[gering]

Summary

• On herdr 0.8.2, agent start returns only once the agent is idle. A native claude:* worker handed /work-system:continue is busy from the first second, so every such kickoff ended blocked=unverified while the worker ran, and herdr never registered the agent name (agent get <label> → agent_not_found).
• herdr-launch.sh now confirms a code=timeout start from process state and names the worker; other errors and an undetected worker stay fail-closed.
• Bundles fix-flaky-herdr-launch-test: seams for the literal 10s probe bounds, pinned in the test harness, plus the residual cause.

Changes

• Timeout confirmation (timeout code only): bounded poll of pane get (WORK_SYSTEM_HERDR_CONFIRM_SECONDS, 10s) for the expected kind in that pane → moved=yes; nothing or a wrong kind → blocked=unverified. Nothing is retried, no second worker.
• Naming: agent rename <pane> <label> after confirmation, best-effort (agent_name_taken is reported on stderr, the launch stays confirmed).
• herdr's timeout error is printed only when confirmation fails.
• Test seams: WORK_SYSTEM_HERDR_PROBE_SECONDS, WORK_SYSTEM_AGENT_PROBE_SECONDS replace the eight literal run_bounded 10 (production default unchanged); Env.__init__ pins them and HA_CALL_TIMEOUT_SECS to 120s.
• Hermetic tests: detected → moved+rename, name taken, nothing detected, wrong kind.
• Knowledge (herdr-kickoff-automation.md): 0.8.2 readiness semantics + the stub-shift gotcha. work-system 1.28.2.

Live measurements (herdr 0.8.2)

• pane get reports agent=claude 0.6s after the start; agent start --timeout 10000 timed out at 10.4s with the agent working. A start whose /work-system:status turn took 16s returned at 17s — readiness = idle.
• After the timeout the name is never registered, even once idle; agent rename <pane> <name> fixes it (idempotent).
• A timeout can also mean nothing was typed (3s start into a still-busy shell → bare prompt) — hence detection, not blanket success.
• agent_not_ready ("blocked during startup") is CC's trust dialog in an untrusted cwd — stays unverified.
• End to end, the patched script on a throwaway worktree: herdr error code=timeout → moved=yes, agent reachable by its label.

Flaky test: cause and measurement

• Not reproducible at current load: baseline 0/12, 0/20, 0/40 concurrent runs before the change.
• Mechanism, measured: the stub counts the Nth call on entry; a bounded call killed after a stall has already advanced the sequence, so every later answer shifts. One 2s pane get under a 1s bound turns wrapper wrong kind: stops immediately from 1 poll into 2; a 0.15s bound reproduces exactly the five modern --grok assertions seen before (the grok models probe). This explains every assertion in the original report.
• Ruled out: concurrent herdr calls within one launch (none), the real cc-harness-agents on PATH (not called by resolve).
• Acceptance after the change: for k in 1 2 3; do <20 concurrent runs>; done → 0/20, 0/20, 0/20 (198–235s each).

Readiness

• ✅ Version 1.28.2 (plugin.json + marketplace.json)
• ✅ CHANGELOG
• ✅ Knowledge
• ✅ scripts/check-structure.py (incl. plugin tests)
• ➖ Build (none)

Test plan

• CI check-structure.py green
• Next real claude:* kickoff on herdr 0.8.2 returns moved=yes and herdr agent get <label> resolves

🤖 Generated with Claude Code

https://claude.ai/code/session_01FqXCdJZeE2VMqC9dZ1Qnh6