Bound terminal output flow across Electron and xterm

[luckeyfaraday]

Summary

Two unbounded-memory paths could drive the Electron main process to V8 heap OOM. This branch bounds both.

Terminal output flow (aafdaac, a5f9b64)

• cap PTY-host and Electron pending terminal output backlogs
• add per-terminal sequence acknowledgements so Electron sends at most one unacknowledged renderer batch
• serialize xterm writes and retain only a bounded 64k pending tail
• bound terminal SSE/backfill paths and clear output state when terminals exit
• ignore local Grok agent configuration via .grok/

Agent session transcript scans (ee83d09)

• add readFilePrefix (512 KB cap via fs.open/read) and replace whole-file readFile(…, "utf8") reads in the codex, claude, and grok session scanners (plus the terminal-restore path)
• add mapWithConcurrency (cap 8) in place of unbounded Promise.all over session files
• memoize the workspace-independent ~/.codex/sessions metadata scan in a 30s cache so it runs once instead of once per workspace

Root cause (transcript scans)

PR #154 (b4c801c, "Improve multi-workspace review handoffs", first shipped in v0.1.9) replaced the single active-workspace session scan with an all-workspaces fan-out:

for (const reviewWorkspace of workspaces) void refreshWorkspaceAgentSessions(reviewWorkspace);

With N saved workspace tabs this issues N concurrent listAgentSessions calls. Each one re-scanned the global ~/.codex/sessions corpus and materialized every file via an unbounded Promise.all(files.map(readFile(…, "utf8"))), so N workspaces produced N concurrent full copies of the same transcript corpus. The per-workspace 30s session cache does not help (it is keyed by workspace), and the utf8 decode path matches the crash stack StringDecoder::DecodeData → NewStringFromUtf8 → NewRawTwoByteString.

Evidence

Terminal output flow

The Electron main process hit V8 heap OOM at 12:45:12, 14:48:07, and 15:15:31 CEST. The 15:15 recurrence ran the initial backlog-cap build, showing retained strings were bounded but Electron IPC and xterm write queues were still unbounded.

Transcript scans (identical corpus, 2 GB heap, v0.1.9)

Scenario	Result	Peak RSS
1 workspace (pre-#154 behavior)	completed	1.40 GB
11 workspaces (#154 behavior)	V8 heap OOM	2.88 GB
11 workspaces (bounded reader)	completed	0.94 GB

Verification

• npm run build:electron
• npm run test:electron — 135 tests across 20 files pass (incl. terminal-buffer, file-prefix)
• npm run build
• packaged client/release/ATHENA-0.1.9.AppImage

Follow-up

The Hermes provider scan had the same global-corpus-rescanned-per-workspace shape; it is deduplicated separately in #170.