State machine lifecycle bugs: messages swallowed, skills halt, plan mode traps

Bug Category

State machine / lifecycle bugs is the #5 bug category (~8 issues). The agent enters an inconsistent state where expected transitions don't fire or fire incorrectly.

Representative issues

• Messages swallowed after ESC interrupt, agents continue running (#30118) — user's interrupt signal doesn't propagate to background agents; subsequent messages are silently dropped
• --print mode hangs with team agents (#30008) — in_process_teammate agents cause the --print pipeline to hang indefinitely
• Parent skill halts after nested skill completes (#30028) — nested skill completion doesn't resume the parent, remaining steps never execute
• Plan mode breaks CLI session (#29725) — entering plan mode puts the session into a state where messages stop reaching the model
• AskUserQuestion auto-completes without user input (#29889, #29774, #30026) — the consent gate fires without blocking, returning empty answers
• Hook continue:false silently ignored (#29991) — PostToolUse hook sets continue=false via Agent SDK, but the control flow ignores it
• Context compaction re-enters plan mode (#29956) — compaction restores stale state, re-entering a mode the user had moved past

Root cause

These are all state machine transition bugs — the system has multiple states (idle, running, plan-mode, interrupted, skill-executing) and the transitions between them have edge cases:

1. Missing transitions: interrupt signal has no transition to "cancel background agents"
2. Stale state restoration: compaction restores a checkpoint from an earlier state instead of the current one
3. Incorrect guard conditions: AskUserQuestion skips the blocking wait under certain permission modes
4. Broken composition: skill nesting (A invokes B, B completes) doesn't correctly resume A

Proposed mitigations

TLA+ modeling is the right tool here. A SessionStateMachine.tla spec could model:

• States: idle, running, interrupted, plan_mode, skill_executing, compacting
• Transitions: user_input, esc_interrupt, tool_complete, skill_invoke, skill_complete, compact_start, compact_end
• Properties:
• S1: After interrupt, no background agent is running (safety)
• S2: After skill completes, parent resumes (liveness)
• S3: After compaction, session state matches pre-compaction state (safety)
• S4: AskUserQuestion always blocks until user responds (safety)

This is exactly the class of bugs TLA+ excels at finding — non-obvious interleavings of concurrent state transitions.

Related issues

#30118, #30008, #30028, #29725, #29889, #29774, #30026, #29991, #29956