feat(plan): full autonomy by default, retire stale model priors

Merged
TI
2mo ago
+109-34

The question budget now defaults to ZERO at every scale — /plan infers, states each inference in one line the user can correct in passing, and goes straight to the plan. Scope/ambition/technical-depth questions are abolished (the model decides and offers a one-line opt-down). Budget above zero comes only from the /plan-tune autonomy dial, which is now actually consumed — users who want more conversation tune it down; nobody configures anything to get the autonomous default.

Also retires the human-era anchors (file-count complexity smells, >15-files→REDUCTION, human-hour estimates, five-question interview gates in spec) and stamps a dated, sourced capability-priors rule into all five plan specialists: never justify narrowing scope by asserting an LLM can't do something — 2026 evidence says pairwise matching is at 92–99 F1 with the error direction inverted vs rule engines; uncertain capabilities get a 30-minute measurement spike, not a product downgrade.

Live-verified against the transcript that motivated this: same prompt, zero questions, ambitious plan with decisions stated and reversible, 5 turns. Free suite green.

🤖 Generated with Claude Code

Summary by cubic

Make /plan fully autonomous by default with a zero-question budget at every scale, so it infers decisions, states them in one line, and moves straight to the plan. Autonomy is opt-down via /plan-tune, and stale file-count/human-hour heuristics are removed in favor of evidence-based capability priors.

  • New Features

    • Chain-wide question budget defaults to zero; decisions are inferred and stated with a one-line opt-down.
    • Autonomy budget comes from ~/.gstack/developer-profile.json using round((1 − autonomy) × 6); absent profile = autonomy 1.0.
    • Consent gates sit outside the budget; if unbundled, skip the action and note it in one line.
    • Allow batching independent questions in one round; still counts per question.

  • Refactors

    • Retired file-count triggers (e.g., “>15 files → reduction”) and human-hour estimates; size by agent sessions + review time.
    • Eng review now flags unjustified structure, not diff size; challenge new components that duplicate existing ones or built-ins.
    • autoplan auto-approves expansions inside blast radius with no new infra or authority boundary; file count no longer a proxy.
    • spec self-answers discovery, marks DECIDED choices, and sizes issues by review-ability, not human build-days.
    • Added dated capability priors across plan specialists: do not narrow scope by claiming model limits; measure uncertain, load-bearing capabilities with a 30-minute spike (~50 labels, 85–90% agreement).