v0.9.0 Harness/Profile cutline: model posture before automatic harness creation

[Hmbown]

Problem

The Agentic Harness Creator is important, but it is not the first harness-profile PR. CodeWhale needs explicit model posture and profile resolution before any agent tries to evolve profiles automatically.

MVP order

1. HarnessPosture enum/policy surface — v0.9.0 HarnessPosture: model-specific context and subagent policy #2693
2. HarnessProfile schema/registry — v0.9.0 HarnessProfile schema: provider/model/posture-aware profile registry #2696
3. HarnessProfile resolver — v0.9.0 HarnessProfile resolver: apply prompt, tool, context, and subagent policy #2697
4. Seed profiles — v0.9.0 Seed HarnessProfiles: PrefixCached DeepSeek/MiMo and LeanRoot generic #2701
   • PrefixCached DeepSeek/MiMo
   • LeanRoot generic/local/open-weight
   • explicit Arcee handling so Trinity Thinking does not get lost in DeepSeek/MiMo assumptions
5. Repo overlay with .codewhale/constitution.json — v0.9.0 Repo Harness Overlay: compose .codewhale/constitution.json with model HarnessProfiles #2702
6. UX/status display — v0.9.0 HarnessProfile UX: show provider, model, profile, and repo law clearly #2703
7. Evidence store — v0.9.0 Harness Evidence Store: profile-tagged traces for harness evolution #2698
8. Candidate manifest — v0.9.0 HarnessCandidate manifest: falsifiable profile changes from trace evidence #2699
9. PromotionGate — v0.9.0 HarnessProfile PromotionGate: replay, eval, and rollback for profile updates #2700
10. Harness Creator workflow — v0.9.0 Harness Creator workflow: inspect, propose, evaluate, promote, rollback #2704
11. Full Agentic Harness Creator epic — v0.9.0 Agentic Harness Creator: evolve per-model CodeWhale harnesses from trace evidence #2695

Acceptance criteria

• The first implementation PR stops after schema/resolver/seed-profile/display unless tests justify more.
• Provider route, endpoint, model id, HarnessProfile, and repo constitution are separately visible.
• DeepSeek V4 and Xiaomi MiMo v2.5 can use cache-heavy posture; generic/HF/local models can use lean/subagent-oriented posture.
• Arcee is explicitly covered in resolver tests and smoke docs.
• No profile is silently promoted or mutated.
• Docs cite AHE/GEPA as inspiration only when discussing future automatic evolution.

Non-goals

• Do not implement autonomous harness mutation before evidence, manifests, replay, and promotion gates exist.
• Do not change provider auth or tool allowlists as part of the first profile schema PR.

[Hmbown]

Progress update: #2844 (efbcc681a) added docs/HARNESS_PROFILE_CUTLINE.md and linked it from README/config docs.

What is now documented:

• HarnessProfile work must start with posture/schema/registry, deterministic resolver, seed profiles, repo constitution input, and status display before any automatic Harness Creator.
• DeepSeek V4 and Xiaomi MiMo are called out for cache-heavy posture while keeping their routes/auth separate.
• Direct Arcee trinity-large-thinking is explicitly separate from the OpenRouter arcee-ai/trinity-large-thinking alias.
• Generic/HF/local/open-weight routes should use lean or standard posture until matched explicitly.
• AHE/GEPA are framed as future inspiration only after candidate proposal, replay/eval, promotion gate, and overlay rollback semantics exist.
• Normal Agent/WhaleFlow runs must not silently mutate or promote harness profiles.

Verification for the docs PR: git diff --check and ./scripts/release/check-versions.sh.

Leaving this issue open because the remaining acceptance criteria need implementation evidence: deterministic resolver tests, seed profiles, Arcee resolver/smoke evidence, and a status surface showing provider, model, profile, and repo constitution separately.

[Hmbown]

v0.9 stewardship follow-up: #2845 is merged as e22a7da53, aligning the release-facing HarnessProfile docs with the #2844 cutline.\n\nWhat changed: the v0.9 acceptance matrix now explicitly says we ship the typed schema/config foundation and defer runtime resolver, seed-profile selection, telemetry, and status display until verified follow-up slices. config.example.toml also has a commented dormant [[harness_profiles]] example so users can see the schema without assuming active runtime behavior.\n\nKeeping this issue open for the remaining acceptance work: deterministic resolver tests, seed profiles, Arcee resolver/smoke evidence, and visible provider/model/profile status surfaces.

[Hmbown]

Status update: #2850 (e60eeb816) added the next safe HarnessProfile foundation slice on the v0.9.0 stewardship branch: a pure ConfigToml::resolve_harness_profile(provider_route, model) helper plus HarnessProfile::matches_route.

What this now proves:

• provider aliases normalize during matching, e.g. mimo resolves against xiaomi-mimo and deepseek-cn resolves against deepseek
• model patterns support deterministic exact / * / ? matching without adding a new regex dependency
• the first matching configured profile wins
• missing route/model combinations return None
• resolving a profile does not change resolve_runtime_options, provider routing, prompts, tools, auth, context, or persisted config

Verification:

• cargo test -p codewhale-config harness_profile --locked
• cargo fmt --all --check
• git diff --check
• cmp -s CHANGELOG.md crates/tui/CHANGELOG.md
• ./scripts/release/check-versions.sh
• ./scripts/release/check-ohos-deps.sh

This keeps the v0.9 cutline intact: schema/resolver foundation can ship, while runtime provider/model posture selection, seed-profile activation, telemetry, and status display remain deferred to later explicit PRs.