docs(design): forward-looking improvement proposals

[fazpu]

What

Adds design/designs/improvement-proposals.md — the forward-looking companion to the success-semantics design doc (#53). Where #53 records decisions already made (and not to be reverted), this doc collects the changes we are considering, curated from the July 2026 review under design/analysis/ and re-prioritized. Every item is marked Proposed, not Accepted. No code changes.

Stacked on #53. Based on the design/success-semantics-and-evaluation branch so its cross-references resolve; GitHub will auto-retarget this to main once #53 merges. Review/merge #53 first.

Framing decision that drives priority

ugm will be driven by the planner/dispatcher double loop from the very beginning — not after a single-loop pilot. That moves the parent/child machinery from "harden later" to harden first, because it's exercised on day one of the flagship use case. The ordering reflects that.

Priorities

P0 — double-loop foundations

• P0.1 Durable session-stack / active-child recovery — fixes the untested crash gap where a coordinator restart reopens the parent and orphans a running child.
• P0.2 First-class paused / waiting_for_human state + typed gate/external-action requests (a governance pause is not an error).
• P0.3 Per-child budgets (today a child inherits max_turns: 120).
• P0.4 Make the PM template runnable from a clean init (it doesn't bundle its inner_outer_eval child set today).

P1 — legibility & safety

• P1.1 events.jsonl + usage/cost ledger, then loopy status --watch.
• P1.2 Deterministic backstop under the judge for high-stakes targets (e.g. ugm) — shells out to the repo's own contract tests; explicitly does not revert the eval decision in docs(design): record success-semantics and evaluation decisions #53.
• P1.3 Fix the stale Agent Skill / "multi-worker" claims (pre-launch blocker).

P2 — hardening & DX

• P2.1 Named model profiles + pin the trio (team-harness<0.3, eval-banana as an extra).
• P2.2 _advance() refactor, folded into the P0.1 state-machine work.
• P2.3 Richer failure taxonomy + per-workflow failure cap.
• P2.4 doctor / validate / session-aware status + stop --force.

Explicitly deferred: parallel loopy workers, web dashboard, breadth-first child sessions.

Suggested sequencing

P0.1+P0.2+P2.2 (one coherent state-machine piece) → P0.3+P0.4 (double loop safe + runnable) → P1.1 (legible) → P1.2+P1.3 → P2.x rolling. After that, the double loop is ready to drive ugm — the subject of a separate execution design, not this doc.

🤖 Generated with Claude Code

https://claude.ai/code/session_01PN8aFwwxA8FzMy9LgpbQFt

[fazpu]

Superseded by a clean, generic version based on main (see the new improvement-proposals PR). Closing this stacked branch.