fix(agent): per-provider default subagent models; refresh retired Anthropic ids

[ericleepi314]

Problem

Subagents resolved their models through a stale global alias table (haiku → claude-3-5-haiku-20241022, sonnet → claude-sonnet-4-20250514). Both ids have been retired from the live Anthropic API (verified 2026-08-12 against GET /v1/models), so:

• Every Explore spawn on an Anthropic session died with 404 not_found_error — and the natural retry with model: "sonnet" died the same way (the goal screenshot: seven parallel verification agents, all 404).
• On DeepSeek sessions, haiku resolved to nothing servable and fell back to inherit, so every subagent fan-out silently ran (and billed) the expensive deepseek-v4-pro session model.

Fix — per-provider subagent defaults (reference-project mechanism)

Both references resolve subagent models per provider: TS Claude Code through its getDefault{Opus,Sonnet,Haiku}Model() functions, opencode through Provider.getSmallModel + the small_model config knob. Each provider's PROVIDER_INFO row may now declare:

provider	subagent_model (unspecified default)	subagent_tier_models
anthropic	claude-haiku-4-5 (cheapest current tier, $1/$5)	fable/opus/sonnet → claude-{fable,opus,sonnet}-5, haiku → claude-haiku-4-5
deepseek	deepseek-v4-flash	opus → v4-pro, sonnet/haiku → v4-flash

get_agent_model precedence: CLAUDE_CODE_SUBAGENT_MODEL env > tool model param > agent-def model: > provider default subagent model > inherit.

• Explicit inherit always forces the session model — Plan and fork agents pin it, the coordinator path pins it at the tool layer (workers can't pass a model param and do implementation work), and the workflow runner fills its slot as spec.model or agent_definition.model or "inherit" so agent() calls keep the documented "inherits the main-loop model" contract without clobbering an agentType agent's declared tier.
• The unspecified-model default is a deliberate divergence from both references (they inherit), per explicit user directive: fan-out subagents default to the provider's cheap tier. Escape hatches: model: inherit in a def or tool call, providers.<id>.subagent_model in config (an alias or inherit resolves like any user pin; providers is an untrusted-tier-blocked config key, so a committable repo config can't hijack it), or CLAUDE_CODE_SUBAGENT_MODEL.
• Same-tier-alias-keeps-parent precedes the table (a claude-sonnet-4-6 session asked for sonnet keeps 4-6). ANTHROPIC_DEFAULT_{OPUS,SONNET,HAIKU}_MODEL pins are honored on the anthropic provider only (a pin must not leak onto another vendor's wire). Custom Anthropic endpoints disable the table and inherit for haiku/sonnet (TS checkIsClaudeNativeProvider parity, incl. its opus asymmetry).
• Alias spellings are never shipped raw: known aliases resolve to their canonical target and take the availability gate (degrading to inherit, never a 404); only ids no alias table knows are trusted literally (proxy deployment names).
• Routing is visible: the Agent tool result and task registry carry the resolved model, the agent-progress emits (including terminal ones) carry it, and the TS bridge forwards it so the /agents overlay shows the real model instead of inherit.

BaseProvider.provider_id (new, set on all concrete + registry-generated classes) keys the lookup; fusion wrappers are unwrapped first.

Model-table refresh (same bug class)

• aliases: sonnet → claude-sonnet-5, opus → claude-opus-5, haiku → claude-haiku-4-5 (bare id resolves server-side — probed), new fable; claude-4-sonnet/s4 → claude-sonnet-4-6, claude-4-opus/o4 → claude-opus-4-8 (newest live 4.x); explicit legacy pins (claude-3.5-haiku, h35) kept and now degrade to inherit for subagents via the gate.
• registry: claude-sonnet-5 (1M ctx per docs, 128K output cap probed live, permanent $2/$10 launch pricing — a new tier, not 3/15) and claude-haiku-4-5-20251001 (200K window and 64K output cap probed live; prefix base covers the bare spelling). Prefix-fallback placement preserved — unknown 4.x sonnets keep the conservative 200K row.
• availability catalogs (both PROVIDER_INFO and AnthropicProvider.get_available_models) pruned to the live 2026-08-12 catalog + the probed bare haiku id: a listed-but-dead id turns the gate's "degrade to inherit" into a shipped 404. An invariant test pins every subagent target against both lists.
• Stale side-call defaults refreshed: settings (model → claude-sonnet-4-6, small_fast_model → claude-haiku-4-5), fast-mode, session-title/rename side calls, exec-agent-hook, count-tokens.
• Agent tool model enum: ["sonnet", "opus", "haiku", "fable", "inherit"], with the description noting the default is a fast/inexpensive tier and inherit is the full-capability escape.
• Deleted src/models/agent_routing.py — a dead parallel resolver whose inherit-parent semantics (and passing tests) contradicted the shipped path.

Verification

• Live DeepSeek e2e: real headless run; session per-model usage shows the main loop on deepseek-v4-pro and the Explore subagent on deepseek-v4-flash ($0.0014 for the subagent turn).
• Live Anthropic probes: /v1/models catalog (old ids gone, new ids served), output caps read off the API's own 400s, bare claude-haiku-4-5 proven to resolve server-side. (A full Anthropic e2e was blocked by an account extra-usage quota gate — a billing 400, not a model 404.)
• Full suite: 10052 passed, 0 failed, 12 skipped; ui-tui vitest 91 passed; tsc --noEmit clean.
• Critic loop: six adversarial review rounds to APPROVE. They caught real bugs at every stage — the config knob shipping alias spellings/inherit verbatim onto the wire, the workflow runner first taking the cheap default then (in my fix) clobbering agent-def tiers, non-family alias spellings trusted raw, the TS bridge dropping the model field so the overlay would claim inherit, retired ids surviving in the availability catalogs, coordinator workers with no opt-out, and cross-provider env-pin leakage. Each fix is pinned by a test (18-test TestPerProviderSubagentDefaults, runner slot test, bridge test, reporting tests, invariant test).

Follow-ups (deliberately out of scope)

• models/bedrock.py has no Claude 5-era rows (bedrock routing untouched by this path).
• The workflow journal records spec.model (often None) rather than the resolved model.
• LocalAgentTaskState.model is recorded but no backend endpoint serves it yet (the TUI type DelegationStatusResponse.active[].model is ready for it).
• Main-session default model stays claude-sonnet-4-6 (live, works); bumping it to sonnet-5 is a product decision.

🤖 Generated with Claude Code

[ericleepi314]

CI note: the test (ubuntu-latest, 3.11) failure is test_opencode_compat_providers.py::test_xai_requests_go_to_chat_completions — no HTTP request was attempted, which is pre-existing on main: the identical failure appears on merged PR #830's run (main's current HEAD, run 31571128350) and is an openai-v3-in-CI environment issue with a dedicated fix already in flight (fix/ci-windows-and-openai3). Local full suite on this branch: 10052 passed, 0 failed (4 consecutive runs). The Windows legs are the known pre-existing failures.