Consult LLM

Reference for invoking the consult-llm CLI. Workflow skills delegate here for mechanics; they focus on orchestration.

Invocation

Run consult-llm with the prompt on stdin, using a quoted heredoc.

cat <<'__CONSULT_LLM_END__' | consult-llm -m <selector> -f src/foo.rs -f src/bar.rs
<prompt body>
__CONSULT_LLM_END__

Rules:

• Run Bash in the foreground (synchronous, no run_in_background). Only background the call when the caller explicitly passes --background. Always set timeout: 1800000 (30 minutes) — LLM calls routinely exceed the 2-minute default.
• ALWAYS use <<'__CONSULT_LLM_END__' (quoted, with this exact terminator). The single quotes prevent shell expansion of $var, backticks, and escapes. The specific terminator __CONSULT_LLM_END__ is chosen because it won't appear in model responses — never use EOF or PROMPT which commonly appear in code samples and would silently truncate the prompt.
• Fallback to --prompt-file <path> if the prompt contains __CONSULT_LLM_END__, or on Windows/PowerShell. Write the prompt to a temp file with $(mktemp), then pass it via consult-llm --prompt-file "$f" ….
• Stdout layout. First line is [model:<id>] [thread_id:<id>], then a blank line, then the response body. In --web mode the prefix is just [model:<id>] (no thread).
• Multi-turn. Read [thread_id:xxx] from line 1 and pass it back with -t <id> on the next call. Thread IDs are opaque strings — don't modify them. Not portable across backends.
• Stderr carries progress/spinner output. Ignore it.
• Exit codes. 0 success, 1 backend/network error (includes thread-not-found), 2 usage error, 3 configuration error (missing API key, unsupported backend).

Models

Selectors and allowed models resolvable in this environment (availability depends on which API keys are configured):

!`consult-llm models`

Pass a selector or exact model ID to -m only when overriding defaults. With no -m, consult-llm uses configured default_models when that config key is present and non-empty, preserving order and duplicates. If default_models is unset or empty, it falls back to default_model, then the built-in fallback model. For same-prompt multi-model calls, omit -m to use those configured defaults; use repeated -m only for explicit overrides. For --run, create one --run model=<model>,prompt-file=<path> entry per desired run; --run does not consume omitted--m defaults. -m is ignored when --web is used.

Multi-model: repeat -m to consult multiple model positions in parallel (e.g. -m gemini -m openai, max 5 total runs). You may repeat the same selector/model (e.g. -m openai -m openai) to get independent calls with the same prompt. The response is a group format: first line is [thread_id:group_xxx], each model's answer under a ## Model: <id> header preceded by [model:<id>] [thread_id:<per-model-id>]. When the same resolved model appears more than once, only those duplicate sections use ## Model: <id>#K and [model:<id>#K] labels. Pass -t group_xxx to resume all group positions together on the next turn; pass an individual per-model thread ID with a single -m <model> to resume just that model outside the group context.

Task modes

Pick a --task mode based on the kind of question. Omit for neutral general-purpose.

Web mode

--web copies the formatted prompt (system prompt + user prompt + file context) to the clipboard and exits 0 instead of calling an LLM. Only use when the user specifically asks for browser/web mode. After invoking, wait for the user to paste the external LLM's response back — do not continue implementation on your own. -m is ignored in this mode.

Prompt authoring

Ask neutral, open-ended questions. Do not suggest specific solutions in the prompt body - that biases the analysis. Let the LLM form its own view.

Present attached context as starting evidence, not an exhaustive set. The consulted model works from that evidence and may append a request for exact additional context when a material gap emerges during analysis.

Context request loop

Apply this loop after every consult-llm response, before presenting, synthesizing, feeding it to another model, or acting on it.

Every response is a bounded answer. An unfenced final ## Context request section means the consultant found missing context that could materially change a stated conclusion. Each item identifies its kind, the exact context needed, and which conclusion it could change.

1. For an artifact request, gather the exact file, command output, log, or diagnostic. Prefer raw output and attach it with -f.

2. For a clarification request, answer from the caller's conversation context when possible. If only the user can answer, ask the user when the enclosing workflow permits interaction. Otherwise tell the consultant that the information is unavailable.

3. Gather every requested item in the same follow-up, then resume only the requesting model's per-model thread with the same model and -t <thread_id>. Attach only the additional artifacts and provide any clarification inline. Use this continuation prompt:

   Here is the requested context. Requested artifacts are attached where applicable.
   
   [clarifications or unavailable items]
   
   Revise your original answer using this context. Say plainly which conclusions change. If material uncertainty remains, state the unresolved gap. Do not issue another context request.
   

4. Perform at most one context follow-up per model per consultation stage. Treat the revised answer as final. If it still ends with a context request, preserve that request as unresolved uncertainty and continue the workflow without another round.

For multi-model output, handle each requesting model independently using the thread ID from its section. Keep responses from models that did not request context. Do not resume the whole group just to satisfy one model.

In web mode, ask the user to provide the requested context in the existing browser conversation and paste back the revised answer.

Flags

Run consult-llm --help for the authoritative flag list.

Diff context adapts to the resolved backend. API and web runs receive the diff contents. CLI runs receive the repository, base, and path scope, then inspect the diff with their read-only repository tools.

File context (-f) best practices

The consulted LLM has no access to your conversation history. Anything it needs - source files, logs, command output, traces, timelines, error messages - must be attached with -f.

• Include conversation artifacts. If the current session already produced diagnostic output relevant to the question (log excerpts, traces, reproduction steps, command output), attach it as a temp file. Prefer raw evidence over prose summaries when both exist.
• Re-run the original command piping to a temp file (cmd > /tmp/artifact.txt) instead of writing output from memory. This is cheaper, faster, and preserves the exact output.
• Source files and diagnostic artifacts are both first-class -f inputs. Do not limit context gathering to source code.
• Follow material context requests using the bounded context request loop above.

Per-model runs

Use --run when a workflow needs to query multiple models in parallel with different prompt bodies. Do not use it for ordinary multi-model calls where the same prompt goes to every model — repeat -m for that.

GEMINI_PROMPT=$(mktemp)
CODEX_PROMPT=$(mktemp)

cat <<'__CONSULT_LLM_END__' >| "$GEMINI_PROMPT"
[prompt for Gemini]
__CONSULT_LLM_END__

cat <<'__CONSULT_LLM_END__' >| "$CODEX_PROMPT"
[prompt for Codex]
__CONSULT_LLM_END__

# First call — no existing threads yet
consult-llm \
  --run "model=gemini,prompt-file=$GEMINI_PROMPT" \
  --run "model=openai,prompt-file=$CODEX_PROMPT"

# Subsequent calls — continue each per-run thread
consult-llm \
  --run "model=gemini,thread=$GEMINI_THREAD,prompt-file=$GEMINI_PROMPT" \
  --run "model=openai,thread=$CODEX_THREAD,prompt-file=$CODEX_PROMPT"

# Duplicate resolved models are allowed; use distinct prompt files and distinct per-run threads.
consult-llm \
  --run "model=openai,prompt-file=$PROMPT_A" \
  --run "model=openai,prompt-file=$PROMPT_B"

Each --run value accepts model=<selector-or-id>, prompt-file=<path>, and optionally thread=<id>. Use mktemp for temporary prompt files and always use __CONSULT_LLM_END__ as the heredoc terminator. Use >| to overwrite temp files in zsh (avoids noclobber errors).

Constraints: max 5 total runs, cannot combine with -m/-t/--prompt-file/--web, duplicate resolved models are allowed, duplicate explicit thread=<id> values are rejected, thread=group_* is rejected because --run uses per-run thread IDs, shared -f and --diff-* context applies to every run, prompt-file paths with commas are unsupported.

Output is the same group format as multi-model -m calls. Extract per-run thread IDs from each section header for subsequent --run thread=... turns.