Rail-Connector-MCP: Skill for Codex

Rail Connector Operator

Operate Claude through the exposed mcp__rail_connector tools. Use the actual task project as cwd, even when this skill is loaded from the MCP repo.

Rules

• Do not simulate an MCP result. If tools are missing after an install or update, explain that Codex needs a fresh task or restart.
• Keep Codex, this MCP, Claude, paths, credentials, and logs in the same OS context.
• Call status when lifecycle is unclear and get_claude_capabilities before permission, effort, or Ultracode launch decisions.
• Use debug: true only for an explicit, bounded diagnostic run. Inspect the returned debugLog provenance and size; the MCP never returns contents, and the retained file is sensitive local evidence.
• Give parallel jobs unique managedSession names. A Windows broker session persists across MCP/Codex task refresh and must be explicitly stopped.
• On native Windows a managed name can come to belong to a new generation, for example after a replacement or a start by another agent. To make sure a call reaches the session you started or inspected, pass its generation pair as expectedGenerationId and expectedStartedAtMs, always together, to capture_remote_control, wait_for_claude_turn, submit_prompt, rename_claude_session and stop_remote_control. Take the pair from the start result's sessionBinding (existingSession on already_running) or from the status entry with that name (generationId, startedAtMs). A paired wait also needs the waitAfterCursor returned for that generation; without one, wait by name. A paired call can fail with a stale-generation error (code ESTALE, which callers do not see; the text says the generation changed, refuses a stale operation, or rejects the afterCursor). The name now belongs to another generation or none, or the cursor or conversation does not match. Such a call reached no other generation, and one refused at its start did nothing at all. Only if Claude exited during the call and another start took the name can input sent earlier in that call have reached the old session (from the code; not seen live). Call status; do not repeat the call by name. Other tools take no pair. On Linux and macOS leave it out: tmux rejects it.
• Treat managed terminal name, Remote Control remoteName, conversation title, and conversation UUID as different identities.
• Prefer submit_prompt plus wait_for_claude_turn. Use terminal capture for attention states and low-level diagnosis, not as the clean answer channel.
• Claude spinner phrases are variable. Trust the returned structured state, completion ordering, and transcript cursor rather than matching a particular phrase.
• Treat workflowPending: true as busy even when the composer appears idle. The MCP combines the exact dynamic-workflow wait control line with sanitized session-log task lifecycle evidence and blocks submit, low-level text/Enter, rename, replacement, and ordinary stop with reason workflow_pending. On Claude 2.1.289 the Waiting for N dynamic workflows to finish row stays on screen after the workflow ends. Once a complete, certain session log (workflowObservationCoverage: "full") has seen the workflow launch, the log decides and the stale row is neither pending nor busy evidence. With head_tail coverage, or before the log has seen a launch, the row still holds workflowPending. Then confirm in Claude that the workflow finished; with authorization, and only when terminalState is otherwise idle and the composer is empty, repeat the blocked submit_prompt, send_text, send_key, rename_claude_session, or stop_remote_control call with force: true (a replacement uses forceKillExisting: true; wait_for_claude_turn takes no force).
• Read terminalState separately from the combined state. Workflow activity can make state busy while terminalState remains idle. When workflowPending is false, workflowPendingEvidence is intentionally empty; use workflowObservationCoverage and workflowObservationSkippedBytes to tell a full scan from a bounded head/tail recovery of a large session log. Treat workflowObservationUncertain: true as a fail-closed pending state: skipped-middle history cannot prove current workflow completion or current model, effort, or permission posture. A current explicit workflow counter or bound interruption in the observed tail can restore current-state certainty, but a counter alone cannot clear known pending workflows. The skipped middle is never replayed, and coverage stays head_tail.
• Treat C-m, C-j, KPEnter, and other documented Enter equivalents as submissions. They use the same pending-workflow gate as Enter and Return.
• Do not infer workflow completion from a final assistant record, worker results, an idle composer, or a pending count. The MCP releases a tracked workflow whose launch bound a task ID and a run ID only on current evidence for it: a direct terminal record with the matching run ID, a completed TaskOutput retrieval (older CLIs; Claude 2.1.289 has no TaskOutput tool) or Claude's own task notification for that task (a run ID may be absent but must match if present; notification statuses completed, failed, killed, or an older terminal word release, while blocked and unknown statuses stay pending), or the workflow snapshot of a fresh native-Windows managed conversation. A current bound interruption can also clear the tracked set. Keep waiting until workflowPending clears or handle a bounded timeout. If Claude shows the workflow finished but it stays pending, use the authorized force recovery described above.
• Treat tool_use as in progress. Check textTruncated, textSha256, and textCharacters before assuming a large report was returned in full. Use get_claude_result with resultId for exact chunks until hasMore is false.
• Never type over a busy or non-empty composer unless the user deliberately authorizes an inspected force action. Native-Windows keyword mode never crosses a draft or paste placeholder: pasteMode: "keyword" needs an empty composer even with force: true, and otherwise returns preflight_blocked (active_input_not_empty or paste_pending, plus forceRefused: true when force was requested) without sending anything.