perf(meta-phases): let the model decide when to browse, not the prompt

[AbirAbbas]

Why

Each meta-phase (meta_semantic, meta_mechanical, meta_systemic) was burning 15–40 minutes apiece in production, even though Kimi K2.6 is fast per-turn. Profiling traced this back to the prompts themselves, not to the model or harness.

Two pieces of prompt language were doing the damage in all three lenses:

1. "Investigation Protocol" — explicitly told the model "browse the actual source files... Read the changed functions. Then find their callers. Trace how data flows..."
2. "Quality Gate" — explicitly rejected diff-only answers: "If your rationale says 'visible in the diff' or 'based on the patches', you have not investigated enough."

That turned every phase into a 20–40 turn opencode session of file reads, when the diff patches plus the intake/anatomy summary already contain almost everything these phases need to produce dimensions. The model wasn't slow — it was being instructed to do expensive exploration before it was allowed to answer.

What

Replaced the mandate with neutral framing in all three meta-phase prompts:

• "Working with Context": the diff + intake/anatomy summary is usually enough. Repo access is still available (cwd=repo_path); use it sparingly, only when the diff genuinely doesn't show what you need (e.g., to confirm an unchanged caller's signature).
• "Quality Bar": each dimension must be SPECIFIC — named files, function names, line ranges. The diff already carries those, so a dimension grounded in the diff is good. Vague dimensions are still rejected. What's no longer penalized is correct dimensions produced from the diff in 2 turns instead of 30.

Each lens keeps its existing bullets (semantic = logic/contracts/concurrency/security, mechanical = signatures/imports/decorators, systemic = patterns/coverage/migrations), the target_files requirement, and the output schema. The structural contract is unchanged.

Also in this PR

Bumped the in-code default model openrouter/moonshotai/kimi-k2.5 → kimi-k2.6 in config.py and .env.example to match what's actually deployed via Railway env. Local dev was silently running on an older model than prod.

Expected impact

This is one half of a two-PR change. The other half (in agentfield) addresses three orthogonal harness issues:

1. The 600s opencode timeout, which kills slow-but-progressing calls
2. Auto-retrying the entire opencode call on timeout (currently TRANSIENT_PATTERNS includes "timeout", so a 10-min timeout silently triggers up to 3 more 10-min retries)
3. Schema validation failures rerunning the entire opencode subprocess instead of just reformatting the existing text output

This PR alone should chop meta-phase wall time roughly in half by simply not forcing useless exploration. With the agentfield fixes layered on top, total review runtime should drop from 60–110+ min to well under an hour for typical PRs.

Test plan

• PYTHONPATH=src pytest — 25 passed (1 pre-existing test_cost_tracker failure on main, unrelated to this PR)
• Smoke imports of anatomy_phase, meta_semantic, meta_mechanical, meta_systemic resolve cleanly
• After merge: trigger a real review on a multi-file PR and compare per-phase durations against baseline (anatomy: 903s, meta_*: 800–1000s each in the run that prompted this PR)

🤖 Generated with Claude Code