Extract prompts for CanyonOS

Create or update prompts.yaml from the project's actual LLM call paths, following Prompt management. Preserve prompt content and application behavior; do not invent or improve prompts.

1. Locate the source and config

• Work in the project root: .canyon when the project folder has one, otherwise the project folder itself. Find config/workflow_config.yaml there and write prompts.yaml beside it. Respect an explicitly chosen destination.
• Trace the manifest's agent entrypoints, relative to the root, to their LLM calls.
• Search for message roles, prompt constants, template constructors, formatting functions, and referenced prompt files. Follow imports and builders to the text actually sent to the model; a variable named prompt alone is not evidence.

2. Extract templates without changing meaning

• Name each prompt <agent>.<function> using the agent name and the function making the LLM call, rather than an internal string builder. Keep existing prompt names when refreshing a file.
• A prompt is the call's system prompt: the instructions, without the runtime input. The gateway sends the live version as the system prompt of every call the function makes, replacing any the code set, so the code keeps sending only its own message (the request, the figures, the conversation).
• Preserve wording, whitespace, punctuation, examples, and literal braces. Combine adjacent string literals as the source does. When the instructions live in the user message, split them out as the system prompt and leave the runtime input in the code; report the split.
• Never insert a real request, retrieved content, conversation, or secret into the YAML. Runtime values stay in the code's own message, not as placeholders in the prompt.
• Do not flatten multiple system messages, assistant/tool messages, multimodal content, or framework-specific templates into one string when that changes semantics. Report those call paths as unrepresented rather than claiming full coverage.

3. Write versioned YAML

The top-level prompts mapping contains a list of versions for each prompt name. Every version has version and content; updated_at is an optional, quoted UTC ISO 8601 timestamp, and live: true marks the version agents are sent.

• Compute the hash exactly as the dashboard does: the first eight lowercase hex characters of SHA-256 over the UTF-8 content text.
• Use <agent>-<function>-<hash>-v<n> for version, replacing every dot in the prompt name with a hyphen. Start n at 1; on a changed prompt, append a version with one more than the highest trailing v<n>. With no live: true, the highest n is live, regardless of list order or timestamps.
• Leave unchanged prompts unchanged. Preserve old versions, unrelated prompts, and other configuration fields; do not overwrite the file with a fresh list.
• Use YAML quoting or literal block scalars that preserve the extracted text. Check the parsed strings before hashing, including their trailing newlines. If no real LLM prompts are found, report that instead of fabricating entries.

4. Verify and report

• Parse the resulting YAML and verify version lists, required string fields, hashes, and revisions. Check that each new entry maps to an actual LLM call.
• Render representative inputs using the original builder and the extracted template, then compare the messages exactly. Cover conditional sections and formatting that affect the text. Use isolated string builders or mocks so this verification does not call model providers or start the application.
• Report the destination, extracted names, source locations, placeholder bindings, unrepresented call paths, and validation results.

Creating the file makes prompts available to the running project's dashboard after controller startup. The gateway sends each function's live version as the system prompt of its LLM calls, so the code must stop sending those instructions itself; remove them from the code only when requested, and keep the runtime message it sends. Once a prompt is stored by the dashboard, later changes to it in prompts.yaml are ignored; a prompt new to the file is stored the next time the dashboard reads prompts. Dashboard edits are not written back to the YAML.