ask_council
Send a question to a council of AI models and get a structured answer in five modes (individual, categorized, deconflicted, pooled, dialectic). Supports image questions and git-based repo reviews.
Instructions
Send a question to the model council and get a structured response. Mode: individual (each model answers separately), categorized (judge groups responses into agreement/complementary/conflicting), deconflicted (iterative loop — judge orchestrates re-questioning until conflicts resolve, returns a deconfliction score 0–100%), pooled (Delphi-style — members reconsider against a neutral, deduplicated, attribution-free pool of answers; no winner is forced, so genuine divergence is preserved), or dialectic (thesis/antithesis/synthesis — members defend their pick and critique the rest, the judge compiles a pros/cons dossier per option, then members re-select a ranked top-3). Attach images to ask a vision question — only auto-detected vision-capable members are queried; the rest are skipped and reported in visionRouting. For a repo review, pass git_ref (e.g. "uncommitted", "main..HEAD") instead of hand-listing files — the server runs git diff locally and attaches it as context. For a full repo-wide review (not just a diff), full_repo_access (default false, WARNING: read access to the whole repo — see its param description) grants claude-cli/codex-cli members read-only browse/read access.
Input Schema
| Name | Required | Description | Default |
|---|---|---|---|
| mode | No | Response mode override for this call only. | |
| files | No | Optional local file paths to read and attach as labelled context (default caps: 512 KB/file, 1.5 MB total, 32 files). Text only — use "images" for pictures. | |
| images | No | Optional local image paths (png/jpg/jpeg/gif/webp). Auto-detected vision-capable council members are queried with the image(s); members without vision support are automatically skipped for this call (see visionRouting in the result). Caps: 8 MB/image, 24 MB total, 6 images. | |
| context | No | Optional background text prepended to the question for every member. | |
| git_ref | No | Auto-attach a local `git diff` as context for a repo review, instead of hand-listing every changed file via "files". One of "uncommitted" (staged+unstaged vs HEAD), "staged", "unstaged", or a git revision/range (e.g. "main..HEAD"). Errors clearly on a bad ref, no changes, or a diff too large to attach (> 512 KB). | |
| verbose | No | deconflicted → include the initial categorization and per-round detail; pooled/dialectic → include the initial (round-0/thesis) raw member responses. | |
| git_repo | No | Repo directory to run git_ref in. Defaults to the working directory. | |
| question | Yes | The question or prompt to send to all council members. | |
| full_repo_access | No | WARNING: grants claude-cli/codex-cli repo exploration for a repo-wide review — ENFORCED DIFFERENTLY per provider. claude-cli: Read/Grep/Glob CONFINED to the repo root (real enforced boundary). codex-cli: cwd points at the repo, but its read-only sandbox does NOT confine reads to it — can read anywhere the OS user can (pre-existing behavior, not added by this flag; writes stay blocked everywhere). Defaults false; confirm with the user before setting true for an interactive call (an unattended review step you already control, e.g. end-of-workflow, may set it autonomously). Other members are unaffected. Repo root: git_repo, else cwd. | |
| reasoning_effort | No | How hard every member AND the judge think, for this call only — overrides the configured default. Higher levels give deeper answers at real cost in time and subscription quota, multiplied across members x rounds. A level the backend does not support is clamped to its nearest supported one ("max" runs as "high" on Ollama, "none" as "low" on claude-cli), so one setting works across a mixed council and no member is dropped for asking. Omit to use the configured default. | |
| max_deconflict_rounds | No | Max deconfliction rounds override for this call only. |