/loop

/thoth:loop

Generated Surface

This file is generated from thoth.command_specs.COMMAND_SPECS. Do not hand edit.

Real Runtime Execution

The repo-local Thoth runtime command for this slash command has already been executed before Claude sees this prompt.

"${CLAUDE_PLUGIN_ROOT}/scripts/thoth-claude-command.sh" loop --host claude $ARGUMENTS

Response Contract

• Treat the structured bridge payload above as the only authority for this command invocation.
• Do not invent or hand-roll alternate .thoth layouts, migrations, run ledgers, or host projections.
• If bridge_success is false, explain the exact failure from the bridge payload and stop.
• If bridge_success is true and packet.dispatch_mode is live_native, the command is NOT finished yet: execute packet in this Claude session using native tool use or subagents as needed.
• For run and loop, use packet.strict_task.goal_statement, packet.strict_task.implementation_recipe, and packet.strict_task.eval_entrypoint as the task contract; do real code edits, run the relevant validators, and do not stop after merely restating the packet.
• Prefer running packet.strict_task.eval_entrypoint.command exactly as provided; do not hand-roll a parallel service lifecycle when the packet already gives a validator entrypoint.
• If packet.executor is codex, the substantive execution must really flow through Codex from this Claude session. Do not silently do the work yourself and then claim Codex parity.
• For packet.executor == codex, use the installed Codex surface or codex exec from Bash to perform the requested run/review work, keep the Thoth ledger writes in this session, and preserve the same packet / acceptance shape.
• While executing a live packet, keep .thoth updated only through the internal runtime protocol commands included in packet.protocol_commands.
• End the command only after the protocol reaches a terminal state via complete or fail; leaving the run in prepared or running without a terminal protocol write is a contract violation.
• If packet.dispatch_mode is external_worker, do not duplicate the work locally; report the run id, worker mode, and the correct follow-up (status, watch, dashboard, or report).
• For review, inspect packet.target, produce structured findings matching packet.required_review_shape, and finish by writing them through the protocol rather than free-form narration only.
• For review with packet.executor == codex, make Codex inspect packet.target, then write the resulting structured findings through packet.protocol_commands.complete instead of returning prose only.
• If you only summarize the packet, list the task, or say what should happen next without executing it, treat that as failure and call fail with the exact blocker.

Scope Guard

CAN:

• Create or resume a durable loop packet
• Attach/watch/stop through the same runtime state machine
• Delegate work to Codex without changing authority write shape

CANNOT:

• Use detached live execution without --sleep
• Depend on subagents or hooks for correctness

Runtime Contract

• Durable: yes
• Codex executor allowed: yes
• Hooks required for correctness: no
• Subagents required for correctness: no
• Lifecycle: prepare -> live-native|external-worker -> protocol-update -> attach/resume/watch/stop -> acceptance
• Acceptance: Loop lifecycle is durable and recoverable through attach/resume/watch/stop, live mode stays inside the current host session, and heavy acceptance expects bounded rounds/time rather than unbounded supervisor churn.

Interaction Gaps

• Strict task id

Shared Authority

Both Claude and Codex surfaces must write through the same .thoth authority tree. Host differences are interaction-only and must not change ledger shape.