[hassiebp]

Fixes #119 (Linear LFE-17701, langfuse/langfuse#18995).

Problem

Every generation's input was built from the entire history of all earlier turns, and each input is serialized onto its own span. Memory therefore grew with steps × session size. On long sessions, the Stop hook crashed with a V8 heap OOM. That crash is a hard abort() that skips the fail-open try/catch, and it happens before markTurnUploaded, so every later Stop hook had an even bigger backlog to process and crashed too.

Fix

• New option max_history_chars (default 200000), set in langfuse.json or with LANGFUSE_CODEX_MAX_HISTORY_CHARS. For each turn, the history from earlier turns is built once. It walks back from the most recent message until the budget is used, and replaces everything older with a single [N earlier message(s) omitted …] system marker.
• Generations whose history was cut get codex.history_omitted_messages metadata, so the loss of detail is visible in the trace.
• The current turn's own user message and earlier steps are never dropped.
• The old code eagerly copied the full history prefix for every turn ([...seen]). That is replaced by per-turn message arrays and a bounded history computed lazily, only for turns that are actually emitted.
• Invalid values (negative numbers, non-numbers) are ignored and the default is used. The rest of the config file is still kept.
• The new option is documented in the README.

Verification

• pnpm test: 125 tests pass, including new ones for the truncation and the marker/metadata and for config parsing. pnpm lint passes.
• Offline repro with the endpoint unreachable, on a synthetic 21 MB rollout (24 turns × 15 steps × 60 KB tool outputs, all turns pending):

Build	Heap limit	Result
main	4096 MB	completes, peak RSS 3.4 GB, 58 s
main	1024 MB	SIGABRT, heap OOM after 5 s
this PR	4096 MB	completes, peak RSS 746 MB, 7.6 s
this PR	1024 MB	completes, peak RSS 744 MB

Not in this PR

Some follow-ups from the issue are left out: flushing between pending turns, and capping individual tool outputs. Memory within a single very long turn still grows with that turn's own steps.

Linear Issue: LFE-17701