fix(dispatcher): keep unclassified work on configured worker role #503

Merged
karsten merged 1 commit from fix/classifier-worker-fallback into main 2026-11-15 19:38:10 +00:00
Owner

Concrete Minerva latency evidence

All four occupied worker slots were observed on homelab-main/deepseek-v4-flash/medium, with classifierReceipt.phase=abandoned, despite configured dedicated Qwen worker roles. Thus worker, Main and validator traffic compete on DeepSeek. An isolated fast dispatcher probe does not reveal this productive routing fallback.

Fix for NEW, not-yet-intended work

  • Unknown/interrupted/invalid classification is NOT a request to run productive work on Main. Resolve the configured high-capability WORKER role (do not invent a high complexity verdict). Configurations without a usable high role retain their legacy default.
  • The direct classifier's HTTP fallback is kept on the chosen dispatcher pair; a 429 cannot turn dispatcher/off into Main/medium. No confidence threshold is weakened and no unknown classifier call is replayed.
  • Existing productive intents, exact pins and rules remain untouched and retain their stored route. This PR deliberately does NOT migrate the four currently running DeepSeek-bound workers or edit any live ledger.
  • Actual worker selection remains durable/visible through existing intent and execution observations.

Three new tests cover unknown, invalid and already-abandoned classifiers, one classification at most, no invented verdict, correct durable worker pair and no settled-effect replay after reload. Existing role/pin/stop/cancel/recovery tests remain.

Typecheck/build, 33 focused tests and independent review PASS. Exact rebased full suite:2642 passed, one existing optional skip, zero failures. Exact-head CI pending at last check.

End-to-end Minerva latency is not claimed fixed by this source PR: old productive route bindings and backend prefill/cache contention still require safe reconciliation/drain, not a silent runtime model switch.

## Concrete Minerva latency evidence All four occupied worker slots were observed on homelab-main/deepseek-v4-flash/medium, with classifierReceipt.phase=abandoned, despite configured dedicated Qwen worker roles. Thus worker, Main and validator traffic compete on DeepSeek. An isolated fast dispatcher probe does not reveal this productive routing fallback. ## Fix for NEW, not-yet-intended work - Unknown/interrupted/invalid classification is NOT a request to run productive work on Main. Resolve the configured high-capability WORKER role (do not invent a high complexity verdict). Configurations without a usable high role retain their legacy default. - The direct classifier's HTTP fallback is kept on the chosen dispatcher pair; a 429 cannot turn dispatcher/off into Main/medium. No confidence threshold is weakened and no unknown classifier call is replayed. - Existing productive intents, exact pins and rules remain untouched and retain their stored route. This PR deliberately does NOT migrate the four currently running DeepSeek-bound workers or edit any live ledger. - Actual worker selection remains durable/visible through existing intent and execution observations. Three new tests cover unknown, invalid and already-abandoned classifiers, one classification at most, no invented verdict, correct durable worker pair and no settled-effect replay after reload. Existing role/pin/stop/cancel/recovery tests remain. Typecheck/build, 33 focused tests and independent review PASS. Exact rebased full suite:2642 passed, one existing optional skip, zero failures. Exact-head CI pending at last check. End-to-end Minerva latency is not claimed fixed by this source PR: old productive route bindings and backend prefill/cache contention still require safe reconciliation/drain, not a silent runtime model switch.
fix(dispatcher): keep unclassified work on configured worker role
All checks were successful
decision-interface-ci / full-validation (pull_request) Successful in 4m3s
078f4a6e39
Sign in to join this conversation.
No reviewers
No labels
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
harness/core!503
No description provided.