S
Description
No description provided
Readme
298 KiB
Languages
TypeScript 100%