fix(routing): gate agentic pool to harness-capable models so the price dial can demote Opus

[steventohme]

Problem

The QualityBias price/quality dial could not bias cheaper for the agentic/coding clusters — Opus 4.8 stayed in rotation at every dial position.

Root cause: the guard was on the wrong axis. applyDialAlpha floors the per-cluster quality weight (alpha[i] = max(dialAlpha, floor[i])) at 0.88 for agentic clusters. A single scalar weight can't express "cheaper but still harness-capable": the agentic quality band is compressed while cost spans the full deployed range, so alpha has only two regimes — high → Opus pinned, low → cheapest-incapable model wins (the original minimax-m3 "grepped for a skill instead of running it" cliff). 0.88 (from #501) avoided the cliff by disabling the dial for agentic entirely.

Fix — move the guard from the weight to the candidate pool

1. Catalog: new AgenticUse capability axis (AgenticLow) — stricter than ToolUseLow: a model emits valid tool calls but can't sustain the skill/tool orchestration loop. Flags minimax-m3, gemini-3.1-flash-lite, deepseek-v4-flash, qwen3-next-80b-instruct. AgenticLowSet() mirrors ToolUseLowSet(). (mimo-pro left eligible — the catalog's documented sweep says the pro variant is stable.)
2. Scorer: on has_tools turns, subtract AgenticLowSet from the eligible pool via the existing soft-filter seam (empty-pool fallback preserved, so a decision is always returned).
3. v0.70 metadata: relax agentic alpha_floor 0.88 → 0.30 (in place, matching fix(routing): per-cluster alpha_floor so the price/quality dial keeps every cluster at its best-for-budget model #501's pattern) so the dial walks the full harness-capable ladder.

Validated behavior (probed the real v0.70 bundle)

Agentic demotion ladder is now:

dial	model
quality end	claude-opus-4-8
mid	gemini-3.1-pro-preview
cost end	gpt-5.4-mini

The gate is load-bearing at the cost end: without it the pool falls through to gemini-3.1-flash-lite (incapable); with it the cheapest capable coder wins. Covered by TestScorer_HasToolsExcludesAgenticLowFromArgmax, the rewritten TestApplyDialAlpha_AgenticStaysOnCapableModelAtLowDial, and catalog set tests.

Known follow-up

The dial preview (RoutingDistribution) has no has_tools signal, so it scores over the full pool and will under-represent the gate at low dial (shows gemini-flash-lite for agentic clusters where real tool turns get gpt-5.4-mini). Actual routing is correct. Fully unifying them means making the gate cluster-level in blendScoresV2.

Notes

• v0.70 geometry (centroids / quality_means) is unchanged — only routing knobs + a pool gate. Deploys via the normal gitlink bump.
• Offline sign-off: cmd/routing-report --quality-bias.

🤖 Generated with Claude Code

[github-actions]

🧭 Cluster routing review

ℹ️ Informational — candidate artifact (latest unchanged); no sign-off required.

Cluster routing report — v0.70

Routed 180 labeled probes through the real Scorer at top_p=2, embedder jina-v2-base-code-int8. Premium = opus-4-7/4-8, gpt-5.5, gemini-3.1-pro, sonnet-4-6.

Register → model (this artifact)

register	Opus %	premium %	top models
conversational	33%	43%	deepseek-v4-pro:16 claude-opus-4-8:10 gemini-3.1-pro-preview:2 gemini-3.5-flash:1
trivial_nl	13%	50%	deepseek-v4-pro:13 gemini-3.1-pro-preview:10 claude-opus-4-8:4 gpt-5.5:1
knowledge_qa	70%	77%	claude-opus-4-8:21 glm-5.2:4 deepseek-v4-pro:2 gemini-3.1-pro-preview:1
easy_code	73%	87%	claude-opus-4-8:22 deepseek-v4-pro:3 gemini-3.1-pro-preview:2 gpt-5.5:2
hard_code	60%	80%	claude-opus-4-8:18 gpt-5.5:6 glm-5.2:6
agentic_tool	87%	93%	claude-opus-4-8:26 gpt-5.5:2 glm-5.2:2

What each cluster represents (probe register mix by nearest centroid)

cluster	probes	dominant	register breakdown
0	36	agentic_tool	agentic_tool:12 knowledge_qa:9 hard_code:8 conversational:4 easy_code:2 trivial_nl:1
1	8	easy_code	easy_code:3 agentic_tool:2 conversational:1 knowledge_qa:1 trivial_nl:1
2	2	conversational	conversational:2
3	1	knowledge_qa	knowledge_qa:1
4	3	conversational	conversational:1 easy_code:1 trivial_nl:1
5	14	agentic_tool	agentic_tool:5 conversational:2 easy_code:2 hard_code:2 knowledge_qa:2 trivial_nl:1
6	1	easy_code	easy_code:1
7	6	agentic_tool	agentic_tool:2 easy_code:2 hard_code:1 knowledge_qa:1
8	9	knowledge_qa	knowledge_qa:4 agentic_tool:2 hard_code:2 easy_code:1
9	10	conversational	conversational:4 trivial_nl:4 easy_code:1 knowledge_qa:1
10	24	conversational	conversational:12 trivial_nl:7 easy_code:3 knowledge_qa:2
11	4	trivial_nl	trivial_nl:2 easy_code:1 knowledge_qa:1
12	13	easy_code	easy_code:9 knowledge_qa:2 agentic_tool:1 trivial_nl:1
13	25	hard_code	hard_code:17 agentic_tool:6 easy_code:1 knowledge_qa:1
14	7	knowledge_qa	knowledge_qa:4 easy_code:2 trivial_nl:1
15	17	trivial_nl	trivial_nl:11 conversational:4 easy_code:1 knowledge_qa:1

[greptile-apps]

T-Rex Logs

What T-Rex did

• T-Rex attempted the Go test for TestTrexAgenticLadderV070HasTools, first running the base command and then the head-target command; both runs failed with exit code 127 and go: command not found.
• The environment blocker was identified: the PATH does not contain a go binary and network installation was disallowed, preventing Go from being installed.

_{Ran code and verified through T-Rex}

_{Reviews (1): Last reviewed commit: "fix(routing): gate agentic pool to harne..." | Re-trigger Greptile}