fix: doom loop not detected for repeated identical tool validation errors

[bug-ops]

Bug

When using orchestrator with Claude as default provider, the write tool is called via native tool_use path but Claude omits the content parameter, causing repeated execution failed: missing required parameter: content errors.

Investigation

1. Orchestrator correctly routes chat_with_tools to Claude (default = "claude")
2. Claude supports_tool_use() -> true, so native tool_use path is used
3. Tool schema is generated via schemars::schema_for!(WriteParams) and serialized with serde_json::to_value(&def.schema) which produces a RootSchema wrapper (includes $schema, title, definitions metadata)
4. Claude API input_schema expects a clean JSON Schema object with type, properties, required at the top level
5. The extra schemars wrapper may confuse Claude into not recognizing required fields

Likely root cause

tool_def_to_definition() in streaming.rs:757 passes the full RootSchema as parameters, but Claude API expects just the schema body. The schemars RootSchema has a nested structure:

{
  "$schema": "...",
  "title": "WriteParams",
  "type": "object",
  "properties": { "path": {...}, "content": {...} },
  "required": ["path", "content"]
}

If the extra fields cause Claude to misinterpret the schema, it may not enforce required parameters properly.

Proposed fix

1. Verify the actual JSON being sent to Claude API (add trace logging for tool definitions)
2. If schemars wrapper is the issue: extract .schema from RootSchema or strip $schema/title fields before passing to Claude
3. Add a test that validates the tool definition JSON structure matches what Claude API expects

Secondary issue

TUI header shows Model: glm-4.7-flash:latest instead of the actual Claude model because main.rs:267 sets model_name from config.llm.model (Ollama model) rather than from the orchestrator's active provider. This is cosmetic but misleading.

Steps to reproduce

1. Configure orchestrator with Claude as default and Ollama model in [llm]
2. Start TUI, ask agent to write a file
3. Observe repeated missing required parameter: content errors

[bug-ops]

Log analysis

From zeph.log, the orchestrator correctly delegates to Claude (default_provider=claude), and native tool_use path is used.

Timeline:

• Iterations 0-2: Claude calls bash tool correctly (git branch, git log, cargo doc)
• Iterations 3-7: Claude calls write with only path parameter, no content — fails 5 times with missing required parameter: content
• Iteration 8: Claude finally sends write with both path and content (full audit report) — succeeds

Root cause confirmed: Claude Sonnet is attempting to send a very large content field (~5KB markdown) and fails to include it in the first several attempts. The is_error: true tool result is correctly returned each time, but Claude keeps retrying with the same incomplete call.

This suggests the issue is model-side (content truncation or generation limits per turn), not schema-related. However, the doom loop detection should catch this pattern and break out earlier.

Action items:

1. Check doom loop detection for tool validation errors (repeated identical tool errors)
2. Consider adding the exact error message (which parameter is missing) to help Claude self-correct
3. Verify schemars output still matches Claude API expectations (secondary concern)

[bug-ops]

Correction

The write tool eventually succeeds on iteration 8 — PROJECT_AUDIT_REPORT.md is created. The problem is not broken functionality but wasted iterations: Claude makes 5 identical failed write calls (path-only, no content) before finally including content.

The doom loop detector should catch this pattern: same tool + same params + same error repeated N times. Currently it does not, letting Claude burn 5 unnecessary API calls (~60 seconds of wall time).

Revised fix

1. Enhance doom loop detection in check_doom_loop() to track repeated tool validation errors (same tool name + same error message)
2. After 2 consecutive identical tool errors, inject a hint into the tool result: e.g. "Error: missing required parameter: content. The 'write' tool schema requires: {path: string, content: string}"
3. This gives Claude the schema info needed to self-correct on the next attempt instead of retrying blindly