cross-session-message

name	cross-session-message
description	Preview or send a bounded instruction into another known live zellij-backed coding-agent session or Codex Desktop thread outside Symphony.

Cross-Session Message

Use the repo-managed send-zellij-message helper when the user wants to nudge another live zellij-backed task/session with a bounded instruction. When running inside Codex Desktop and the target is a Codex Desktop thread, use the app thread tools as the Desktop transport.

For UI-only work routed to the Mac in-app Browser or native Computer Use, use mac-ui-worker instead of treating a generic cross-session message as a UI-job protocol.

Within an accepted Mac UI job, do not use this skill to relay routine clarification, login, MFA, credential entry, approval, or handoff between the worker and source. Those checkpoints stay between Sourya and the worker by default. The only approval exception is an exact, visibly pre-declared, single-use source delegation. Cross-task sends must match a callback, dependency result, correction, cancellation, or terminal receipt allowed by the complete initial mac-ui-worker/v1 envelope.

Invocation

$cross-session-message <task slug|tracker ref|session name>
/cross-session-message <task slug|tracker ref|session name>

Command

send-zellij-message --project-root "<project-root>" --text "<message>" <target>

Useful flags:

• --execute
• --submit enter
• --tab-name <name>
• --replace-claude-suggestion
• --json

Codex Desktop thread path, when codex_app tools are available:

0. If the thread tools are not already visible, use the runtime's tool discovery to load list_threads, read_thread, and send_message_to_thread.
1. Use list_threads with the tracker id, task slug, or visible title.
2. If the result is ambiguous, stop and surface the candidates.
3. Use read_thread if you need to confirm the current target state.
4. Send with send_message_to_thread, passing threadId and hostId.
5. Omit model and thinking unless the user explicitly asks to override the destination thread's settings.

Behavior

• Run the preview first. The helper is dry-run by default.
• Add --execute only after the resolved session/tab target looks correct.
• Use --submit enter only when you intentionally want the target session to receive Enter after the text write.
• The preview resolves one supported Claude Code or Codex pane and inspects its composer without mutation. An unsafe, non-empty, busy, ambiguous, or unsupported composer blocks execution.
• Claude Code autocomplete is a narrow exception only with explicit --replace-claude-suggestion. The helper requires visible suggestion text while the cursor remains at the empty-prompt position. A flagged preview reports preview_probe_required; execution sends a non-text end-of-line cursor probe. A real buffered draft moves the cursor, is restored without changing text, and fails closed. Only an empty buffer behind autocomplete may proceed.
• Execution writes only to the resolved pane id. It verifies staging before any Enter, and reports delivered: true only after the staged text is observed as a new agent turn.
• Receipts distinguish preview_safe, preview_unsafe, unsafe_composer, staged, delivered, and unverified. executed: true means a mutation was attempted; it does not mean delivery. Gate follow-through on outcome: delivered and delivered: true.
• Long or multiline Codex prompts receive a delayed confirm Enter only when the intended payload is still visibly staged after the first Enter.
• Stop on ambiguous tab selection, cross-machine targets, or missing tracked session data unless the user is intentionally doing local debug work with an explicit --zellij-session.
• If a non-interactive shell lacks XDG_RUNTIME_DIR, the helper keeps an active default zellij namespace. When that namespace has only a missing or exited target and /run/user/<uid> has the exact live session, it selects the live systemd namespace and records that choice in the receipt. It never switches away from an active default target.
• Explicit-session debug sends may identify a uniquely supported wrapped Claude or Codex UI from a read-only screen fingerprint when process-command metadata is unavailable. Ambiguous or multiple matches fail closed.
• For Codex Desktop threads, the preview step is target resolution and, when useful, read_thread; send_message_to_thread is the actual send.

Manual fallback

If the helper reports an unsafe or unverified composer, stop automated sends. Attach to the resolved session, inspect the exact agent pane, and let the human decide how to handle any existing draft, selector, or unknown UI state. Do not clear it, append a retry, or send Enter through the helper. After the composer is manually returned to a visibly empty supported state, start again with a fresh preview.

If the human confirms that Claude's visible text is autocomplete rather than a real draft, repeat the preview with --replace-claude-suggestion. Require preview_probe_required, then execute with the same flag. Execution may stage text only after suggestion_probe: verified_empty_buffer and composer_preflight: safe_claude_suggestion_override. Keep the message short enough for the whole staged payload to remain visible and verifiable. If staging is unverified, do not send Enter.

See ~/pro/botfiles/docs/cross-session-orchestration-contract.md for the shared contract.