v0.9.0 HarnessPosture: model-specific context and subagent policy

[Hmbown]

Goal

Make CodeWhale's harness strategy explicit per provider/model route instead of assuming every model wants the same amount of up-front system context.

This came out of v0.8.53 testing: DeepSeek V4 and Xiaomi MiMo v2.5 appear likely to benefit from a cache-heavy/prefix-stable starting prompt, while many other routes may do better with a lean root context, immediate exploration, subagents, and compact handoff packets.

This is related to #2672, but this issue is specifically about runtime harness posture: how much context to load up front, when to encourage subagents, when to compact, and when to start fresh from a handoff.

Current Shape

• Plan mode already has a read-only design-first prompt in crates/tui/src/prompts/modes/plan.md.
• Plan mode currently resolves via update_plan and opens an approval prompt from crates/tui/src/tui/ui.rs / crates/tui/src/tui/plan_prompt.rs.
• Tool exposure is already mode-sensitive in crates/tui/src/core/engine/turn_loop.rs.
• Provider/model routing work is happening around the config/provider registry, but there is not yet a first-class policy object that says how a given route should use context, subagents, and handoffs.

Proposal

Add a central HarnessPosture policy chosen from provider + model + user override.

Possible initial values:

• PrefixCached: stable, rich system/context prefix; prefer cache byte stability; minimal churn. Good default candidates: DeepSeek V4, Xiaomi MiMo v2.5.
• LeanRootExplore: minimal starting prompt; strongly prefer quick repo orientation, subagent exploration, and on-demand docs/skills.
• PlanHandoffReset: plan in one context, then launch implementation from a compact approved handoff packet in a fresh context.
• SmallContextLocal: aggressive summarization, narrow tool surface, small prompts, cheap parallel probes.

The posture should affect:

• System prompt assembly and context injection volume.
• Whether Plan mode recommends handoff reset by default.
• Subagent encouragement and default delegation copy.
• Context compaction thresholds.
• Which docs/skills are eagerly injected vs available on demand.
• Telemetry labels so we can evaluate posture choice rather than argue from vibes.

Acceptance Criteria

• A single registry/policy layer maps provider+model route to default HarnessPosture; no scattered provider-name conditionals.
• User config can override posture without changing provider/model identity.
• Tests assert at least:
  • DeepSeek V4 and MiMo v2.5 select a cache-heavy posture by default.
  • Generic OpenAI-compatible/OpenRouter/local routes can select a lean or handoff posture.
  • Prompt/tool catalog byte stability remains protected for prefix-cached routes.
• Docs explain provider vs model vs harness posture as separate concepts.
• Telemetry/logging records posture for later outcome analysis.

Non-goals

• Do not hardcode benchmark-specific behavior.
• Do not claim one posture is globally better until we have eval data.
• Do not remove the existing global identity/constitution preamble as part of this work.