feat(models): SWE_MODEL_LOW/MED/HIGH per-tier model env vars

[AbirAbbas]

Summary

Brings the claude_code-style tier approach to every runtime (open_code, codex, claude_code) and both implementations (Python + Go): three env vars map the 17 agent roles to models by capability tier, instead of enumerating 17 models.<role> keys.

SWE_MODEL_HIGH=openrouter/z-ai/glm-5.2#high        # planning-heavy: pm, architect, tech_lead, replan
SWE_MODEL_MED=openrouter/deepseek/deepseek-v4-pro  # coding/review/QA (coder, qa, code_reviewer, ...)
SWE_MODEL_LOW=openrouter/deepseek/deepseek-v4-flash # mechanical: qa_synthesizer, git

Precedence (lowest → highest): runtime base defaults → SWE_DEFAULT_MODEL/AI_MODEL/HARNESS_MODEL → tier env vars → models.default → models.<role>. With no tier vars set, resolution is byte-identical to today. SWE_MODEL_HIGH also becomes the default for the standalone plan pipeline (its reasoners are high-tier).

The optional #variant reasoning-effort suffix rides on the model string and is consumed by the harness providers (opencode --variant, codex model_reasoning_effort) — Agent-Field/agentfield#801, which implements it in all three SDKs (Python/Go/TS). Plain tier models work with current SDK releases; once #801 ships, the agentfield floor here should be bumped so Docker builds pick it up.

Changes

Python: swe_af/execution/schemas.py — MODEL_TIERS, ROLE_TO_TIER, TIER_MODEL_ENV_VARS, _tier_models_from_env(); tier layer in resolve_runtime_models(); _default_planning_model() consults SWE_MODEL_HIGH first. .env.example documented. 15 new tests (tests/test_model_tiers.py).
Go: go/internal/config/resolve.go — RoleToTier + tierModelsFromEnv() at the identical precedence point; DefaultPlanningModel() consults SWE_MODEL_HIGH first. 15 subtests mirroring the Python contract 1:1 (modeltiers_test.go); env-isolation lists extended in config_test.go / orch/plan_test.go.

Validation

• Python: make check (literal CI steps, fresh 3.12 venv, committed tree): 1052 passed, 1 skipped
• Go: CI go job reproduced literally (GOWORK=off, Go 1.23 toolchain, pinned sparse SDK clone): gofmt/build/vet/test -race — all 26 packages ok
• Live end-to-end (runtime=open_code, tier vars above): all 17 roles resolved to their tier's model, models.coder override still wins, and opencode session records confirmed architect served by z-ai/glm-5.2 (variant high), coder by deepseek-v4-pro, qa_synthesizer by deepseek-v4-flash

Notes

• Fast mode (swe_af/fast/) has no env cascade at all today, so tiers were deliberately not mirrored there — separate decision.

---

🤖 Generated with Claude Code