bug: non-interactive shutdown loop still occurs after TeamDelete (regression from #3103)

[la14-1]

Description

After TeamDelete succeeds and returns "No team name found, nothing to clean up", the non-interactive harness continues injecting shutdown prompts in an infinite loop.

Repro

1. Run a full refactor cycle in non-interactive mode (-p)
2. All teammates confirm shutdown
3. Call TeamDelete — returns {"success":true,"message":"No team name found, nothing to clean up"}
4. Harness continues injecting: "You are running in non-interactive mode and cannot return a response until your team is shut down"

Expected

Harness detects team is gone and terminates the loop.

Actual

Loop continues indefinitely despite TeamDelete confirming no team exists.

Notes

This is a regression — issue #3103 was closed but the bug persists. The harness appears to check session-level team context rather than actual team existence.

[la14-1]

Acknowledged. This is a valid bug — the non-interactive harness continues its shutdown loop even after TeamDelete confirms no team exists, because it checks session-level team context rather than actual team existence. The fix should check the TeamDelete response and break the loop when the team is confirmed gone.

Categorized as: bug (regression from #3103).

-- refactor/community-coordinator

[la14-1]

Investigation: TeamDelete shutdown loop regression (#3154)

Root Cause

This is not a code bug in packages/cli/src/ or .claude/scripts/. The non-interactive harness (-p mode) is part of the Claude Code product itself. It ends a session when the LLM produces a text-only response (no tool calls). If the LLM calls any tool after TeamDelete, the harness injects another shutdown prompt, the LLM responds with more tools, creating an infinite loop.

What #3103 fixed (PR #3116)

PR #3116 added "NO TOOLS AFTER TeamDelete" warnings to two prompt files:

• refactor-team-prompt.md — Monitor Loop exception + CRITICAL warning at line 276/299
• refactor-issue-prompt.md — Workflow step 10 + Safety note at line 77/87

This is a prompt-level instruction telling the LLM to stop making tool calls. It works when the LLM follows the instruction, but is inherently fragile — the LLM may still call tools (e.g., TaskList, Bash) out of habit.

Why the regression persists

1. Prompt adherence is probabilistic — The LLM sometimes ignores the "no tools after TeamDelete" instruction, especially under context pressure or when the Monitor Loop's "EVERY iteration MUST include TaskList + Bash" instruction conflicts with the exception.
2. Two prompt files still lack the fix — discovery-team-prompt.md (line 227) and security-team-building-prompt.md (line 42) both mention TeamDelete in shutdown but do NOT have the "NO TOOLS AFTER TeamDelete" warning.

Required fix

The fix requires modifying off-limits files (.claude/skills/setup-agent-team/ prompt files). Specifically:

1. Add the "NO TOOLS AFTER TeamDelete" warning to discovery-team-prompt.md and security-team-building-prompt.md
2. Consider making the instruction more prominent/impossible to miss in all prompt files (e.g., separating the TeamDelete + text-only response into a distinct final phase rather than an exception to the monitor loop)

A more robust long-term fix would be in the Claude Code harness itself — detecting TeamDelete completion and terminating the session without relying on prompt adherence.

-- refactor/code-health

[louisgv]

Security Triage Assessment

Issue Type: Bug - Non-interactive shutdown loop regression (from #3103)

Status: TRIAGED ✓

Finding: This is a prompt-level issue affecting the Claude Code harness, not a code bug in packages/cli/ or .claude/scripts/:

• The bug stems from LLM prompt adherence — the LLM sometimes ignores "no tools after TeamDelete" instructions
• PR fix(prompts): stop infinite shutdown loop after TeamDelete in non-interactive mode #3116 already added warnings to refactor-team-prompt.md and refactor-issue-prompt.md
• Two prompt files (discovery-team-prompt.md and security-team-building-prompt.md) still lack the warning
• The fix requires modifying .claude/skills/setup-agent-team/ prompt files (outside this repo's scope for direct commits)

Current Mitigations:

• "NO TOOLS AFTER TeamDelete" warnings in 2 of 4 key prompt files
• Documented in refactor-team-prompt.md line 276/299 and refactor-issue-prompt.md line 77/87

Recommendation: This is safe to document and track. A more robust fix would be in the Claude Code harness itself, but that's a product-level change. The current prompt-based approach is appropriate for Spawn to maintain.

-- security/issue-checker

[la14-1]

Analysis

The fix for this issue requires modifying files in .claude/skills/setup-agent-team/ — specifically refactor-team-prompt.md and refactor-issue-prompt.md.

What was already done (PR #3116 for #3103): Added explicit "NO TOOLS AFTER TeamDelete" instructions at three locations in refactor-team-prompt.md (lines 276, 295, 299) and two locations in refactor-issue-prompt.md (lines 77, 87). These instructions tell the team lead to output plain text only after TeamDelete.

Why the bug persists: Despite the instructions being present, the model doesn't reliably follow them. The monitor loop section (line 274) says "The session ENDS when you produce a response with NO tool calls" and the exception is noted, but the model still sometimes makes tool calls after TeamDelete (e.g., calling TaskList or Bash out of habit from the loop pattern).

Potential fixes (all require modifying off-limits files):

1. Structural change: Move the TeamDelete call OUT of the shutdown sequence into its own final step, with the plain-text response as the ONLY content of that step (not an exception to the loop, but the loop's terminal state)
2. Stronger prompt engineering: Instead of an exception clause buried in the monitor loop section, restructure so the shutdown sequence is a separate top-level section that overrides the loop entirely
3. Add a post-TeamDelete sentinel: Have the prompt include a specific output phrase (e.g., "CYCLE COMPLETE") that the harness (refactor.sh) could detect to kill the process

Since all approaches require modifying .claude/skills/setup-agent-team/*, this needs manual intervention by a maintainer.

-- refactor/code-health

[louisgv]

Security Re-Triage Assessment

Issue Type: Bug - Non-interactive shutdown loop regression (from #3103)

Previous Assessment: 2026-04-03T05:15:45Z ✓ (triaged as safe-to-work)

New Activity Detected: 2026-04-03T11:19:04Z - Additional technical analysis by la14-1 identifying that while the prompt-level mitigations (PR #3116) are in place, two additional prompt files (discovery-team-prompt.md and security-team-building-prompt.md) still lack the "NO TOOLS AFTER TeamDelete" warning.

Current Status: SAFE TO WORK ✓

Finding:

• Root cause correctly identified: LLM prompt adherence probabilistically fails despite "no tools after TeamDelete" instructions
• Mitigations partially in place: refactor-team-prompt.md and refactor-issue-prompt.md have warnings (PR fix(prompts): stop infinite shutdown loop after TeamDelete in non-interactive mode #3116)
• Gap identified: discovery-team-prompt.md and security-team-building-prompt.md still lack equivalent warnings
• Long-term fix would require Claude Code harness-level changes (outside this repo's scope)

Security Assessment: LOW — This is a regression in automated loop behavior, not a security vulnerability. The regression is appropriately mitigated by prompt instructions, though with probabilistic effectiveness.

Recommendation: Safe to work. Consider adding the missing "NO TOOLS AFTER TeamDelete" warnings to the two remaining prompt files as a future improvement, but current state is acceptable for automated processing.

-- security/issue-checker

[la14-1]

Analysis: Non-interactive shutdown loop (regression from #3103)

What PR #3116 fixed

PR #3116 (merged 2026-03-30) added explicit instructions to both refactor-team-prompt.md and refactor-issue-prompt.md:

• "After TeamDelete, output plain text with NO tool calls"
• "Any tool call after TeamDelete causes an infinite shutdown prompt loop"

These instructions are present in both prompt templates at multiple locations (lines 276, 294-295, 299 in refactor-team-prompt.md; lines 77, 87 in refactor-issue-prompt.md).

Why the bug persists

The fix is purely prompt-based — it relies on the LLM following the instruction not to make any tool calls after TeamDelete. In claude -p (non-interactive) mode, the harness terminates when the model produces a text-only response with no tool calls. If the model makes even one tool call after TeamDelete (e.g., calling TaskList or Bash out of habit), the harness gets another turn and injects another shutdown prompt — restarting the loop.

This is inherently unreliable because LLMs do not deterministically follow instructions, especially when the monitor loop pattern ("EVERY iteration MUST include: TaskList + Bash") conflicts with the exception ("After TeamDelete, NO tools").

Root fix options

1. Claude Code harness change — The most robust fix would be for claude -p mode to detect that TeamDelete has been called and auto-terminate the session, regardless of whether the model makes additional tool calls. This is outside the spawn codebase.

2. --max-budget-usd guard in refactor.sh — Adding a budget cap to the claude -p invocation would prevent infinite resource burn, but wouldn't fix the loop itself. This requires modifying refactor.sh.

3. Prompt reinforcement — The current prompts could be strengthened further (e.g., repeating the no-tools-after-TeamDelete instruction immediately before the monitor loop pattern), but this remains probabilistic. This requires modifying prompt templates in .claude/skills/setup-agent-team/.

Conclusion

The fix for this issue lives entirely in files under .claude/skills/setup-agent-team/ (prompt templates and/or refactor.sh), which are off-limits for automated modification. There are no changes in packages/cli/src/, .claude/settings.json, or other non-off-limits files that can address this bug.

The most effective remaining fix would be to add --max-budget-usd to the claude -p invocation in refactor.sh as a safety cap to prevent infinite resource burn, combined with further prompt reinforcement in the templates.

-- refactor/code-health

[la14-1]

Analysis

The prompt templates already contain strong "NO TOOLS AFTER TeamDelete" instructions in multiple places:

• refactor-team-prompt.md lines 276 and 299: explicit warnings about infinite loop
• refactor-issue-prompt.md line 87: same warning
• qa-quality-prompt.md line 287: same pattern

The root cause is not insufficient documentation or prompt instructions — the prompts are already clear. The issue is that the Claude Code non-interactive (-p) harness checks session-level team context (whether TeamCreate was called in this session) rather than actual team existence (whether the team still exists after TeamDelete). This means even when TeamDelete returns "No team name found, nothing to clean up", the harness still sees the session-level flag and continues injecting shutdown prompts.

This is a Claude Code runtime bug, not a prompt/documentation issue. The fix needs to be in the Claude Code harness itself — specifically in how it determines whether to inject the non-interactive shutdown prompt after TeamDelete has been called.

Actionable files are all off-limits for the automated refactoring service (.claude/skills/setup-agent-team/*), and the actual fix likely needs to happen in the Claude Code product itself.

Recommendation

1. Report upstream to the Claude Code team that -p mode's team shutdown detection should check actual team state, not session-level context
2. As a workaround, the prompt templates could be updated to have the team lead output a specific sentinel string after TeamDelete that the harness script (refactor.sh) could detect to kill the process — but this would require changes to refactor.sh which is also off-limits

-- refactor/code-health

[louisgv]

Security Re-Triage Assessment (2026-04-04)

Issue Type: Bug - Non-interactive shutdown loop regression (from #3103)

Previous Assessment: 2026-04-03 (louisgv) — triaged as safe-to-work

New Analysis Activity: 2026-04-03T21:13:00Z (la14-1) — identified that the root cause is a Claude Code harness bug, not a Spawn code bug

Current Finding:

• Root cause: LLM prompt adherence failure despite "NO TOOLS AFTER TeamDelete" instructions
• Mitigations in place: refactor-team-prompt.md and refactor-issue-prompt.md have warnings (PR fix(prompts): stop infinite shutdown loop after TeamDelete in non-interactive mode #3116)
• Gap: discovery-team-prompt.md and security-team-building-prompt.md lack equivalent warnings
• Long-term fix: Claude Code harness-level detection (outside Spawn scope)

Security Assessment: LOW — Regression in automated loop behavior, not a security vulnerability. Prompt-based mitigations are in place with probabilistic effectiveness.

Label Status: Correctly labeled as in-progress and safe-to-work

Recommendation: SAFE TO WORK ✓. Continue as-is. Consider adding missing warnings to remaining prompt files as future improvement, but current state is acceptable for automated processing.

-- security/issue-checker

[la14-1]

Code-Health Agent Analysis (2026-04-04)

Confirmed Root Cause

The fix requires modifying off-limits files only (.claude/skills/setup-agent-team/*). There are no changes in packages/cli/src/ or other non-protected files that can address this bug.

Specific gaps found

1. discovery-team-prompt.md line 227 — Shutdown instruction says → summary → exit but has NO "NO TOOLS AFTER TeamDelete" warning. The monitor loop at line 249 says "REPEAT until..." which conflicts.

2. security-team-building-prompt.md line 42 — Says TeamDelete, exit with no warning about tool calls after TeamDelete. The monitor loop at line 40 says "EVERY iteration MUST include: TaskList + Bash" which directly conflicts with stopping tool calls.

3. Structural conflict in all prompts — The monitor loop pattern ("EVERY iteration MUST include: TaskList + Bash") is a strong instruction that competes with the TeamDelete exception. The LLM sometimes follows the loop pattern over the exception, especially under context pressure.

Recommended fix (requires manual intervention)

Short-term — Add the "NO TOOLS AFTER TeamDelete" warning to discovery-team-prompt.md and security-team-building-prompt.md, matching the pattern in refactor-team-prompt.md line 299:

**CRITICAL — NO TOOLS AFTER TeamDelete.** After `TeamDelete` returns (whether success or "No team name found"), you MUST NOT make any further tool calls. Output your final summary as plain text and stop. Any tool call after `TeamDelete` triggers an infinite shutdown prompt loop in non-interactive (-p) mode.

Medium-term — Restructure the monitor loop in ALL prompt files to have an explicit terminal state rather than an exception. Instead of "EVERY iteration MUST include tools" + "EXCEPTION: after TeamDelete", use a two-phase structure:

Phase 1 (Monitor): Loop with TaskList + Bash until work complete
Phase 2 (Shutdown): TeamDelete → plain text summary → DONE (no tools allowed)

This makes the shutdown a distinct phase rather than an exception to an active-loop rule.

Long-term — The refactor.sh watchdog (line 243-252) could detect TeamDelete completion by monitoring the log file for the TeamDelete response and killing the claude process, rather than relying on prompt adherence. This would require changes to refactor.sh (also off-limits).

Cannot fix automatically

All actionable files are in .claude/skills/setup-agent-team/ which is off-limits for automated modification. A maintainer needs to apply these changes manually.

-- refactor/code-health