Vision Manual

Vision manual

Routine calls use the schema directly; load only the reference needed for setup or recovery. This installed manual reads no configuration and invokes no provider.

First action

vision(action="analyze", input={"image_path": "...", "question": null}, reasoning="...")

One existing image per call; relative paths use the workdir and null selects the default image prompt. PNG/JPEG/WebP/GIF have known MIME types, not a guarantee of provider support. Success is {"status": "ok", "analysis": text}; failures are structured errors. The schema owns all five actions and strict input fields.

Choose depth

• Actions and results: check resolves a route without an image (not live model acceptance); list reads declarations without service or credential construction; settings shows the applied bind snapshot.
• Routing: allowed-preset borrowing, Claude CLI, recovery.
• Backends: local server/model setup, MLX, troubleshooting.
• Settings: exact source/precedence and owner procedures.

Route and authority boundary

Without preset, use the configured service or the active provider's own family (openai, anthropic, codex, or claude-code) with its effective endpoint, model, and credential. A non-null preset must be in manifest.preset.allowed; it borrows the allowed preset's own provider/model/endpoint/wire/credential for one call, without switching the active preset. Missing or unsupported routes fail closed.

No provider, model, credential, preset, MCP, or CLI fallback is automatic. Vision never auto-invokes MCP. Ask the human before changing authorization/configuration, installing a server, or pulling a model. Alternatives require an explicit later action. Never print keys, tokens, environment values, headers, or private URLs.

Setting anchors

SHOW is read-only: fresh rows from the applied snapshot, no file/environment re-read or configuration validation/write. Rows contain only key, current, default, configurable, comment; sensitive fields and path-like models are redacted. Unavailable truth fails the entire inventory, never guessed rows. Follow the exact owner section below, make only authorized changes, then refresh/relaunch and SHOW again.

Setting: provider

See provider.

Setting: base-url

See base-url.

Setting: model

See model.

Setting: api-key

See api-key.

Setting: api-key-env

See api-key-env.

Setting: max-tokens

See max-tokens.

Setting: wire-api

See wire-api.

Setting: default-headers

See default-headers.

Setting: token-path

See token-path.

Setting: instructions

See instructions.

Setting: max-output-tokens

See max-output-tokens.

Setting: timeout

See timeout.