fix(dreamer): forward task fallbackModels & timeout, increase subagent maxSteps and verify batch size

IZ
2d ago
+98-87

Approved issue: #784

Harness coverage (required)

For each entry select: covered / not applicable — surface does not exist there / NOT covered.

  • OpenCode 1: covered — shared Dreamer plugin logic has unit coverage
  • OpenCode 2: covered — shared Dreamer plugin logic has unit coverage
  • Pi: covered — Pi Dreamer and subagent runner tests.
  • OMP: covered — OMP runner unit cases
  • Rust module: not applicable — no Rust Dreamer task runtime/config surface.

Summary

  • Forward task-level fallbackModels and timeoutMinutes to the Pi subagent runner, honoring configured fallbacks and timeouts.
  • Increase dreamer-docs and dreamer-memory-mapper maxSteps from 60 to 400.
  • Increase memory verification VERIFY_BATCH_SIZE from 5 to 10.

Testing

  • Verified configured fallbackModels and timeoutMinutes are passed through to the runner.
  • Pi subagent runner tests: 165 passed; Plugin agent/verify tests: 50 passed; Pi Dreamer/telemetry tests: 64 passed.
  • Both package typechecks, targeted Biome checks, and git diff --check passed.

fix #784

RetriggerView in GreptileConfidence Score: 4/5

Update the stale fallback assertion before merging so the Pi Dreamer tests pass.

Findings

  1. P1 Fallback test now fails ▶

Summary

The PR forwards task timeouts to Pi, raises the documentation and memory-mapper step limits to 400, and increases verification batches to ten memories.

  • Pi Dreamer runs use each task’s configured time limit.
  • Two read-only Dreamer agents can take up to 400 steps.
  • Memory verification checks up to ten memories per batch.
  • The unavailable-provider test fixture uses a clearer name.

Diagram

 
%%{init: {'theme': 'neutral'}}%%
flowchart LR
    A[Task model and fallbacks] --> B[Shared retry loop]
    B --> C[Pi runner: one model per attempt]
    C --> D[Return result and token spend]
    D --> B

Reviews (2) · Last reviewed commit: "Update packages/pi-plugin/src/dreamer/in..." · Reviewed by Greptile