spec: opt-in decision-model gate for subagent delegation and (model, effort) tier selection

[Yeachan-Heo]

Migrated from #5842 (originally opened by @Yeachan-Heo).

Summary

Spec-only issue. Requested by @probe4206 for local prototyping — no implementation is being started here.

Two related asks:

1. Nudge (or gate) toward subagent delegation when a turn should have delegated but didn't.
2. Decide subagent (model, effort) dynamically instead of pinning it per role.

Both reduce to one cheap classification call per decision point, which is what a small decision model (Jev / the open-source Kev clone) is built for.

Why effort must be decided at the subagent boundary, not mid-session

Changing effort inside a live session invalidates prompt cache on both major providers:

• Anthropic: "Changing thinking parameters (switching modes, or changing the budget in extended mode) invalidates cached message prefixes, and can invalidate cached system prompts and tools as well, because the thinking configuration is rendered into the prompt." output_config.effort is documented the same way.
• OpenAI: "Keep the top-level reasoning.effort at its original value as changing that setting can rewrite instructions in the hidden system instructions." Measured in Changing reasoning level results in cache miss openai/codex#35416 as an effectively full miss; opencode observed 100%.

A subagent starts its own session and shares no prefix, so per-subagent (model, effort) selection costs nothing in cache terms. That is the safe place to be dynamic.

(OpenAI GPT-6+ also exposes a configuration_update input item that changes effort mid-conversation while preserving the cached prefix. Out of scope here, but relevant if in-session dynamic effort is revisited.)

Existing surfaces to build on

• packages/coding-agent/src/config/autorouting-tier-map.ts — tier assignments already validate an optional assignment.effort against an EFFORTS set, enforce at most one assignment per tier, and check provider/tier rank collisions. The (model, effort) pair is already the unit of routing.
• packages/coding-agent/src/hooks/events.ts — HOOK_EVENT_SCHEMAS, HookAuthority, and HookErrorBehavior (isolate | fail-closed) already distinguish advisory from blocking hooks.

Proposed shape

Decision model exposes three primitives — noul (yes/no probability), choice (named options), score (ordered scale) — so one request covers both asks:

• noul: "should this work be delegated to a subagent?"
• choice: which tier, resolved through the existing tier map to a concrete (model, effort) pair.

choice scores arbitrary option sets against the question, so adding or renaming a tier needs no retraining — unlike a fixed-label classifier.

Tiers stay a small fixed set (e.g. fast / balanced / strong). Continuous per-request effort is explicitly rejected: it produces a unique cache lineage per request.

Role defaults follow call frequency — executor is called most often and defaults low; planner/architect are called once or twice per task and can default high.

Three modes, default off

Mode	Hook behavior	Effect
off	not registered	default; zero overhead
hint	isolate	surfaces a delegation suggestion; a hook failure never blocks the turn
enforce	fail-closed	blocks direct edits when the gate fires

enforce must stay closed until a threshold is validated on our own data. Kev's published limitation: Kev-4B assigns >=90% confidence to a wrong answer on 8.2% of new-source development questions (Kev-9B: 7.5%), and its README states plainly: "Test it on your own data before choosing a probability threshold." At that rate a blocking gate would reject correct direct edits often enough to be worse than no gate.

Decision criteria must be observable

Whatever triggers the noul question has to be computable from turn state — files touched, consecutive edit count, diff size. "Looks complex" is not implementable.

Domain adaptation

Public checkpoints will not match this repo's categories. Kev documents kev.train --init_from, which continues from a released checkpoint's LoRA adapter and pointer head. Their reported before/after on 836 labeled examples: training from the base model collapsed to 0.33 on the held-out set, while --init_from held 0.83 there and reached 0.88 on the new domain. A few hundred labeled decisions is the realistic budget.

Risks to size before building

• Latency. Kev-9B is ~2s per five-question request on an M5 in bf16 (Qwen3.5 generation lacks fast Gated DeltaNet kernels on Apple GPUs); the older Qwen3-based generation is ~300ms. Per-tool-call invocation makes this user-visible.
• Local server has no authentication by default and binds 127.0.0.1. Fine for single-machine prototyping, must be closed before anything wider.
• Option order can change the answer — question isolation does not isolate options within a question.
• Accuracy does not transfer across task types: Jev leads on SemIf's 144 authored decisions (0.965 vs 0.917) while Kev-9B leads on scienthoon's 900 support tickets (0.952 vs 0.897). Measure on our own representative task.

Acceptance for a future implementation

• Mode defaults to off; hint and enforce are explicit opt-in.
• Tier resolution reuses the existing tier map rather than introducing a parallel table.
• Threshold selection is backed by measurements on this repo's decisions, not a borrowed default.

Sources: Anthropic prompt-caching / extended-thinking docs, OpenAI prompt-caching guide, openai/codex#35416, and explainx.ai's write-up of jaredpalmer/kev.

—
[repo owner's gaebal-gajae (clawdbot) 🦞]

—
[repo owner's gaebal-gajae (clawdbot) 🦞]

[Yeachan-Heo]

Closing per maintainer decision: not actionable at this time. Can be reopened if the need resurfaces.

—
[repo owner's gaebal-gajae (clawdbot) 🦞]