v0.1.27

Changed

• Bump claude-agent-sdk to >=0.1.40: Updated from 0.1.39 to latest SDK version

v0.1.24

Fixed

• Increase SDK buffer size to 5MB: Added max_buffer_size=5*1024*1024 to all ClaudeAgentOptions calls to prevent CLIJSONDecodeError on large files (SDK issue #98)
• Disable hooks globally: Set self.hooks = {} (empty dict, not None) to prevent "Stream closed" errors from hook callbacks in Claude Code v2.1.39+

v0.1.23

Refactored

• Simplified credential manager: Removed manual token refresh logic - now handled by Claude Agent SDK
  • Removed refresh_access_token() method (84 lines)
  • Removed _save_credentials() method (16 lines)
  • Removed httpx dependency from credentials module
  • Reduced from 140 to 99 lines (29% reduction)
  • Credential manager now only loads and validates credentials from disk
  • Token refresh is fully automated by the Claude SDK/binary
  • Updated documentation in CLAUDE.md to clarify SDK handles refresh
  • Removed obsolete test files: test_credential_manager_refresh.py, test_credential_manager_save.py
  • Updated integration tests to match new behavior
  • All 4,634 tests pass with 95.65% coverage on credentials.py

v0.1.22

Enhanced

• Test pattern extraction in coding-style.md: Coding style generation now discovers and documents test patterns

  • Finds E2E/integration test locations automatically (**/e2e/**/*.spec.ts, **/*.test.ts)
  • Extracts test naming conventions, run commands, and example files to follow
  • Critical for [debugging-qa] tasks - workers now know exactly where to write tests
  • Increased coding style limit from 600 to 800 words to accommodate test patterns
  • Planning prompt emphasizes test patterns from coding-style.md for debugging-qa tasks

• Improved [debugging-qa] guidance: Clarified debugging-qa workflow and emphasize automated tests

  • Explicit workflow: 1) Manual test to find issues → 2) Fix bugs → 3) Write integration tests → 4) Verify
  • Emphasizes both manual exploration (finds UX issues) AND automated tests (prevents regressions)
  • Context sublists now reference test patterns from coding-style.md
  • Examples show complete flow from manual testing to automated test code

v0.1.21

Changed

• Dependency updates: Updated all dependencies to latest stable versions
  • claude-agent-sdk: 0.1.28 → 0.1.35
  • typer: 0.21.1 → 0.22.0
  • fastapi: 0.128.2 → 0.128.7
  • hypothesis: 6.151.5 → 6.151.6
  • bcrypt: relaxed pin from <4.1.0 to <5.0.0 (now resolves to 4.3.0)
  • pydantic: bumped minimum to >=2.12.0
  • httpx: bumped minimum to >=0.28.0
  • uvicorn: bumped minimum to >=0.40.0
  • pytest: bumped minimum to >=9.0.0
  • pytest-cov: bumped minimum to >=7.0.0
  • pytest-asyncio: bumped minimum to >=1.3.0
  • mypy: bumped minimum to >=1.19.0
• Added requirements.txt with pinned versions for reproducible installs

v0.1.20

Added

• [debugging-qa] complexity level (#93): New task complexity for CI failures, bug tracing, visual QA, and log analysis
  • Routes to Sonnet with 1M context window for deep analysis of large logs and traces
  • New DEBUGGING_QA enum value in TaskComplexity
  • Updated planning prompts with [debugging-qa] tag documentation
  • Config support for debugging_qa model and context window settings
  • Comprehensive test coverage for new complexity level

v0.1.19

Fixed

• Stronger CI log warnings with explicit bad examples: Agents were still reading all log files despite v0.1.18 prompts
  • Added "⚠️ CRITICAL" warning in intro (first thing agents see)
  • Shows explicit BAD examples with ❌ marks of what NOT to do
  • Moved CI workflow above file list so it can't be missed
  • Explicit prohibition: "NEVER spawn tasks to read all logs"
  • Stronger language: "NEVER READ ALL FILES" instead of "DO NOT"
  • Warning now appears in 3 places: intro, file section, and examples

v0.1.18

Fixed

• More concise CI debugging prompts: Simplified workflow instructions to prevent agents from reading all log files
  • Added explicit ls step as first action to see file count before searching
  • Condensed CI debugging section from 20+ lines to 10 lines
  • Simplified PR review feedback workflow from 6 steps to 5 steps
  • More direct language emphasizing the correct workflow: ls → Grep → Read specific files
  • Same safety features and information, 50% less text to parse

v0.1.17

Fixed

• Context overflow from CI logs (#92): Prevent "Prompt is too long" fatal errors during CI debugging
  • Changed log splitting from line-based (500/file) to character-based (20KB/file)
  • Ensures predictable file sizes (~5,000 tokens) instead of unpredictable 30KB+ files
  • Added explicit warnings in work prompts: "⚠️ DO NOT read all log files at once"
  • Enhanced prompts with step-by-step Grep instructions and concrete examples
  • Prevents agents from reading all log files in parallel (which caused 443KB context overflow)
  • Updated tests for character-based splitting (29/29 pass, 98.26% coverage)

v0.1.16

Fixed

• Task timing always displayed: Fixed timing not showing for states created before v0.1.15
  • Always display timing even when task_start_time is None
  • Uses session_duration as fallback for backward compatibility
  • Ensures [claudetm HH:MM:SS] Task #N took X.X minutes always shown
• Enhanced CI log error reporting: Added detailed error logging for troubleshooting
  • Logs repository name and run ID before downloading
  • Shows job count when logs successfully downloaded
  • Includes full exception tracebacks for debugging
  • Helps identify gh CLI, API, and download failures