Module codex 

Codex CLI: native command hooks in ~/.codex/hooks.json ($CODEX_HOME), each running <borg> hook codex <event>. The file has Claude Code’s shape, and so do the events and the answers Codex reads back (hookSpecificOutput.additionalContext, {"decision":"block","reason"}), except the edit tool: apply_patch sends a patch, which Codex::normalise turns into a canonical Edit.

Codex runs a new or changed hook only once the engineer trusted it (/hooks in codex), and skips an untrusted one without a word: borg install says so, and borg doctor asks Codex itself which of borg’s hooks are not trusted (codex app-server, hooks/list). borg never writes that trust.

Structs

Codex

Constants

CONTEXT_CHARS

Protocol (Codex’s default additionalContextLimit of a hook): an additionalContext up to this many characters reaches the model whole as a developer message; a longer one is cut to its head and tail with a “truncated output” warning, the whole text saved under hook_outputs. Measured on Codex 0.153.2 against a local mock of the Responses API (2026-10-28): 10,000 characters arrived whole, 10,010 cut.

EVENTS

Codex hook events and the canonical borg events they become. Codex has no failure event (a failed tool call fires no PostToolUse at all), so denials are not captured. Interrupt (an Esc) ends the turn of a session that listens for wakes (listen.rs), which a turn with no stop and a tool call that never ended would hold off for at least listen::TOOL_MS.

RETIRED

Hooks an older borg installed and this one does not.

TIMEOUTS

Guards (protocol once trusted): seconds Codex gives each hook, written explicitly (its own default is 600 s, and 1 s for SessionEnd, which it clamps to 3 s), so a hung borg holds Codex’s turn no longer than this at each event. The stop’s (STOP_TIMEOUT_S) is also its budget: the stop’s model call and the expert drone keep inside it. The trust Codex keeps covers the timeout too: changing one here makes every engineer trust borg’s hooks again.

Functions

patch_edit

One apply_patch as the canonical edit: the first file the patch names (*** Add|Update|Delete File:) as file_path with its added lines as new_string, each further file in edits the same way. A rename (*** Move to: after an Update) names its destination as a further file, which gets the hunk’s added lines. Headers are matched on the trimmed line, as Codex does. Relative paths are taken from cwd. None for a patch naming no file.

rollout_of

Thread id’s rollout under sessions: a file named after it whose session_meta names it.

sessions

Where Codex keeps every thread’s rollout, <sessions>/YYYY/MM/DD/rollout-<time>-<thread>.jsonl: $CODEX_HOME/sessions, else home/.codex/sessions.

spawn_runs

One run per session whose agent spawned others (spawn_agent), oldest first: of session only when given, from sessions that started at or after since. Every thread keeps its own rollout (<sessions>/YYYY/MM/DD/rollout-*.jsonl), whose first line, session_meta, names a spawned thread’s parent (source.subagent.thread_spawn.parent_thread_id, with its agent_path, nickname and role; checked against the serde names in codex 0.153.2). An agent spawned by a spawned agent belongs to its root session’s run, which is Run::loose. Codex has no structured output, so findings come from an agent’s text. This reader is built from codex’s serde names and hand-made rollouts, not checked against a real spawned one yet.

trust_findings

borg doctor’s findings from a hooks/list answer: borg’s hooks (hooks recognises their commands) that Codex would skip (untrusted, changed since trusted, disabled), and Codex’s load warnings.