rlm-auto

rlm-auto routing skill

You only run this skill when a rlm-auto: system reminder appears earlier in the turn. The reminder will tell you:

• the verdict (rlm / ambiguous / direct)
• the signals
• the cost estimate
• the decision_id
• the auto-approve cap

This skill never invokes a sibling plugin's scripts by path. It invokes the public slash commands the user already has registered (/rlm:rlm-plan and /rlm:rlm) so the rlm plugin's install location stays opaque.

Workflow when verdict = rlm

1. Plan the run silently. Invoke the planning slash command:

   /rlm:rlm-plan <target-from-prompt> "<user's question>"
   

   The command outputs the same Markdown + fenced JSON block as rlm_plan.py. Parse the JSON block.

2. Check the cap deterministically. Pipe the plan's JSON into the bundled cap-check helper. Do not eyeball the math:

   py -3 "${CLAUDE_PLUGIN_ROOT}/scripts/check_cap.py" --plan-json '<plan-json>'
   

   The script reads ~/.rlm/auto-config.json for auto_approve_cap_usd, compares it to plan.est_cost_usd.high, and prints one line: decision=auto_approve or decision=ask reason=<reason>. Exit code matches: 0 for auto-approve, 1 for ask. Trust the exit code.

3. If decision=ask, emit exactly one line and wait:

   rlm-auto: estimated cost up to $X.XX exceeds the auto-approve cap ($Y.YY). Run anyway? (yes/no)

   Do not paraphrase, do not ask follow-ups. On yes, proceed; on anything else, fall back to a normal direct flow.

4. Execute via the public slash command. Once approved (auto or explicit), invoke:

   /rlm:rlm <target> "<question>"
   

   That command handles health probe + native fallback automatically.

5. Produce the answer. Render the final answer to the user. Append exactly ONE footer line built from the estimate the system reminder gave you (or re-derived via the rlm-estimate skill):

   [rlm-auto] used RLM mode=<min|default|max> (est $X.XX-$Y.YY,
   ~Z% vs direct, +Wpp acc projected; log: .rlm/logs/<id>.jsonl;
   decision: <decision_id>)
   

   This footer is the user-visible transparency signal. Do NOT skip it unless show_footer: false is set in ~/.rlm/auto-config.json.

6. Attach an accurate outcome. If the runner reported a real cost figure, write it to the decision log so the evaluator grades against reality, not the projection:

   py -3 "${CLAUDE_PLUGIN_ROOT}/scripts/decision_log.py" attach-outcome \
     --decision-id <decision_id> \
     --outcome-json '{"ran":"rlm","actual_cost_est_usd":0.22,"rlm_fanout":12,"tool_calls":N}'
   

   Otherwise the Stop hook fills in what it can from session telemetry.

Workflow when verdict = ambiguous

Do not auto-route. Decide for yourself based on the user's actual question:

• If the question requires reasoning across the inputs (compare, aggregate, audit, summarize the whole), invoke /rlm:rlm-plan and follow the same cap-check + execute flow as above.
• If it's a lookup-style question on one of the inputs, proceed normally.

Either way, no footer is required when going direct.

Workflow when verdict = direct

Do nothing. The hook already logged the decision; the evaluator will grade it on session end.

Critical rules

• Never explain RLM unless asked. This skill is meant to be invisible. The one-line footer is the only transparency signal.
• Never put the user's raw input into your chat context. Pass paths to the RLM scripts. RLM's whole point is to keep big context outside the prompt.
• Honor the cap script's exit code. Even if you're confident, never bypass check_cap.py. The cap exists so a misclassification costs at most $cap.
• Never write to the user's settings.json. Configuration changes go through /rlm-auto:config.
• Don't grade the decision. That is the evaluator's job. Just route and write the outcome.

What this skill does NOT do

• It does not run the planner if the user has already typed /rlm:rlm directly. In that case, defer to the base rlm skill.
• It does not collect telemetry or send anything off the machine beyond what already flows to the API via the normal Claude Code request.
• It does not retry on RLM failure. If the RLM run fails or returns ok: false, fall back to a normal Read/Grep flow and write that outcome to the log so the evaluator can mark it.

Required dependency

This skill calls /rlm:rlm-plan and /rlm:rlm. If those commands are not registered, the rlm plugin is not installed. Tell the user to install it (claude plugin install rlm) and abort the auto-routing - fall back to a normal direct flow for the current turn.