feat(routing): task-aware reweighting for interactive tasks #12

Merged

epistemophiliac merged 3 commits from develop into main 2026-08-26 04:20:20 +00:00

Conversation 0 Commits 3 Files changed 6 +252 -14

epistemophiliac commented 2026-08-26 04:19:42 +00:00

Owner

Copy link

Summary

Shifts intelligence weight toward speed/reliability for chat/qa/writing so a slow Frontier model (20-108s TTFB) no longer wins interactive routes. Code and reasoning keep intelligence leading. Operator-toggleable via routing_task_aware_enabled (default ON).

Motivation

Production diagnostics showed xkiro/stealth/ox-alpha-free winning 10/25 recent routes at 20-77s latency while fast mid-tier models sat unused. Root cause: the aexora-default preset (i=0.45, s=0.15, r=0.40) is intelligence-led and task-agnostic.

Changes

• task-routing.ts: per-task weight shifts
• router.ts: taskAdjustedWeights wired into orderChain
• fallback.ts: taskAwareEnabled field on GET/PUT /routing
• diagnostics.ts: strategy/weights/taskAwareEnabled exposed in overview
• task-routing.test.ts: unit tests

Verification

• 34/34 tests pass
• chat: speed 0.15 → ~0.26, intelligence 0.45 → ~0.34
• code/reasoning: no shift

## Summary Shifts intelligence weight toward speed/reliability for chat/qa/writing so a slow Frontier model (20-108s TTFB) no longer wins interactive routes. Code and reasoning keep intelligence leading. Operator-toggleable via `routing_task_aware_enabled` (default ON). ## Motivation Production diagnostics showed `xkiro/stealth/ox-alpha-free` winning 10/25 recent routes at 20-77s latency while fast mid-tier models sat unused. Root cause: the aexora-default preset (i=0.45, s=0.15, r=0.40) is intelligence-led and task-agnostic. ## Changes - task-routing.ts: per-task weight shifts - router.ts: taskAdjustedWeights wired into orderChain - fallback.ts: taskAwareEnabled field on GET/PUT /routing - diagnostics.ts: strategy/weights/taskAwareEnabled exposed in overview - task-routing.test.ts: unit tests ## Verification - 34/34 tests pass - chat: speed 0.15 → ~0.26, intelligence 0.45 → ~0.34 - code/reasoning: no shift

epistemophiliac added 2 commits 2026-08-26 04:19:42 +00:00

feat(routing): task-aware reweighting for interactive tasks ... 0d12628622

Shifts intelligence weight toward speed/reliability for chat/qa/writing so a
slow Frontier model (20-108s TTFB) no longer wins interactive routes. Code and
reasoning keep intelligence leading. Operator-toggleable via
routing_task_aware_enabled (default ON), exposed in diagnostics and /routing.

Production diagnostics showed xkiro/stealth/ox-alpha-free winning 10/25 recent
routes at 20-77s latency while fast mid-tier models sat unused. Root cause:
the aexora-default preset (i=0.45, s=0.15, r=0.40) is intelligence-led and
task-agnostic, so the bandit always prefers the highest-intelligence model
regardless of task type or observed TTFB.

Fix: after the base preset + peak-hours adjustment, apply a per-task convex
rebalance that moves a fraction of intelligence weight to speed (and slightly
to reliability). The re-normalized weight vector stays in [0,1]. For chat,
speed rises from 0.15 to ~0.26 and intelligence drops from 0.45 to ~0.34 —
enough for a fast 1-3s model to outscore a 30s frontier one.

merge task-aware-routing into develop a3dfb0492d

epistemophiliac added 1 commit 2026-08-26 04:20:00 +00:00

chore(release): bump version to v1.3 4865de84fa

epistemophiliac merged commit 05dbb614fd into main 2026-08-26 04:20:20 +00:00

epistemophiliac referenced this pull request from a commit 2026-08-26 04:20:21 +00:00

Merge pull request 'feat(routing): task-aware reweighting for interactive tasks' (#12) from develop into main

Sign in to join this conversation.

Reviewers

No reviewers

Labels

Clear labels

No items

No labels

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

1 participant

Notifications

Subscribe

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

epistemophiliac/LLMAPI!12

No description provided.