Skip to content

CHANGELOG.md

Latest commit

 

History

History
321 lines (219 loc) · 11.4 KB

CHANGELOG.md

File metadata and controls

321 lines (219 loc) · 11.4 KB
Raw

Changelog

All notable changes to Codi are documented in this file.

[0.17.0] - 2026-01-26

Breaking Changes

  • Ollama Cloud Provider Removed: The ollama-cloud provider has been removed. Use --provider ollama with OLLAMA_HOST=https://ollama.com instead. The regular ollama provider uses the OpenAI-compatible API which works correctly with Ollama Cloud.

  • Web Search Tool Migration: Replaced WebSearchTool with EnhancedWebSearchTool

    • Tool name remains web_search for backward compatibility
    • Supports Brave API (primary), Google Custom Search, Bing API, and DuckDuckGo fallback

Features

  • Workflow System Phase 8 - Production Ready (#173):

    • AI-assisted workflow building with natural language
    • Multi-step pipelines with variable substitution
    • Git and PR action steps (commit, push, create-pr, review-pr)
    • Comprehensive test coverage (E2E and unit tests)
    • Template system for common workflow patterns

  • Context Debug Command (#179):

    • New /compact debug subcommand for inspecting context window state
    • View message counts and token estimates
    • Analyze working set and indexed files
    • Debug context compaction behavior

  • Enhanced Web Search - Phase 2 (#165, #170):

    • Multi-engine support with automatic fallback
    • Search templates for docs, pricing, and error queries
    • Relevance scoring algorithm (domain, content match, quality)
    • Per-engine rate limiting with graceful degradation
    • Template-aware caching TTLs

  • Memory Monitoring (#167):

    • Proactive context compaction based on memory pressure
    • Automatic cleanup when approaching limits

  • Symbol Index Multi-Language Extension (#172):

    • Improved symbol extraction across languages
    • Better TypeScript/JavaScript support

Improvements

  • Technical Debt Cleanup (#181):

    • Extracted ~400 lines from index.ts into src/cli/ modules
    • Added discriminated union types for workflow steps
    • Added 9 type guards for type-safe step handling
    • Reduced as any usage from 30+ to 10 instances

  • Debug Logging for Error Handlers (#185):

    • Added logger.debug() calls to 20 previously silent catch blocks
    • Enables troubleshooting with --debug flag
    • Files: memory.ts, session.ts, history.ts, usage.ts, agent.ts, diff.ts, spinner.ts

  • Plugin System Investigation (#181):

    • Created comprehensive docs/PLUGIN-INVESTIGATION.md
    • Security analysis and recommendations
    • Phased re-enablement roadmap

Bug Fixes

  • UI Freeze During Compaction (#174): Fixed UI becoming unresponsive during context compaction
  • Workflow E2E Tests (#180): Resolved flaky PR review tests with improved mock agent
  • Model Display Updates (#175): Fixed model display not updating when provider changes during workflow
  • IPC Disconnect Race (#164): Graceful IPC disconnect prevents race condition in orchestration
  • Ink UI Stability (#163): Added tool call display and improved visual stability

Tests

  • 2151 tests passing
  • Added workflow E2E tests for PR review workflow
  • Added rate limiting tests for enhanced web search
  • Improved test stability with buffer flush helpers

[0.13.0] - 2026-01-18

Features

  • GitHub Pages Enhancement: Complete redesign of documentation site

    • Added demo GIF with terminal window styling and animated gradient border
    • Added badges for version, license, Node requirement, and default model
    • Added Tools section showing all 12 built-in tools
    • Added Usage & Models command section
    • Visual improvements: animations, glow effects, smooth scrolling
    • Open Graph meta tags for better social sharing
    • Improved mobile responsiveness

  • Default Ollama Model: Changed default from llama3.2 to glm-4.7:cloud

    • Applies to both local Ollama and Ollama Cloud providers

Configuration

  • Updated codi-models.yaml with current defaults:
    • Opus updated to claude-opus-4-5-20251101
    • Added gpt-5 and gpt-5-nano for OpenAI
    • Renamed llama3 to glm using glm-4.7:cloud
    • Fallback chain now prioritizes opus → sonnet → haiku → gpt-5 → glm

[0.12.0] - 2026-01-18

Features

  • Claude Opus 4.5 as Default: Changed default Anthropic model from Claude Sonnet 4 to Claude Opus 4.5 (claude-opus-4-5-20251101)
    • Added Opus 4.5 pricing to usage tracking ($15/$75 per 1M tokens)
    • Added Opus 4.5 to static model registry

[0.11.0] - 2026-01-18

Documentation

  • Development Process: Updated CODI.md to require tests before merging PRs
    • AI agents must run pnpm build && pnpm test before merging
    • Added self-review template with build/test status
    • Ensures code quality and prevents regressions

Bug Fixes

  • Fixed getAllCommands() function call in index.ts
  • Fixed IndexErrorCallback to expect Error objects instead of strings

[0.10.0] - 2026-01-18

Features

  • Dynamic Context Configuration: Tier-based context settings
  • README Accuracy Update: Comprehensive documentation review
    • Added 8 missing CLI options
    • Documented /init command
    • Added missing command sections (Usage, Planning, RAG, Approvals)
    • Expanded .codi.json example with all config options

[0.9.1] - 2026-01-18

Bug Fixes

  • JSON Parser: Fixed tool call parsing for models that output raw newlines in JSON strings

    • Models like glm-4.7 via Ollama Cloud can now correctly parse multiline commands
    • Added escapeNewlinesInStrings() preprocessing before JSON parsing

  • Bracketed Paste: Rewritten paste handling to capture and display summary

    • Now shows [pasted N lines, M chars] like Claude Code

  • Documentation: Fixed README and CODI.md discrepancies

    • Added ollama-cloud to CLI provider options
    • Updated OpenAI models to include GPT-5

[0.9.0] - 2026-01-17

Command Consolidation

Commands are now organized under /git and /code prefixes for better discoverability:

Git Commands → /git <action>

  • /git commit, /git branch, /git diff, /git pr, /git stash
  • /git log, /git status, /git undo, /git merge, /git rebase

Code Commands → /code <action>

  • /code refactor, /code fix, /code test, /code doc, /code optimize

Backward Compatibility: Popular commands still work as standalone aliases:

  • /commit, /branch, /pr (git)
  • /refactor, /fix, /test (code)

Documentation

  • Added CODI.md as the main documentation file (CLAUDE.md symlinked for AI compatibility)
  • Added ollama-cloud provider to docs and README
  • Synced all commands between README and GitHub Pages site
  • Added demo GIF to README

Other Changes

  • Updated branding to "AI coding wingman"
  • Set up GitHub Pages documentation site

[0.8.6] - 2026-01-17

Bug Fixes

  • Fixed arrow key navigation - Resolved issue where arrow keys showed escape sequences (^[[A) after commands completed. Root cause was ora spinner's discardStdin setting interfering with readline.

Improvements

  • Added spinner for /compact command - Visual feedback while context compaction runs

Documentation

  • Added git worktree documentation - CLAUDE.md now includes guidance for working with multiple worktrees

Removed

  • Removed interactive command picker - Feature was causing terminal handling conflicts; may revisit in future

[0.8.4] - 2026-01-16

New /codi Command

Generate comprehensive AI context files automatically:

/codi generate   # Analyze codebase and create CODI.md
/codi show       # Display current context file
/codi edit       # Get AI help to improve it

Auto-Generated CODI.md Includes

  • Project type, language, and framework detection
  • Directory structure (2 levels)
  • Package.json scripts as quick reference
  • Key files with purposes
  • Top dependencies with descriptions
  • Test framework detection
  • Coding conventions based on language

Updated /init Command

/init              # Now creates .codi.json, codi-models.yaml, AND CODI.md
/init --context    # Create only CODI.md template

Test Reliability

  • Added retry for flaky e2e tests
  • Disabled file parallelism to avoid resource contention
  • Increased default test timeout

Other Improvements

  • Bun lockfile detection for package manager
  • Improved pytest detection (checks pyproject.toml content)
  • Performance: cached stat results in directory walk

[0.8.3] - 2026-01-16

Adaptive Context Window Calculation

Replace fixed 40% context allocation with adaptive calculation that maximizes available context:

Model Context Window Old Limit (40%) New Adaptive Limit
Claude (200k) 200,000 80,000 ~189,000
GPT-4o (128k) 128,000 51,200 ~117,000
GPT-4 base (8k) 8,192 3,276 2,457 (30% floor)

Formula: contextWindow - systemPrompt - tools - outputReserve - buffer

  • Falls back to 30% minimum when overhead exceeds available space
  • Warns when message budget is below 5k tokens
  • New constants: MAX_OUTPUT_TOKENS, CONTEXT_SAFETY_BUFFER, MIN_CONTEXT_PERCENT, MIN_VIABLE_CONTEXT

Documentation

  • Added Git Workflow guidelines to CLAUDE.md and CONTRIBUTING.md
  • Established branch/PR workflow (never push directly to main)

[0.8.2] - 2026-01-16

Token Estimation Improvements

  • Content-aware estimation: Different ratios for prose (~4 chars/token), code (~3), and JSON (~3.5)
  • Complete context accounting: System prompt and tool definitions now included in token counts
  • Calibration from API responses: Exponential moving average improves accuracy over time

New /compact Command

  • Proper slash command (converted from built-in handler)
  • Alias: /summarize
  • Shows before/after token counts

Code Detection Optimizations

  • Pre-compiled combined regex for single-pass matching
  • Minimum length guard (15 chars) skips short text
  • More specific patterns reduce false positives

[0.8.1] - 2026-01-16

Bug Fixes

  • Tool parsing: Fixed nested parentheses in quoted strings (e.g., response.cookies.get('token')) being truncated
  • Tool extraction: Fixed tool calls not being detected when model sends all content via "thinking" field

The regex-based tool extraction [^)]* stopped at the first ) even inside quoted strings. Replaced with a state machine that properly tracks string context and escape sequences.

[0.8.0] - 2026-01-16

Dynamic Context Window

  • Context limit now adapts to the model being used (40% of model's context window)
    • Claude 200k → 80k threshold
    • GPT-4o 128k → 51k threshold
    • GPT-4 base 8k → 3.2k threshold
  • Context limit automatically recalculates when switching providers mid-session

Compression Improvements

  • Disabled by default - was confusing models that output E1/E2 symbols
  • Enable with --compress flag when needed
  • Only applies compression if it actually saves space (including legend overhead)
  • Output is now decompressed so you never see raw entity references

RAG-Enhanced Compaction

  • Code relevance scoring - messages discussing indexed files get higher importance during compaction
  • File context in summaries - extracted file paths included in summarization prompts
  • Semantic deduplication - similar messages grouped together using embeddings (when RAG enabled)

Code Quality

  • Vector utilities extracted to src/utils/vector.ts for reusability
  • Fixed model matching to prevent "gpt-4" incorrectly matching "gpt-4o"
  • Static imports for better performance

For earlier versions, see the GitHub Releases page.