[Bug]: One-shot text generation (commit/PR/branch/title) ignores the provider's configured model and falls back to claude-haiku-4-5, which backend guardrails can block

Before submitting

• I searched existing issues and did not find a duplicate.
• I included enough detail to reproduce or investigate the problem.

Area

apps/server

Environment

• T3 Code 0.0.42 (linux-x64 runtime at ~/.t3/runtime/versions/0.0.42)
• Arch Linux (kernel 7.2.3-arch1-3)
• Provider: claudeAgent instance with binaryPath pointing at a wrapper that runs the Claude CLI through OpenRouter (ANTHROPIC_BASE_URL=https://openrouter.ai/api), custom model z-ai/glm-5.3-flash
• OpenRouter workspace guardrails block Anthropic models (only the GLM model is allowed)

Steps to reproduce

1. Use a claudeAgent provider whose backend refuses claude-haiku-4-5 (e.g. OpenRouter with workspace guardrails that exclude Anthropic models), with a working custom model configured for threads.
2. Leave textGenerationModelSelection unset (default settings).
3. Press Commit on any repo with staged changes.

I captured the exact spawn t3 issues (via a logging wrapper around the configured binary):

claude-ori -p --output-format json \
  --json-schema {"type":"object","properties":{"subject":{"type":"string"},"body":{"type":"string"},"branch":{"type":"string"}},...} \
  --model claude-haiku-4-5 \
  --settings {"disableAllHooks":true} --tools '' --disable-slash-commands \
  --strict-mcp-config --permission-mode dontAsk

→ exit 1. stdout contains the real failure:

{"api_error_status":400, "is_error":true, "result":"API Error: 400 0 endpoints out of 4 requested are available matching your guardrail restrictions and data policy. … Model blocked by guardrail: 4 endpoints excluded", ...}

stderr contains only an informational line from our credential wrapper ("Using the OpenRouter credential from the global credential ~/.ori/credentials.json.").

Expected behavior

Commit-message generation should use a model that is actually configured/available on the provider instance — e.g. the instance's defaultModelSelection / custom model — instead of a hardcoded fallback slug that the backend may refuse.

Actual behavior

Every one-shot text-generation operation fails: the Commit button shows

Text generation failed in generateCommitMessage: Claude CLI command failed: Using the OpenRouter credential from the global credential ~/.ori/credentials.json.

i.e. the stderr of the spawn (the wrapper's informational notice) is shown as the entire error, while the real API error sits in stdout and is discarded. Thread-title generation fails the same way (I observed ~36 failing title spawns over several hours, 0 successes) and generateBranchName / generatePrContent share the same path.

Root cause in the 0.0.42 bundle: with settings.textGenerationModelSelection unset, resolveTextGenerationProvider() → fallbackTextGenerationProvider() selects DEFAULT_TEXT_GENERATION_MODEL_BY_PROVIDER[claudeAgent] = "claude-haiku-4-5", ignoring the instance's configured default/custom model. A secondary problem is that runClaudeJson reports stderr when non-empty on non-zero exit, so any wrapper's stderr chatter masks the actual API error from stdout.

Workaround

Setting textGenerationModelSelection explicitly in ~/.t3/userdata/settings.json fixes it immediately (picked up live, no restart needed):

"textGenerationModelSelection": { "instanceId": "claudeAgent", "model": "z-ai/glm-5.3-flash" }

Suggested fixes

1. When textGenerationModelSelection is unset, fall back to the enabled provider instance's configured default model (e.g. defaultModelSelection/custom model) rather than a hardcoded slug; at minimum, skip fallback slugs that don't resolve in the instance's model catalog.
2. When the spawned CLI exits non-zero, include the stdout API error (or at least api_error_status/result from the JSON) in the surfaced error instead of preferring stderr, which may only contain wrapper notices.

Impact

Major degradation or frequent failure

Start in apps/server by tracing resolveTextGenerationProvider() and fallbackTextGenerationProvider(), then inspect runClaudeJson for non-zero exit handling. Verify that unset textGenerationModelSelection uses the enabled provider instance's configured model and that surfaced failures include the stdout API error rather than wrapper-only stderr; check commit, title, branch, and PR generation paths.