Skip to content

CHANGELOG.md

Latest commit

 

History

History
226 lines (183 loc) · 10.3 KB

CHANGELOG.md

File metadata and controls

226 lines (183 loc) · 10.3 KB

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]

Added

  • LLM pre/post hooks wired in agent command handlers for multi-agent coordination (Refs #451)
  • Self-Documentation API exposed via robot CLI subcommand (Refs #1011)
  • ForgivingParser integrated into CLI command dispatch for typo-tolerant command parsing (Refs #1012)
  • MS Teams SDK test suite with comprehensive SDK tests (Refs #1034)
  • Tantivy index for session search with BM25 ranking (Refs #1039)
  • Token budget flags wired to Search command for robot mode output control (Refs #672)
  • JSON format support on roles/config/graph commands in robot mode (Refs #1013)
  • ADF operations guide and blog post for PR fan-out deployment
  • ADF agent fleet reference documentation

Fixed

  • RUSTSEC-2026-0049 eliminated by switching serenity to native-tls (Refs #418)
  • Spec gaps addressed and resolved across ADF orchestrator templates (Refs #1040)
  • Global concurrency limits enforced in orchestrator to prevent task/memory exhaustion (Refs #664)
  • listen_mode test assertion updated to match clap error output (Refs #1044)
  • Robot response formatting corrected in chaining logic
  • MCP latency benchmarks gated behind release builds to prevent debug flake (Refs #672, Refs #987)
  • Pagination and token budget wired into search response with test alignments (Refs #672)
  • Performance benchmark thresholds raised and redundant attributes removed (Refs #987)

Changed

  • Robot mode now honours --format json for consistent machine-readable output
  • Orchestrator uses KG-routed model in Quickwit logs and AgentRunRecord
  • CLI exit codes aligned with F1.2 contract (typed ExitCode, not bare exit(1)) (Refs #860)
  • Learning store implemented on SharedLearningStore with markdown backend
  • Token budget management engine active for robot mode output control

[1.17.0] - 2026-04-27

Added

ADF Orchestrator Framework

  • New terraphim_orchestrator crate (v1.8.0) -- Autonomous Development Flow orchestrator
  • Multi-agent coordination with mention-chain support for Gitea
  • PR lifecycle management -- fan-out to build-runner + pr-reviewer on PR open
  • Push webhook handling with build-runner spawning
  • Compound-RICE scoring with Themis persona for issue prioritisation
  • Runtime guardian, upstream synchronizer, and meta-learning agent templates

Session Connectors

  • OpenCode and Codex JSONL session connectors for real-time session ingestion
  • ClineConnector for VS Code extension sessions
  • SessionConnector::watch() for live session monitoring
  • BM25-ranked session search with search-index feature

Learning Store & Knowledge Graph

  • SharedLearningStore with graph hybrid scoring
  • LearningInjector and learn inject CLI command
  • Knowledge graph validation workflows for pre/post-LLM quality gates
  • NormalizedTerm metadata and learn export-kg command

CI Infrastructure

  • Firecracker-accelerated CI with VM lifecycle management
  • Shared Rust build cache via sccache + SeaweedFS S3 on fcbr0
  • Bigbox runner support with rch dispatch for build queue management
  • Zig 0.16.0 compilation for zlob feature (SIMD glob matching)

Robot Mode & Auto-Routing

  • Token budget management engine for robot mode output
  • Auto-route on search when --role is unset (thesaurus-scored)
  • RobotFormatter wired into CLI search with JSON/JSONL/table output
  • ForgivingParser integrated into REPL for typo-tolerant commands

Fixed

  • zlob compilation -- Zig 0.16.0 build with Darwin linker workaround
  • Security advisories -- RUSTSEC-2024-0421, 2026-0098/0099/0104 resolved
  • REPL role selection -- reads from service instead of hardcoded "Default"
  • CLI exit codes -- listen --server exits with ERROR_USAGE (2) not 1
  • Graceful degradation for roles without KG
  • Flaky tests -- removed data-dependent assertions, fixed timing issues

Changed

  • Workspace version bumped to 1.17.0
  • Default features -- zlob enabled by default on terraphim_file_search and terraphim_mcp_server
  • CI workflow -- migrated to bigbox + SeaweedFS + rch dispatch
  • Dependency updates -- removed dead trust-dns-resolver, ring deps

[1.14.0] - 2026-03-22

Added

Terraphim Symphony Orchestrator

  • New terraphim_symphony crate for automated multi-agent issue dispatch
  • Claude Code runner with full streaming protocol support
  • Dual-mode orchestration combining issue-based and time-based scheduling
  • Workspace management with git operations and context tracking
  • Gitea and Linear tracker integrations for issue lifecycle management

Session File Tracking

  • File access extraction from agent sessions (extract_file_accesses)
  • Files and ByFile subcommands for session file queries
  • Sessions service methods for querying by file path

Agent Workflows E2E

  • Complete workflow implementation with 5 workflow templates
  • Real API integration with Cerebras support
  • Playwright browser tests for all workflows
  • Quality gate integration with merge-to-main workflow

Orchestrator Improvements

  • Proportional fairness scheduling algorithm
  • Dual-mode run loop combining time-based and issue-based dispatch
  • Dependency-aware dispatch with topological sorting
  • PageRank-aware issue sorting for prioritization

Validation Framework

  • Phase 4/5 disciplined verification and validation framework
  • Requirements traceability matrix (REQ -> design -> code -> test)
  • Quality gate reports for release readiness

Fixed

  • axum-test 19.x API changes - Updated test files
  • atty deprecation - Replaced with std::io::IsTerminal
  • Symphony hook timeouts - Increased for long-running operations
  • Liquid template issues - Moved to heredoc approach

Changed

  • TLA+ TypeScript bindings research and design completed
  • Dependency updates for axum-test, env_logger, actions/github-script

1.8.0 - 2026-02-16

Added

Learning Capture System (Complete Implementation)

  • Core capture logic (crates/terraphim_agent/src/learnings/):

    • CapturedLearning struct with full context (command, error output, exit code, timestamp)
    • capture_failed_command() with automatic secret redaction via regex
    • Chained command parsing (cmd1 && cmd2 || cmd3)
    • Markdown serialization/deserialization for persistent storage
    • list_learnings() - list recent learnings with [P]/[G] source indicators
    • query_learnings() - search by pattern (substring or exact match)
    • Hybrid storage model: project-specific (.terraphim/learnings/) with global fallback (~/.local/share/terraphim/learnings/)
    • 15 unit tests covering all functionality

  • CLI Integration:

    • terraphim-agent learn capture <command> --error <msg> [--exit-code N]
    • terraphim-agent learn list [--recent N] [--global]
    • terraphim-agent learn query <pattern> [--exact] [--global]
    • terraphim-agent learn correct <id> --correction <text> (placeholder)

  • Hook Integration:

    • .claude/hooks/learning-capture.sh - PostToolUse hook for automatic capture
    • Native Hook Support (terraphim-agent learn hook):
      • Process hook input directly without bash wrapper
      • terraphim-agent learn hook [--format=claude] - Reads JSON from stdin
      • terraphim-agent learn install-hook <claude|codex|opencode> - One-command setup
      • No external dependencies (no jq required)
      • 156 tests covering hook functionality
    • Automatic capture of failed Bash commands
    • Secret redaction before storage
    • Fail-open design (continues even if capture fails)
    • Environment variable TERRAPHIM_LEARN_DEBUG for debug output

  • Documentation:

    • docs/src/kg/learnings-system.md - User guide with architecture, usage, troubleshooting
    • skills/learning-capture/skill.md - AI agent skill reference
    • docs/verification/learning-capture-verification-report.md - Phase 4 verification
    • docs/validation/learning-capture-validation-report.md - Phase 5 validation

Guard Thesaurus-Driven Matching

  • Replaced 12 static regex patterns with terraphim's Aho-Corasick engine
  • JSON thesaurus files for extensible pattern matching:
    • guard_destructive.json: 32 entries across 13 concept categories
    • guard_allowlist.json: 10 entries across 5 concept categories
  • Coverage for 20+ previously unblocked destructive commands:
    • rmdir, chmod, chown, shred, truncate, dd, mkfs, fdisk
    • git commit --no-verify, git push --no-verify
    • git reset --merge, git restore --worktree, git stash clear
    • git branch -D, rm -fr
  • CLI flags for custom patterns: --guard-thesaurus, --guard-allowlist
  • 36 tests for guard patterns

Performance Improvements

  • HTTP Client: Connection pooling with global static clients
    • 30-50% reduction in connection overhead
    • Connection pooling (10 idle per host, 90s timeout)
    • Functions: get_default_client(), get_api_client(), get_scraping_client()
  • Summarization Worker:
    • Use VecDeque for O(1) removal (was O(n) with Vec)
    • Implemented retry re-queuing with exponential backoff
    • Removed unused active_workers counter

Documentation

  • crates/terraphim_agent/DESIGN-guard-patterns-redesign.md - Guard pattern redesign document
  • Graph embeddings learnings examples
  • Learning via negativa examples for command corrections
  • Validation reports for agent, CLI, and tinyclaw crates

Changed

  • OpenRouter: Removed unreachable truncation logic (dead code elimination)
  • Hooks: Avoid false positives in secret detection

Fixed

  • Fixed regex patterns in learning capture redaction to handle dashes in API keys
  • Workspace: Exclude terraphim_automata_py (PyO3 extension - use maturin develop)

1.7.0 - 2026-02-10

Added

  • Initial learning capture specification
  • Guard patterns for command protection
  • HTTP client improvements

[1.6.0] - 2026-01-XX

Added

  • Base implementation of terraphim-agent
  • RoleGraph functionality
  • Knowledge graph ranking