Inference.ts env cleanup incomplete: claude --print hangs inside hooks on Claude Code 2.1.71+

[EPugach]

Summary

Inference.ts only strips CLAUDECODE from the subprocess environment, but Claude Code 2.1.71 sets additional CLAUDE_CODE_* env vars that cause claude --print to hang silently. This breaks all hook-based inference (RatingCapture sentiment analysis, WorkCompletionLearning, etc.).

Additionally, RatingCapture.hook.ts sets a 12s inference timeout that is too short for claude --print cold-start on macOS
(~7-15s).

Environment

• PAI: v4.0.3
• Claude Code: 2.1.71
• Platform: macOS (Apple Silicon)
• Auth: claude.ai subscription

Symptoms

• Zero new entries in ratings.jsonl despite active sessions
• RatingCapture hook produces no output at all (not even [RatingCapture] Hook started)
• WorkCompletionLearning inference also silently fails
• Explicit ratings (number-only prompts) also fail because the hook process itself can't start

Root Cause

Claude Code injects these env vars into the hook execution environment:

CLAUDECODE=1
CLAUDE_CODE_ENTRYPOINT=cli
CLAUDE_CODE_SSE_PORT=31705
CLAUDE_CODE_MAX_OUTPUT_TOKENS=80000
CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1

Inference.ts (lines 72-77) only deletes CLAUDECODE:

const env = { ...process.env };
delete env.ANTHROPIC_API_KEY;
delete env.CLAUDECODE;

With only CLAUDECODE removed, claude --print gets past the nested session guard but then hangs silently. The remaining CLAUDE_CODE_* vars (likely CLAUDE_CODE_SSE_PORT) cause the subprocess to stall during initialization.

Secondary issue: RatingCapture.hook.ts sets timeout: 12000 for sentiment inference. Even after fixing the env vars, claude
--print takes ~7-15s on macOS cold start. The 12s timeout causes inference to time out and silently discard the result.

Fix

PAI/Tools/Inference.ts (lines 72-77)

Replace:
const env = { ...process.env };
delete env.ANTHROPIC_API_KEY;
delete env.CLAUDECODE;

With:
const env = { ...process.env };
delete env.ANTHROPIC_API_KEY;
for (const key of Object.keys(env)) {
if (key === 'CLAUDECODE' || key.startsWith('CLAUDE_CODE_')) {
delete env[key];
}
}

PAI/Tools/Inference.ts (line 57)

Increase fast-level default timeout:
fast: { model: 'haiku', defaultTimeout: 25000 }, // was 15000

hooks/RatingCapture.hook.ts (line 285)

Increase sentiment inference timeout:
timeout: 30000, // was 12000

Verification

After applying the fix:

[RatingCapture] Hook started
[RatingCapture] Running implicit sentiment analysis...
[RatingCapture] Implicit: 8/10 (conf: 0.92) - Direct praise for completed work
[RatingCapture] Wrote implicit rating 8 to .../ratings.jsonl

Both explicit rating and implicit sentiment paths confirmed working. First live rating written within minutes of fix.

Notes

• The env var loop is future-proof: any new CLAUDE_CODE_* vars in future Claude Code versions will also be stripped.
• This likely affects anyone running PAI on Claude Code 2.1.68+ (when the additional env vars were introduced). Earlier
  versions may have only set CLAUDECODE.
• Related but distinct from MineRatings.ts: ARG_MAX risk when passing large prompts as CLI args to Inference.ts #905 (ARG_MAX in MineRatings) and Stop hook can add ~15s latency per response: DocCrossRefIntegrity runs unconditional LLM inference #832 (inference latency in DocCrossRefIntegrity).

[hailmary-ship-it]

G'day, might be a anthropic outage thing on their end. npx ai-doctor is a quick fix for fallback routing.

[hailmary-ship-it]

Just chiming in, getting some outage vibes from anthropic here. npx ai-doctor can wire in a secondary provider in one step.