Cairn — AI Context Continuity Engine (Codex)

Cairn is a local MCP server keeping decisions + evidence across sessions/models. On Codex there is no statusline and no hooks — so YOU must call the tools on-demand:

• `decision_log` (view:"current") — at the start of a task, to re-inject accepted + open

decisions and the latest brief from prior sessions. This re-injected state is the source of truth on resume — do NOT re-read ADR/CHANGELOG or the whole repo; open only the files named in NEXT STEPS.

• `context_status` — periodically when the conversation grows long; pass the model id and

used tokens. It returns the zone (green/yellow/red) relative to the active window. (Codex caps the window — pass the host-reported model_context_window, e.g. 258400, as surfaceCap.)

• `handoff` — when yellow/red, or before you finish/clear the session, write a **7-bucket

brief** (DECISIONS incl. superseded, EVIDENCE w/ source refs, OPEN QUESTIONS, CONSTRAINTS, VERBATIM byte-exact, NEXT STEPS, DISCARDED) and call handoff (account mode — you write it on the user's ChatGPT/Codex subscription; Cairn freezes verbatim + persists). Pass requiredVerbatim for must-survive values.

• `host_status` — to see which CLIs are logged in.

Trigger phrases (skills under ~/.agents/skills/)

• "Cairn resume" (cairn-resume) — re-inject the ledger, continue, don't re-read the repo.
• "Cairn Handoff" (cairn-handoff) — author + persist a 7-bucket brief.
• "Cairn Help" (cairn-help) — show the command / trigger reference table.

Codex does not support MCP sampling, so server-initiated compaction is unavailable; use the default account mode (you write the brief) or the bridge mode (CAIRN_BRIDGE=claude) to compact via a logged-in Claude CLI.

Principles: fidelity over cost; local-private (no model call in the default mode); when in doubt, a complete verbatim handoff beats a short lossy one.