feat: cw 1.0.1 stall ladder, lean loop and session notifier

[victorkzam]

Summary

• Stall prevention in the agents. reviewer, design-reviewer and direction-reviewer drop memory: project (fresh context is their point; memory writes cost turns). Reviewer turn caps are sized to measured need (40 for deep review, 15 for direction), both reviewer schemas gain an unchecked list so a partial report is structured rather than a stall, and every agent's Output section is a bounded contract: full report to the file the prompt names, final message of at most about 1,500 tokens.
• Stall ladder with one home. rules/orchestration.md replaces the "retry once fresh" bullet with the ladder: ground truth first (git log and the verification command for an implementer, the report file for a reviewer), two SendMessage nudges to the same agent, one briefed respawn, then escalate. The build and design skills point to it in one line each.
• Lean orchestration loop (plan C). /cw:design gains the lite path, a two-round review cap, delegated synthesis research and bounded briefs. /cw:build gains bounded briefs, commit by pathspec, and the stranded-implementer gate.
• Session-labelled desktop notifier. hooks/notify.sh on Notification, Stop and StopFailure: one ping titled <folder> · <session> only when a human is needed; the built-in idle_prompt is suppressed while background agents still run; presence policy ships behind a config flag, off by default; opt-in via a config file, silent under the SDK entrypoint and CW_NOTIFY=0.
• Docs and release. README, WORKFLOW and the design draft describe 1.0.1; plugin version bumped to 1.0.1 with the marketplace description.
• Test fix (added after the first CI run). The four branch-guard speed cases in tests/run.sh judged the hook with bash's integer SECONDS counter against a budget of 1 tick; the heaviest payloads take about 0.65-0.75 s locally, and the macOS runner read 2 ticks on case 85 with the correct exit code, failing the first pull_request run while the push run on the same commit passed. The budget is now 2 ticks (under 3 s of wall clock) with one retry on a timing-only miss; a wrong exit code still fails at once.

Why: three measured problems since 1.0.0, recorded in the plan's Context section. Deep-review agents ended without a verdict in 4 of 4 and then 2 of 4 runs, while a single resumed-turn nudge recovered every one. One four-phase session cost about 1.6M sub-agent tokens plus a full orchestrator window. The notification hook produced 110 banners in a day, most of them idle_prompt while agents were still running, with no session label.

Test Plan

• bash tests/run.sh all on e4b797c: hooks 154 passed, 0 failed; size 64824 bytes (87.9% of baseline); dedupe, settings-keys, trailers and scan all PASS. shellcheck -x hooks/*.sh tests/run.sh clean. /bin/bash tests/run.sh hooks under macOS system bash 3.2: 154 passed.
• Speed-case helper exercised with fake hooks under bash 3.2: fast hook passes first try; wrong exit code fails on attempt 1; a 3.2 s hook is invoked twice and reported as a timing failure with attempt=2.
• Hands-on check: run a deep cw:reviewer review under the new turn cap and confirm a verdict lands, or that nudge 1 of the ladder recovers it in one turn.
• Hands-on check: with the notifier enabled, end a turn while a background agent is running and confirm no idle_prompt banner until the agent finishes; confirm the banner title reads <folder> · <session>.
• Hands-on check: /cw:design lite <description> runs a single review round and skips the direction pass.

Design Reference

docs/plans/stall-handling-1-0-1.md

🤖 Generated with Claude Code

https://claude.ai/code/session_01STufi6776QVKvuDcJv4RPM