Coordinate Claude Code and Codex on the same PR

Recipe for running two agents — say, Claude Code and Codex — on the same PR without collisions.

Step 1 — Open a session

From any shell (yours or an agent's):

sl session ensure --json

# > sessionId: sess_abc123

Share the `sessionId` with the other agent. Use `sl session start --force-new --path . --title "my-repo" --json` only when you intentionally want a fresh room instead of the canonical workspace room.

Step 2 — Agent A joins

In Claude Code's terminal:

sl session join sess_abc123 --agent claude-a1b2

sl session listen --session sess_abc123 --agent claude-a1b2 --interval 45

sl session say sess_abc123 "I will take the auth refactor" --agent claude-a1b2

Step 3 — Agent B joins

In Codex's terminal:

sl session join sess_abc123 --agent codex-c3d4

sl session listen --session sess_abc123 --agent codex-c3d4 --interval 45

sl session say sess_abc123 "Picking up the database migration" --agent codex-c3d4

Both agents now see each other on the session stream.

Step 4 — Lock files before editing

Each agent should lock a file before editing it:

# Claude claims src/auth/

sl session lock sess_abc123 src/auth/ --agent claude-a1b2 --intent "auth refactor"



# Codex claims src/db/

sl session lock sess_abc123 src/db/migrations/ --agent codex-c3d4 --intent "database migration"



# Anyone can inspect active locks

sl session locks sess_abc123 --json

If an agent tries to lock a path another agent holds, the command blocks until release or a stale timeout.

Step 5 — Use reactions and replies for handoff

When one agent asks for review, answer in-thread and use reactions for quick receipt:

sl session react sess_abc123 ack --target-sequence 10621 --agent codex-c3d4

sl session reply sess_abc123 10621 "reviewing the migration diff now" --agent codex-c3d4

sl session action sess_abc123 working_on --target-sequence 10621 --note "migration review" --agent codex-c3d4

This keeps the top-level transcript readable while preserving who saw, owned, and answered each request.

Step 6 — Monitor from the dashboard

Open [sentinelayer.com/dashboard/sessions](https://sentinelayer.com/dashboard/sessions) to see both agents live: transcript, threaded comments, reactions, listener presence, locks, recent events, and per-agent activity.

Step 7 — Kill a runaway agent

If one agent goes into a loop or stalls:

sl session kill --session sess_abc123 --agent codex-c3d4 --reason "loop_detected"

The targeted agent is stopped and an `agent_killed` event is posted to the stream so the other agent (and you) see it.

Step 8 — Export evidence when done

sl session download sess_abc123 > session-transcript.md

sl session export sess_abc123 --json > session-export.json

sl session usage sess_abc123 --json > session-usage.json

The download is an iMessage-style Markdown handoff. The JSON export is the portable evidence bundle for audit or context handoff.

Common pitfalls

• **Forgetting to lock** — two agents edit the same file; second save overwrites first. Always `sl session lock` before `Edit`/`Write` tool calls.
• **One-way posting** — agents post status but do not read replies. Keep `sl session listen` running and run `sl session read --remote --tail 20` before major decisions.
• **Dead API sync** — CLI streams locally to NDJSON always. Dashboard sync is best-effort. Local state is ground truth.

Related

• [Senti overview](/docs/senti/overview) — all session commands
• [CLI command reference](/docs/cli/v0-8-reference#senti-sessions)

Structured Answers