fix(settings): cover slow SessionEnd hooks with CLAUDE_CODE_SESSIONEND_HOOKS_TIMEOUT_MS

[jacobo-ortiz]

Problem

Claude Code v2.1.74+ caps the entire SessionEnd hook chain at 1.5 seconds by default (Claude Code docs; originally surfaced via anthropics/claude-code#33706). The per-hook timeout field (default 600s) is bounded by this chain-wide budget, not in addition to it.

PAI v5.0.0 ships UpdateCounts.hook.ts in SessionEnd position 4. Its handler (hooks/handlers/UpdateCounts.ts) makes two synchronous Anthropic API calls inside refreshUsageCache:

• usage OAuth API — AbortSignal.timeout(3000) at line 214
• cost_report admin API — AbortSignal.timeout(5000) at line 234 (only when ANTHROPIC_ADMIN_API_KEY is set)

Worst case: ~8 seconds. Measured locally on macOS, 5 runs: 3.05 / 3.72 / 7.42 / 6.46 / 3.74 seconds. UpdateCounts alone blows the 1.5s SessionEnd budget by 2–5×, producing the recurring output:

SessionEnd hook [$HOME/.claude/hooks/UpdateCounts.hook.ts] failed: Hook cancelled

When the chain budget expires mid-execution, subsequent hooks (IntegrityCheck, KVSync) never start — statusline counts go stale and Cloudflare KV sync is skipped.

Why this is structural, not a bug in UpdateCounts

• UpdateCounts.hook.ts and handlers/UpdateCounts.ts are byte-identical to upstream v5.0.0 — confirmed via diff.
• Standalone timing (/usr/bin/time -p × 5) shows 3.05 to 7.42 s — the API calls are the bottleneck regardless of session state.
• Sibling SessionEnd hooks measured for comparison: IntegrityCheck 0.60 s, KVSync 0.91 s.

The fix

Add CLAUDE_CODE_SESSIONEND_HOOKS_TIMEOUT_MS=10000 to the env block in Releases/v5.0.0/.claude/settings.json. 10000 ms gives UpdateCounts worst-case (~8 s) headroom while keeping the chain bounded.

   "env": {
     "PAI_DIR": "${HOME}/.claude/PAI",
     "PROJECTS_DIR": "${HOME}/Projects",
     "BASH_DEFAULT_TIMEOUT_MS": "600000",
     "API_TIMEOUT_MS": "1800000",
+    "CLAUDE_CODE_SESSIONEND_HOOKS_TIMEOUT_MS": "10000",
     "PAI_CONFIG_DIR": "${HOME}/.claude/PAI",
     "GOOGLE_WORKSPACE_CLI_CREDENTIALS_FILE": ...
   },

Verification

• jq '.env.CLAUDE_CODE_SESSIONEND_HOOKS_TIMEOUT_MS' settings.json → "10000"
• jq empty settings.json → JSON valid (no parse errors)
• After a clean session exit, no Hook cancelled line for UpdateCounts / IntegrityCheck / KVSync.

Related

• Closes/addresses SessionEnd hooks need hard timeouts to prevent chain blocking #965 ("SessionEnd hooks need hard timeouts to prevent chain blocking") — the framing is inverted: Claude Code already enforces a hard timeout (1.5 s), and that's exactly what's biting users. The env-var override is the documented fix.
• [DOCS] Hooks docs are outdated for SessionEnd timeout override behavior anthropics/claude-code#33706 — docs issue that named the 1.5 s default; closed RESOLVED 2026-03-18.
• [BUG] SessionEnd hook cancelled with "Request interrupted by user" on Ctrl+C in Version 2.1.72 anthropics/claude-code#32712 / #41577 — closed user reports of the symptom; closed NOT_PLANNED / DUP because the 1.5 s budget is by-design.

Notes for maintainers

• If a more conservative value is preferred (e.g., 5000 ms, matching the Claude Code docs example), I'm happy to adjust. 10000 ms is chosen specifically to cover UpdateCounts worst case with margin; 5000 ms would leave the cost_report API path at risk on slow networks.
• Long-term, moving refreshUsageCache to a detached subprocess (returning the hook in <50 ms) would eliminate the budget pressure entirely, but it's a larger change. The env-var addition is the minimal fix that resolves the symptom for every PAI user today.

[danielmiessler]

Hey @jacobo-ortiz, thanks for raising this, and sorry it sat for a while.

We're changing how LifeOS ships. Instead of cloning a full ~/.claude directory and running it as a complete system, LifeOS is becoming a skill you install through an agentic installer. The installer hands integration to your own AI, which reads your actual machine (your OS, your paths, your harness) and wires the hooks and system prompt in where they belong.

That's aimed right at what you hit here. The old "one directory, one layout, hope it matches your setup" approach is exactly what broke for so many people, and the new model should handle it far better because your AI does the integration per machine instead of us guessing.

So we're closing this in prep for that release. If it still bites you once the skill-based version is out, reopen or file a fresh one and we'll jump on it. Appreciate you taking the time.