Start one consultant (target) or several (targets); returns job_id or group_id immediately. Available (name: CLI, model): (none -- no consultant CLI is installed) Pass the user's name/CLI/vendor/model wording; target_ambiguous lists candidates, target_unavailable lists available ones. For a user's second-opinion request without a named consultant, use targets: ["all"]; for a named consultant, use target. targets: ["all"] asks every available consultant, including your CLI (marked same lineage). Each gets an identical brief; one concurrency seat each. Model and limits are operator-owned. concurrency_limit starts nothing. Follow details: wait on already-known holders or retry_after_ms; at most two retries within 60 seconds, only if retryable. Never inspect/cancel unrelated work or replay a start receipt. followup_to: completed job_id or group_id; only question is also needed. Inherits the prior inputs; explicit fields, even empty, replace them. Files are recopied in their current state; add expose_paths or set expose_from_parent: false. retry_of: failed job_id alone replays saved inputs in a new chain, with current files and a link to the failure. Use successful group answers first; retry only failed members after fixing auth/quota causes. No other request fields. mode defaults from context: facts = explore, proposal = review, proposal + counterpoints = debate. For code: expose_diff: { repo, base? } for changes, expose_paths: [absolute paths] for relevant callers, tests and config. Files are copied read-only, secrets masked; symlinks/protected paths skipped. Caps: 20 entries, 500 files, 5 MB. Inspect files_exposed and the manifest; missing premises stay unknown. Other context: facts, artifacts (logs/test output). network: on/off controls web access, not model submission or local reads; boundary states CLI-specific enforcement. caller/caller_model identify you; initiator: user only when asked. Optional prediction: { expected, worry } stays local.

Wait for job_id or group_id (default 45000 ms; wait_ms: 0 polls). Repeat while running; A group wait returns when another member finishes; check available advice while others run, then wait again for new completions. progress is sent every 15 s. completed has advice; failed has failure.kind and no advice. A group includes statuses and comparison.by_target: actual model, stance, concrete recommendation, conditions, points/grounds, redaction observations and truncation markers. Null observations are unknown; matching stances need not recommend the same action. network is this round's web policy; limits.network is the server default. Read grounds already present. For missing/shortened deciding points use job_id + ids: ["f2", "u1"]; job_id alone gets one whole answer. detail: full adds the brief/question. Check deciding claims, batch checked entries in consult_record, then report; unchecked points stay unverified.

Stop a running consultation (job_id) or fan-out (group_id). Other severally servers receive a stop request. The process tree is signalled where the platform permits; on Windows descendants may survive if the CLI already exited.

Save checks against point IDs from consult_get: f1 (finding), u1 (unknown), c1 (next check). verdict = confirmed/refuted/not_applicable/unverifiable/unverified; effect = decision impact; note = checked inputs, environment, evidence and limits. Confirmed is not adoption; refuted means the claim did not hold in the recorded check. Partial checks stay unverified with their scope in note. Record only checked points; same ID replaces its entry. Use records: [{ job_id, entries, reflection? }, ...] for one fan-out batch. Invalid jobs/IDs prevent every write; write failures appear per job in results. Saved history survives restart; verdict/effect and bounded new evidence notes go to follow-ups. Legacy notes remain local until recorded again. Optional reflection describes what the answer added over your prediction.

Render one consultation (chain_id) or fan-out (group_id) as Markdown: the brief as sent, each answer as it came back, and verdicts on each point. Starts with verified points/effects/evidence and unverified count; no inferred adoption or scoring. The text is returned, not written: put it beside the code the decision is about. Reads the history, so an earlier session's consultation can be exported.

Recent consultations, earlier sessions' too (this_session is false), newest first, with their ids, status, summary, verdicts, latest_completed_job_id and rounds_remaining. Optional target, question (case-insensitive substring), chain_id filters apply before limit. Pass filters at the top level, not in request. usage_totals is the tokens/cost reported this session, retries included. A figure a CLI does not report is named under not_reported, not counted as zero.