v0.15.0

0.15.0 — per-agent model routing + preset-aware defaults (2026-05-18)

Token-cost optimization across Claude Code / Cursor / Codex via declarative
per-agent model pinning. 13 ADRs locked in PLAN-model-routing-multi-ide.md.
8 implementation phases shipped via /hm:loop with per-phase /hm:review.

Added

• Per-agent model schema (ADR-001/002) — new HarnessConfig.agent_models: dict[str, AgentModelSpec] with nested {claude, cursor, codex: {model, reasoning_effort}}. recommended_model: str renamed to default_model: str
  (deprecated read-side property kept for 0.15.x / 0.16.x; removed no earlier
  than 0.17.0 per ADR-012).
• Preset-aware defaults (ADR-005) — new src/harness_maker/presets.py
  ships PRESET_AGENT_MODELS for Production (opus on 3 reasoning agents,
  sonnet on 11 reviewers) and Side (sonnet everywhere with downshifted
  reasoning_effort). 3-tier resolve_agent_spec(): explicit override →
  preset map → _spec_from_default_model fallback (never KeyErrors on
  user-authored agents).
• Canonical Cursor ID table (ADR-003) — CURSOR_MODEL_IDS maps aliases
  (opus/sonnet/haiku) to concrete IDs. Users write aliases in
  agent_models; renderer normalizes via this table at render boundary
  (single-point upgrade across Claude releases). Templates lint-enforced
  against raw concrete IDs.
• Codex profiles (ADR-008) — .codex/config.toml now renders
  [profiles.cheap] (reasoning_effort=minimal) + [profiles.deep]
  (reasoning_effort=high). Codex agent TOMLs render
  model_reasoning_effort per-agent (the dominant cost lever; keep model =
  omission per RESEARCH-codex-plan-validator-model-unavailable).
• /hm:health Layer-1 sub-check (ADR-010) — new model_routing dimension
  with 3 advisory signals: Claude #43869 reliance, Cursor alias-form
  warnings, Codex reasoning_effort coverage. Weight 0 (advisory only;
  doesn't change composite).
• CLI --default-model flag + back-compat alias --recommended-model
  (ADR-012) with DeprecationWarning.
• --update cwd guard (ADR-013) — rejects snapshot regen invoked from
  inside .worktrees/<branch>/, turning the documented footgun
  ([fail:snapshot-regen-inside-worktree] count:4) into enforced
  prevention with actionable error message.
• Silent schema migration (ADR-004 + ADR-011) — recommended_model: in
  v1 harness.yaml migrates silently to default_model; INFO log gated on
  schema_version<2 to avoid noise on fresh v2 renders. Multi-doc YAML
  provenance frontmatter handled via io_utils.load_harness_yaml().
• HOW-IT-WORKS docs — new "Agent Models" section with worked example
  covering preset defaults, per-agent override, and the 3-tier resolution chain.

Changed

• HarnessConfig.schema_version: 1 → 2.
• 14 agent .md.j2 templates: hardcoded model: opus|sonnet → model: {{ claude_model }} (context driven by resolve_agent_spec).
• 2 preset YAML templates + 5 foreign-config templates: recommended_model:
  → default_model: rename.

Fixed

• Pydantic dual-key handling for AliasChoices + extra="forbid" —
  model_validator(mode="before") silently drops recommended_model when
  default_model is also present (avoids extra_forbidden on rendered
  output round-trip).
• agent_models parse path catches pydantic.ValidationError in addition
  to TypeError/ValueError so a malformed override drops with a WARNING
  log instead of silently nuking the whole answers_from_harness_yaml
  return (Phase 2 /hm:review consensus-passed P1 fix).
• Migration log message sanitizes newline + ANSI escape sequences in
  user-provided values (security-reviewer P1 fix).
• Migration log includes the harness.yaml path so multi-repo runs can
  identify which file triggered the advisory (code-reviewer P1 fix).