fix(haiku): pass prompt on stdin, not argv — saves of long sessions die with E2BIG

[kays0x]

Summary

call_haiku passes the entire prompt as a single command-line argument (claude -p <prompt>). On Linux a single argv string is capped at MAX_ARG_STRLEN (131072 bytes = 128 KB = PAGE_SIZE * 32). When a session's extract reaches that size, subprocess.run fails at exec() with OSError: [Errno 7] Argument list too long, which surfaces as a RuntimeError and the save is lost.

The failure is silent (only a log line) and permanent for that session — the extract only grows, so every subsequent save fails the same way. It strikes exactly the sessions a memory tool most wants to capture: long ones, recovery of a missed session, or any accumulated delta over 128 KB.

Root cause

pipeline/haiku.py:

cmd = ["claude", "-p", prompt, ...]      # prompt is one argv string
subprocess.run(cmd, ...)                 # exec() → E2BIG when len(prompt) > 128KB

Reproduction (no API call needed — it fails at exec)

import subprocess
big = "x" * (128 * 1024)
subprocess.run(["/bin/true", big])       # OSError: [Errno 7] Argument list too long
subprocess.run(["/bin/cat"], input=big, text=True)  # fine — stdin has no argv limit

Threshold is exact: 131071 bytes argv → ok, 131072 bytes (128 KB) → E2BIG.

Fix

claude -p reads the prompt from stdin when no positional prompt is given. Drop the positional and pipe via input=:

cmd = ["claude", "-p", "--output-format", "json", ...]   # no positional prompt
subprocess.run(cmd, input=prompt, ...)                    # prompt on stdin, unbounded

stdin has no size limit, so prompts of any length deliver. No behavior change for normal-size prompts.

Testing

• Adds test_call_haiku_sends_prompt_on_stdin_not_argv — asserts the prompt is passed via input= and is absent from the command argv. It passes on this change and fails against the old argv form (the existing tests were blind to it: they checked cmd[0] and env, never input=), so the regression can't silently return. tests/test_haiku.py now 31/31.
• Full suite: no new failures vs the pre-change baseline.
• Real-environment: a save whose extract was ~1.9 MB failed with call-haiku error: [Errno 7] Argument list too long before the change, and completed normally (Haiku returned a summary, entry written) after it.

Risk

Minimal. subprocess.run(..., input=, capture_output=True) uses communicate() internally, so there is no stdin/stdout deadlock. input is a str consistent with text=True/encoding="utf-8".

Not changed

Scope is the single argv→stdin swap. No change to flags, parsing, env stripping, or token accounting.

[fdaviddpt]

Clean fix — and a nasty bug well caught.

The failure mode is the scary part: extracts only grow, so once a session crosses 128KB every subsequent save dies, silently (just a log line). That hits exactly the long sessions a memory tool most wants to keep.

The fix is correct and minimal — claude -p reads stdin with no positional, subprocess.run(input=...) uses communicate() so no deadlock, and there's no behavior change for normal-size prompts. Bonus: prompt no longer shows up in the process list.

The test earns its keep too — guards both halves (prompt in input=, absent from argv) and fails against the old form, so the regression can't sneak back. Nice catch on the old tests being blind to it.

Thanks Kevin — repro without an API call, exact threshold, real-env validation on a ~1.9MB extract. That's how a bug fix should read. 🙏