Opus 4.7 is now the default paid model - Fallback chain extended: 4.7 → 4.6 → 4.5 → Sonnet 4.6 → 4.5 - Strict tool use guard removed — all models support it natively

- Agent now recognizes AWS permission errors as infrastructure issues instead of code bugs - Added `date` command to agent's shell toolkit for debugging time-dependent tests

New: git_diff tool for on-demand diffs - Agent now fetches file diffs only when needed instead of carrying the full patch in every message - Large patches are truncated before entering the LLM conversation to keep costs proportional to PR size

Auto-merge bug fix: required check comparison now uses app_id - Fixed a bug where auto-merge compared app names against check context strings, which never matched - Required status checks now correctly matched by app_id from branch protection API - Added 50 tests covering r...

Refactored message optimization into single-responsibility modules - Split 338-line monolith into 6 focused files with proper SDK types - Consolidated duplicate diff extraction functions into one with a marker table - Added pre-commit hooks blocking cast(), private functions, ...

Quality gate now grades tests on case-matrix completeness - New case_coverage category checks whether tests enumerate independent input dimensions and assert exact expected values per cell - Verified the grader distinguishes a 1-case test from a full parametrized matrix on the...

Logger coverage is now enforced automatically - New lint script checks that every return/continue/break has a preceding logger call and every if/elif/else branch starts with one - Wired into pre-commit so violations block the commit with the exact lines to fix - Claude Code no...

Agent now detects a concurrent push to a PR branch and stops cleanly - git_commit_and_push returns a typed result with concurrent_push_detected on non-fast-forward push - Tool wrappers propagate the flag up through AgentResult; agent loop breaks the turn instead of retrying on...

Rate-limit 429s from Google/Anthropic/GitHub now retry cleanly instead of crashing the Lambda - New helper pulls the retry-after delay from any SDK's error shape (Google's message body, Anthropic's retry-after header, GitHub's X-RateLimit-Reset) - handle_exceptions sleeps the ...

Google 499 CANCELLED errors now retry instead of bubbling to Sentry - is_transient_error recognizes Gemini's server-side stream cancellations as retryable - Linear backoff via the existing transient-retry budget (3 attempts, 2s/4s/6s), no new config - Paired with the 429 retry...

Infrastructure-failure classifier no longer trips on app-level console warnings - Console.warn/log blocks stripped before scanning, so app warnings (failed SSM fetches, etc.) can't false-positive the classifier - Regression test uses a real captured CircleCI log that previousl...

Test-coverage PRs on mongodb-memory-server repos now run. Lambda was downloading the amazon2 MongoDB binary, crashing on missing libcrypto.so.10. Agent now picks rhel90 for MongoDB 6.x and reads binary.version from jest-mongodb-config.js.

GitAuto's audit log now covers 100% of model calls. Every Claude Opus 4.7, OpenAI, and Gemini call records to llm_requests against a real user_request id from the first token. Claude structured outputs moved off beta onto the GA output_config=. System prompt persists alongside...

GitAuto's agent loop now keeps the model's reasoning text when pruning duplicate tool calls from history. Stripping whole turns made Claude Opus 4.7 forget its plan and call the same tool again, burning roughly half the per-task budget on duplicate work.

GitAuto's agent now sees `[log source: OUR ...]` or `[log source: CUSTOMER ...]` on every error log. The OUR tag tells the agent the failure is in GitAuto's runtime and to keep doing the customer's task instead of writing infrastructure workarounds into the customer's repo.

GitAuto now starts at $2/PR with Google Gemma 4. Same $24 free credits, up to 12 PRs instead of 3. Claude Opus 4.7 still available at $8/PR for teams that need maximum capability.

GitAuto evaluates test quality across 44 checks in 9 categories: integration, business logic, adversarial, security, performance, memory, error handling, accessibility, and SEO. New docs page details every check.

Your Prompt Is Not Enough. Add a Gate. gitauto.ai/blog/prompts-n…

Some devs run lint and type-check before pushing. Some don't. Pre-commit fixed that gap by putting the gates in code, not in habits. AI agents reopened the same gap at a new layer. GitAuto closes it the same way: phpcs, phpcbf, phpstan run from the orchestrator on every PHP PR.

For teams debating whether to let your agent run free-form shell commands as a fallback (dedicated tools handle file read/edit/git/search; shell is the long-tail catch-all): here's what months of monitoring told us. Every block was complicated-but-safe diagnostic work. We drop...