Memory

2026-03-08

• codex exec --json emits machine-readable JSONL that includes thread.started.thread_id, item.completed entries for agent messages, and turn.completed.usage.
• codex exec resume --json <thread_id> <prompt> can continue the same Codex session non-interactively.
• When integrating with CLI JSONL output, keep a line buffer across stdout chunks; splitting each data chunk independently can corrupt partial JSON lines.

2026-03-09

• In modern Node, http.Agent / https.Agent can be created with proxyEnv so libraries that use core http(s) without an explicit proxy agent can inherit proxy behavior via the global agents.
• discord.js on Node uses undici for REST and ws for the gateway, so full proxy support may require configuring both an undici dispatcher and core http(s) agents instead of relying on only one transport hook.
• codex exec --json failures do not always put the real user-facing error on stderr; sometimes stderr only has an internal warning while the actionable error is printed on stdout, so wrappers should inspect both.
• Claude Code CLI stream-json integrations should keep a line buffer across stdout chunks and inspect both structured result events and stderr, because session ids, final text, and surfaced errors may arrive in different events.
• Claude Code --output-format stream-json in --print mode also requires --verbose; without it, the CLI can fail before the actual model/backend error is reached.
• Claude stores per-project local session logs under ~/.claude/projects/<absolute-path-with-slashes-replaced-by-dashes>/, and assistant events there can include both usage metadata and human-readable limit text such as You've hit your limit ....
• Claude interactive slash commands can be scripted over stdin; for example, printf '/usage\n/exit\n' | claude returns the same human-readable usage/limit text shown in the interactive CLI.
• If real behavior or the correct data source is still uncertain after a few probes, stop self-looping, ask the human for help/context, and avoid speculative production-code changes first.

2026-03-11

• codex CLI marks --ask-for-approval on-failure as deprecated; for non-interactive codex exec wrappers, prefer --ask-for-approval never and pair it with an explicit sandbox mode instead of relying on older approval behavior.

2026-03-12

• Feishu long-connection event handlers still need to finish within 3 seconds; if the real work is a long-running CLI turn, acknowledge the event immediately and continue the job asynchronously, otherwise Feishu can retry the same event.
• Feishu long-connection delivery is cluster-mode, not broadcast: if the same app is connected from multiple runtimes, only one client receives a given event.

2026-03-13

• When an integration starts using a new third-party endpoint or platform capability, treat the required permissions/scopes as part of the feature itself and update setup docs in the same change instead of letting runtime permission errors surface the requirement later.