fix(manager): stop repeated diagnostic loops

[RerankerGuo]

Summary

• add Manager prompt safeguards against repeating the same no-op or malformed troubleshooting command
• define Worker deletion completion as the target disappearing from hiclaw get workers
• update worker lifecycle guidance to avoid looping on Matrix room probes such as copaw channels list after deletion

Fixes #974.

Tests

• git diff --check
• rg -n "Diagnostic loop safeguard|Worker deletion is complete|Delete Completion Boundary|Stop repeated diagnostics" manager/agent/AGENTS.md manager/agent/copaw-manager-agent/AGENTS.md manager/agent/skills/worker-management/references/lifecycle.md manager/agent/skills/worker-management/SKILL.md

Temporary dependency

• This branch is temporarily stacked on test: make multi-worker bob setup deterministic #985 (78e9385), whose full CI passed all 10 integration matrices.
• Relative to test: make multi-worker bob setup deterministic #985, this PR retains only its original scoped changes; the shared test fix disappears from the diff after test: make multi-worker bob setup deterministic #985 merges.
• Fresh checks on 90c9fd4f passed every build and all 10 integration matrices.

[github-actions]

📊 CI Metrics Report

Summary

Metric	Current	Baseline	Change
LLM Calls	95	81	+14 ↑ +17.3%
Input Tokens	2982071	2803871	+178200 ↑ +6.4%
Output Tokens	19839	16791	+3048 ↑ +18.2%
Total Tokens	3001910	2820662	+181248 ↑ +6.4%

By Role

Role	Metric	Current	Baseline	Change
🧠 Manager	LLM Calls	82	68	+14 ↑ +20.6%
	Input Tokens	2702276	2502214	+200062 ↑ +8.0%
	Output Tokens	17398	13725	+3673 ↑ +26.8%
	Total Tokens	2719674	2515939	+203735 ↑ +8.1%
🔧 Workers	LLM Calls	13	13	0 — 0%
	Input Tokens	279795	301657	-21862 ↓ -7.2%
	Output Tokens	2441	3066	-625 ↓ -20.4%
	Total Tokens	282236	304723	-22487 ↓ -7.4%

Per-Test Breakdown

Test	Mgr Calls	Wkr Calls	Δ Calls	Mgr In	Wkr In	Mgr Out	Wkr Out	Δ Tokens	Trend
02-create-worker	7	0	-5 ↓ -41.7%	186447	0	862	0	-171313 ↓ -47.8%	✅ improved
03-assign-task	11	6	+2 ↑ +13.3%	303434	115858	2262	851	-51251 ↓ -10.8%	⚠️ regressed
04-human-intervene	18	0	+5 ↑ +38.5%	468392	0	2872	0	+38266 ↑ +8.8%	⚠️ regressed
05-heartbeat	9	0	+2 ↑ +28.6%	290119	0	1863	0	+16730 ↑ +6.1%	⚠️ regressed
06-multi-worker	37	7	+10 ↑ +29.4%	1453884	163937	9539	1590	+348816 ↑ +27.2%	⚠️ regressed

Trends

✅ 1 test(s) improved (fewer LLM calls)
⚠️ 4 test(s) regressed (more LLM calls)

---

Generated by HiClaw CI on 2026-07-13 16:03:01 UTC

---

📦 Download debug logs & test artifacts

[github-actions]

❌ Integration Tests Failed (llm-interaction / mgr=copaw / wk=hermes)

Commit: 38e4c45
Workflow run: #1503

Test Results

No test output captured.

Debug Log (tail)

No debug logs available.

---

📦 Download full debug logs & test artifacts

[github-actions]

❌ Integration Tests Failed (llm-interaction / mgr=copaw / wk=copaw)

Commit: 38e4c45
Workflow run: #1503

Test Results

No test output captured.

Debug Log (tail)

No debug logs available.

---

📦 Download full debug logs & test artifacts

[RerankerGuo]

@shiyiyue1102 Hi! Both SHARD_A copaw/copaw and copaw/hermes failed. Same pattern as #984 — in the copaw/hermes case the hiclaw-worker-bob session directory does not exist at all and Manager never invoked hiclaw create worker --name bob — Manager only created alice. The copaw/copaw failure is similar (bob got created but a downstream assertion failed). The failure mode is a test-prompt interpretation flake, not anything in this PR's Manager prompt + worker-management guidance changes.

This PR is docs-only + prompt guidance; it doesn't touch runtime code paths. Could you re-run CI or merge if you agree the flake is unrelated?

[shiyiyue1102]

LGTM. The current main branch does not contain an equivalent diagnostic-loop safeguard. The conflict resolution keeps only the scoped Manager prompt and Worker lifecycle guidance, updates the deletion check to the current agt get workers contract, and removes the temporary #985 test diff. All image builds pass. The fork-only controller shards cannot start because the untrusted pull_request run has no LLM secret; they fail at installer preflight before executing tests, not in this PR path.