feat: auto-detect API keys and default model for zero-config startup

[Nivesh353]

Problem

gitagent required an explicit model configuration (model.preferred in
agent.yaml, or --model on the CLI) to start — even when a valid
provider API key was already present in the environment. Users had to
manually configure a model before gitagent would run at all.

Fix

New file: src/auto-detect-model.ts
Checks known providers in a fixed order and returns a sensible default
model for the first one with usable credentials, using pi-ai's
getEnvApiKey() rather than a hardcoded env-var check — so ambient
credential sources are picked up too, not just direct API keys.

Providers checked, in order:

1. Direct API-key providers (explicit, deliberate signal): anthropic,
   openai, google, xai, groq, mistral
2. Ambient/enterprise credential sources (checked after, since these can
   be present incidentally): amazon-bedrock, google-vertex,
   azure-openai-responses, github-copilot

All defaults use versionless model aliases (e.g. claude-sonnet-4-6,
gpt-4o) rather than dated model IDs, so they don't go stale as new
point releases ship.

src/loader.ts

• Wires autoDetectModel() in as the final fallback in model resolution,
  preserving existing priority: --model flag > agent.yaml
model.preferred > env config override > auto-detected from API key.
• Logs the choice on auto-detect: Using <model> (auto-detected from <provider> credentials).
• Updated the "no model configured" error to mention the new option.
• Related fix: manifest.model is now normalized to { fallback: [] }
  right after parsing agent.yaml, and model.preferred is now correctly
  typed as optional. Previously, a model: block missing entirely (not
  just missing preferred) crashed downstream code (index.ts, sdk.ts)
  with Cannot read properties of undefined (reading 'constraints') —
  masked before this change because a missing model always hit the "no
  model configured" error first; auto-detection now gets further and
  exposed it.

.env file loading was already implemented separately (src/index.ts
loads .env from ~/.gitagent/.env and the agent directory before model
resolution runs), so no changes were needed there.

Testing

• npm run build — compiles cleanly, no type errors.
• npm test — all 27 existing tests pass, no regressions.
• Verified live, end-to-end, against a real agent directory:
  • No model configured + a real provider key present → correctly
    auto-detects and logs the choice, proceeds to a real API call.
  • No model configured + no keys at all → correct friendly error,
    no crash.
  • No model: block in agent.yaml at all → previously crashed with
    Cannot read properties of undefined (reading 'constraints'); now
    resolves correctly via auto-detect.

Closes #14

[shreyas-lyzr]

Clean implementation of zero-config model auto-detection. The fallback chain is well-ordered, the normalization fix for manifest.model is correct and needed, and the build + all 27 tests pass.

One thing worth knowing: getEnvApiKey('github-copilot') in pi-ai resolves against COPILOT_GITHUB_TOKEN, GH_TOKEN, and GITHUB_TOKEN. Since gitagent itself requires a GitHub token, github-copilot will match on any machine where no other LLM provider key is set. The PR already places it last in the priority list, so it only fires as a last resort — but it means a user with only GITHUB_TOKEN set and no actual Copilot access will see github-copilot auto-selected and then get an auth error at call time rather than a clean config error. Worth a note in the README or error message so it's not confusing.

Minor: claude-sonnet-4-6 (dash) for anthropic/bedrock vs claude-sonnet-4.6 (dot) for github-copilot — looked correct against the pi-ai model registry, so not a bug.

No tests for autoDetectModel() itself — the function is simple and pure enough that the risk is low, but a few unit tests covering key detection order and the undefined-when-no-keys path would be a useful addition.