Phase 04: Hook Lifecycle — Stop Hook Loop and Output Isolation

The Stop hook creates an infinite loop because its summarize output gets interpreted as new instructions by Claude Code (#987). The fix is ensuring Stop/Summary hooks return suppressOutput: true and never emit stdout. Secondary: stderr from hooks shows as error UI in Claude Code (#1181). These are specific, targeted fixes — not a rate limiter framework.

Issues resolved: #987, #984, #975 (Stop hook loop), #1181 (stderr as errors), #598 (conversation pollution), #784 (agent output leaks)

Root Cause Validation

Stop hook loop — The Stop hook generates summarization output. Claude Code interprets any stdout from a hook as instructions, creating a feedback loop. The fix is suppressOutput: true in the hook response, not a rate limiter.

stderr as UI errors — Claude Code displays stderr content to users as error messages. The logger or hook code writes to stderr for diagnostic purposes, which users see as errors. Fix: suppress stderr in hook context.

Tasks

• Fix the Stop hook infinite loop:

  • Find the Stop/SessionEnd handler — search for session-complete in src/cli/handlers/
  • Ensure the handler returns { continue: true, suppressOutput: true } — no stdout content
  • Find the hook response builder (src/cli/adapters/claude-code.ts or src/hooks/hook-response.ts) and ensure Stop/Summary hook types always set suppressOutput: true
  • Fix the "Unknown event type: session-complete" error (#984) — find the event type dispatcher and add session-complete as a recognized type
  • This is a property set, not a new system
  • DONE: All already implemented — session-complete handler exists with suppressOutput: true, adapter defaults to suppressOutput: true, unknown event types return no-op with exit 0.

• Fix stderr showing as errors in Claude Code UI (#1181):

  • Check src/utils/logger.ts — if it writes to stderr, add a check: when running in hook context (detect via env var or process context), suppress stderr output
  • The simplest fix: in hook entry points (the built scripts), redirect stderr to a log file before any imports: process.stderr.write = () => true or pipe to the log file
  • Do NOT create a "hook-mode flag" system — just suppress stderr at the entry point
  • DONE: Added process.stderr.write = (() => true) at start of hookCommand() with finally block restore. Converted console.error() to logger.warn()/logger.error() in hook-command.ts and handlers/index.ts.

• Fix conversation history pollution (#598, #784):

  • This is the same root cause as the Stop hook loop: hook output leaking to the conversation
  • Audit all hook handlers in src/cli/handlers/ — every handler that returns output should set suppressOutput: true unless it's specifically injecting context (like SessionStart context injection)
  • For the --continue case (#784), the memory agent's internal processing output should never be visible — ensure the summarization flow uses suppressOutput: true
  • DONE: Audited all 7 handlers — all return suppressOutput: true. Adapter defaults to suppressOutput: true. Context handler uses hookSpecificOutput (correct for context injection). stderr suppression from task 2 covers the remaining conversation pollution vector.

• Run npm test and fix any failures

  • DONE: 10 new tests added in tests/hook-lifecycle.test.ts, all passing. 52 total hook-related tests pass. Full suite: 954 pass, 21 fail (all pre-existing baseline failures, none from TRIAGE-04 changes).