thoth:auto

/thoth:auto

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.

THOTH_AUTO_ARGUMENTS_FILE="$(mktemp -t thoth-auto-arguments.XXXXXX)"
trap 'rm -f "$THOTH_AUTO_ARGUMENTS_FILE"' EXIT
cat > "$THOTH_AUTO_ARGUMENTS_FILE" <<'THOTH_AUTO_ARGUMENTS_EOF'
$ARGUMENTS
THOTH_AUTO_ARGUMENTS_EOF
"${CLAUDE_PLUGIN_ROOT}/scripts/thoth-claude-command.sh" auto --host claude --thoth-arguments-file "$THOTH_AUTO_ARGUMENTS_FILE"

Response Contract

• Treat the structured bridge payload above as the only authority for this command invocation.
• If bridge_success is false, report the exact bridge failure and stop.
• If run or loop is missing --work-id, show returned candidates and stop; do not invent, create, compile, or guess a work item.
• If bridge_success is true and runtime events are present, summarize progress, terminal status, and risk from those events only.
• Do not hand-edit .thoth or manually call runtime protocol commands; the Thoth RuntimeDriver advances phases.
• If packet.dispatch_mode is external_worker, do not duplicate the work locally; report the run id, worker mode, and the correct follow-up only.
• If you only describe what should happen next instead of reporting the executed runtime result, treat that as failure.
• Substantive execution must flow through packet.executor; by default this matches the host unless the user explicitly supplied --executor.
• Runtime lifecycle is plan -> execute -> validate -> reflect; execute owns the official validator receipt, validate confirms it mechanically.
• Execute must actively produce canonical acceptance evidence: missing artifacts, metrics, logs, receipts, benchmark output, service state, or files are execution work until produced or until a concrete root cause, blocker, or budget boundary is captured.
• For long-running work, first-artifact evidence proves startup only; continue or resume canonical execution until acceptance evidence, a concrete blocker, or a real budget boundary exists.
• Do not let healthy work be stopped just because a short observation window has not yet produced canonical evidence; stop or restart only as explicit debugging or cleanup with logs and a next action.
• If runtime budget expires before acceptance closes, preserve continuation evidence and the exact next command instead of presenting the work as passed.
• Live monitor should observe sparsely around every 288s; on clear runtime/env mistakes, append or interrupt guidance instead of only narrating.
• Trailing text/live corrections are temporary guidance only; never rewrite authority or validators.
• If the bridge payload exposes body.monitor_command, observe that command instead of executing work directly in the Claude session.
• Prefer the Claude Monitor tool with persistent=true for body.monitor_command when available; otherwise use Bash to run the same watch command in the foreground.
• Treat the monitor/watch JSONL stream as the only live progress authority; summarize progress and risks from those events only.
• If the live observer is interrupted, do not stop the auto controller unless the user explicitly requests /thoth:auto --stop <controller_id>.