imladris-deliberate

name

imladris-deliberate

description

Convene a panel of independent AI models to deliberate on a hard question via the Imladris MCP server (stdio), returning structured analysis (consensus, contradictions, partial coverage, unique insights, blind spots) plus all raw panel answers. Use for research, expert critique, "compare and contrast", architecture/design trade-offs, or any task where being wrong is costly. You remain the final author.

Imladris Deliberation

When to use

Reach for Imladris when a task is hard enough that one model's answer is risky: research questions, design/architecture trade-offs, "compare X and Y", expert critique, or anything where being wrong is costly. For quick lookups, answer directly — deliberation costs ~N panel calls + an analysis call.

Quick Start

Imladris is an MCP stdio server — there is no HTTP endpoint, port, or token. Your harness spawns imladris-mcp; call its tools directly. The /council prompt wraps the one-shot flow below. Use /council-session when the user wants a sustained back-and-forth with the same council.

Call the imladris tool:

• prompt (required), context (optional extra context for every panellist)
• preset or panel — which of your configured providers answer
• analysis_model — override which provider runs the analysis judge
• thread_id — present means session mode; Imladris reads/writes the local ~/.imladris/sessions/<thread_id>.json file and reconstructs messages[]
• prior_answer — in session mode, pass your previously authored answer for this thread so Imladris can store it as the assistant turn

Inspect configuration and health with imladris_status (resolved roster, roles, presets, budget thresholds, budget estimates — never secrets or env-var names). Clear local session files with imladris_clear_sessions.

Workflow

1. If meta.ok == 0, every panellist failed and the tool returns a clean meta.failure — say so instead of inventing evidence; let the harness fall back.
2. Send the user's question as prompt. Add context for files/constraints.
3. Pick a preset (your configured presets — use a local-only preset when the data must not leave the network) or an explicit panel.
4. You are the final author. Read the analysis — especially blind_spots and contradictions — and the raw_answers, then write the answer, preserving genuine disagreements and caveats rather than smoothing them away.
5. For a session, use the same thread_id on each turn and pass your last answer as prior_answer. If the host drops it, Imladris backfills from stored analysis.
6. When the analysis flags a claim as low-confidence, verify it before relying on it.

Response Handling

• panel[] — per-provider status metadata (id, status, latency, tokens). The raw answer text is returned in raw_answers[]; partial results are normal.
• analysis — consensus / contradictions / partial_coverage / unique_insights / blind_spots / confidence_notes. Do not treat consensus as proof.
• raw_answers[] — all successful panel answers, returned unconditionally.
• meta — timings, ok/failed counts, analysis provider id, cost estimate, budget status, session warnings, and a typed failure if the deliberation degraded.
• thread_id / compacted — present on session responses; compacted means older local history was omitted from the provider request.
• text — a Markdown rendering of the same content.