返回 CodeWhale
READ_MEDIA.md
根目录 / docs / READ_MEDIA.md
1 # `read_media` Operator Guide
2
3 `read_media` is a safe, first-class image reading and preprocessing tool for Codewhale v0.9.10. It allows vision-capable coding models to inspect visual assets (diagrams, UI mockups, screenshots, rendered graphs) with strict security, memory bounds, and privacy guards.
4
5 ---
6
7 ## 1. Overview and Scope
8
9 - **Supported Formats:** PNG, JPEG, GIF, WebP.
10 - **Out of Scope:** Video files, audio streams, and background/automatic screenshot watching are deliberately excluded.
11 - **Provider-Neutral Wiring:** Decoded and normalized images are converted into native image parts across all supported providers (OpenAI Chat Completions, Anthropic Messages, and OpenAI Responses API).
12
13 ### Show the agent a screenshot
14
15 - Paste a clipboard image into the composer with the normal terminal paste
16 shortcut, or run `/attach <path>` for an existing PNG, JPEG, GIF, or WebP.
17 - A visible attachment row appears above the composer before the turn is sent.
18 Temporary macOS `NSIRD_screencaptureui` paths are copied into Codewhale's
19 stable attachment store when ingested.
20 - Ask the agent to inspect the screenshot. The image is sent only as part of
21 that explicit turn action; merely having a screenshot path or artifact does
22 not trigger background analysis.
23
24 `read_media` is the corresponding agent-side path for inspecting another
25 image later in the task without requiring the operator to attach it again.
26
27 ---
28
29 ## 2. Activation and Catalog Policy
30
31 - **Default-Off / Deferred Loading:** To preserve model context budgets, `read_media` is registered as a deferred tool (`defer_loading = true`) rather than occupying active slots in the default core catalog.
32 - **Explicit Model Invocation:** The tool is invoked explicitly by name when the model or user requests media inspection.
33 - **Eager Configuration:** Operators who want `read_media` always pre-loaded in the model tool catalog can configure:
34
35 ```toml
36 [tools]
37 always_load = ["read_media"]
38 ```
39
40 ---
41
42 ## 3. Tool Parameters and Schema
43
44 ```json
45 {
46 "path": "docs/architecture.png",
47 "crop": {
48 "x": 100,
49 "y": 50,
50 "width": 800,
51 "height": 600
52 },
53 "detail": "auto"
54 }
55 ```
56
57 | Parameter | Type | Required | Description |
58 |-----------|------|----------|-------------|
59 | `path` | string | **Yes** | Workspace-relative or trusted external path to the image file. |
60 | `crop` | object | No | Optional pixel bounding box `{ "x": u32, "y": u32, "width": u32, "height": u32 }` (0-indexed). |
61 | `detail` | string | No | Resolution target: `"auto"` (max 2048px, default), `"low"` (max 1024px), `"high"` / `"original"` (up to 4096px). |
62
63 ---
64
65 ## 4. Safety, Privacy, and Guardrails
66
67 ### 4.1. Workspace Boundary and Credential Protection
68 - **Workspace Containment:** Paths must resolve within the workspace or user-approved trusted external paths (`/trust`). Symlink escapes outside trusted roots are rejected.
69 - **Credential Protection:** Codewhale configuration (`config.toml`, `.codewhale/`, `.deepseek/`, secrets directory) cannot be read via `read_media` and will fail with `PermissionDenied`.
70
71 ### 4.2. Decompression-Bomb and Memory Limits
72 - **Source Byte Limit:** Maximum source file size before decoding is **20 MiB** (`MAX_SOURCE_IMAGE_BYTES`). Oversized files are rejected before allocation.
73 - **Dimension Guards:** Maximum permitted image width and height is **8192 px** (`MAX_IMAGE_DIMENSION`).
74 - **Pixel Budget:** Maximum total pixel count is **33,554,432 pixels** (~33.5 megapixels, `MAX_IMAGE_PIXELS`).
75 - **Memory Ceiling:** Safe memory allocation during decode is capped at **64 MiB** (`MAX_DECODE_ALLOC_BYTES`).
76 - **Wire Payload Limit:** Re-encoded image payload is capped at **5 MiB** (`MAX_WIRE_IMAGE_BYTES`), matching provider constraints.
77
78 ### 4.3. Active Route Vision Checks
79 - Before reading an image, `read_media` inspects `context.route_capabilities.image_input`.
80 - If the active model route explicitly lacks vision support (`CapabilityState::Unsupported`), the tool returns an actionable error directing the operator to switch to a vision model:
81
82 ```text
83 read_media: the active model route does not support image input. Switch to a route marked vision-capable with /model, or configure the route's image_input capability, then try again.
84 ```
85
86 Only a known `Unsupported` capability blocks the explicit tool call. An
87 `Unknown` capability is admitted deliberately, matching normal attachment
88 routing: custom and self-hosted providers often do not publish modality
89 metadata, so their provider response remains authoritative. Operators who
90 need a fail-closed route can set its `image_input` capability explicitly.
91
92 ---
93
94 ## 5. Typed Receipts and Wire Integration
95
96 Each successful execution yields:
97 1. **Human-Readable Receipt (`content`):** Summarizes original format, source and final dimensions, crop details, and byte sizes.
98 2. **Typed JSON Metadata (`metadata`):** Contains structured dimension, crop, and byte information without exposing credentials or internal tokens.
99 3. **Rich Image Content Block (`content_blocks`):** Attaches a standardized `ToolResultContentBlock::Image` which provider adapters wire into outbound requests:
100 - **OpenAI Chat Completions:** `image_url` block in following user message.
101 - **Anthropic Messages:** `image` base64 source inside `tool_result` block.
102 - **OpenAI Responses:** `input_image` block inside function call output.
103
103 lines MARKDOWN