ai: unify claude JSON instructions and error handling

[derekbarbosa]

Centralizes JSON instruction logic in src/ai/mod.rs by adding AiResponseFormat::format_json_schema_instruction().

Improves claude-cli in src/ai/claude_cli.rs:

• Introduces ClaudeCliError implementing ClassifyAiError for better retry handling.
• Switches build_prompt to use the new centralized JSON instructions.

Unifies Claude API in src/ai/claude.rs:

• Injects the centralized JSON instructions into the system prompt.
• Removes redundant/duplicate formatting logic.

Updates error classification in src/ai/mod.rs to support the new ClaudeCliError.

fixes #171

[HouMinXi]

Tested PR #174 on Vertex AI with claude-sonnet-4-6. Still fails at Stage 1:

Error: Stage 1 failed to produce valid 'concerns' array after 3 attempts

The response_format injection into system prompt looks correct (Phase 0 should work now), but Stage 1 uses tool use, and the tool schemas in src/worker/prompts.rs still have uppercase type names:

// prompts.rs:486-490
"type": "OBJECT",
"type": "ARRAY",
"type": "STRING"

Claude API requires lowercase ("object", "array", "string"). Gemini accepts both cases but Claude rejects uppercase with:

tools.0.custom.input_schema.type: Input should be 'object'

This is the same issue as #107. PR #108 (closed) fixed both problems (response_format + uppercase schemas), but only the response_format half made it into this PR.

Quick fix -- in src/worker/prompts.rs, change all inline JSON schema types to lowercase:

-"type": "OBJECT"
+"type": "object"
-"type": "ARRAY"
+"type": "array"
-"type": "STRING"
+"type": "string"

Also check src/worker/tools.rs if it has the same pattern (I didn't find uppercase there, but worth double-checking).

[derekbarbosa]

Reworded some commits, added some more error handling for claude_cli, and unified JSON prompt building/formatting for claude responses.

[derekbarbosa]

Addressed some review-pr concerns, mainly:

src/ai/claude.rs:L322-326: 🟡 risk: system_blocks.push appends a new text block. Some Anthropic-compatible proxies (e.g. AWS Bedrock, Google Vertex AI) might reject multiple consecutive
text blocks in the system prompt array. Consider appending the instruction to the last text block if one exists.

src/ai/mod.rs:L95: 🔵 nit: if serde_json::to_string fails (unwrap_or_default returns ""), the schema instruction becomes invalid JSON matching this schema: .. Since schema is dynamic,
might be safer to handle serialization errors gracefully (e.g., fallback to schema-less instruction).

Tested using us-east5 endpoint, E2E patch submission successful with sonnet & opus models via vertex.

ready for review! @rgushchin

[derekbarbosa]

Hi @rgushchin, sorry for the noise. Seems like there are others encountering the JSON encoding issue. Kindly asking for a review :-)

[anders-heimer]

Hi Derek,

Thanks for the PR. I ran a check against the JSON-format failures collected from my various benchmark runs. The response-format side looks useful and centralizing the JSON response-format instruction should help cases where the model returns prose or JSON that misses the required top-level schema fields.

I also replayed a small matrix of observed JSON/format failure classes. The PR helps with the instruction/schema side.

Br, Anders

[derekbarbosa]

Hey Anders,

Thank you for posting your findings, it certainly helps. I'm glad the proposed changes address some of the formatting issues. Hopefully pairing this with TodoWrite may increase the consistency wrt failures across the board