- Exec Subcommand:
hawk exec "prompt"— full engine non-interactive mode with--output-format json,--autoautonomy levels,--worktreeisolation,--agentpersonas,--session-idresume, stdin piping - Daemon Server:
hawk daemon start/stop/status— background HTTP server with JSON + SSE streaming on/v1/chat,/v1/health,/v1/sessions - Mission Mode:
hawk mission "prompt"— multi-agent orchestration decomposing work into parallel features executed in isolated git worktrees.--dry-runfor planning only - Session Search:
hawk search "query"— full-text search across all saved sessions with--jsonoutput - Custom Agents:
hawk agent list/create/show/remove— markdown persona definitions in~/.hawk/agents/with YAML frontmatter (name, description, model) - Snapshot System: Shadow git tracking of every file change.
hawk snapshot list/restore/diff+/snapshotslash command. Auto-snapshots on every Write/Edit tool call - Waza Workflows:
/think(plan before code),/hunt(root-cause diagnosis),/check(pre-ship review with auto-fix),/design(screenshot-driven UI iteration) - Structured Compaction: Summary template with Goal/Constraints/Progress/Files/Decisions/Errors/Instructions/Next sections for better intent preservation
- Doom Loop Detection: Lowered threshold to 3 (from 4). Two-tier escalation: first detection injects redirect prompt, doom loop hard-stops with "ask user for help"
- Session Persistence for exec: All exec runs saved to
~/.hawk/sessions/and searchable viahawk search
mission/— Multi-agent orchestration with worktree-based parallel workersdaemon/— HTTP server with session factory, SSE streaming, PID file managementagents/— Markdown persona loader with frontmatter parsingsnapshot/— Shadow git repository for granular file-level undo
- Factory Droid v0.117.0 (exec mode, daemon, mission orchestration, custom agents)
- OpenCode (structured compaction, snapshot system, doom loop escalation)
- Waza by tw93 (engineering-habit workflows: think, hunt, check, design)
- Model Cascade Router: Cost-aware routing that classifies prompts and selects optimal model tier (simple→Haiku, debug→Sonnet, generation→Opus). Supports frugal mode for aggressive cost savings. Tracks routing decisions for analytics.
- Dynamic max_tokens: Adaptive output budgets based on task type and recent tool-call patterns. Reduces output token costs 15-25% by not over-allocating.
- Cheap Compaction Model: Conversation summaries now use the cheapest available model (Haiku/gpt-4o-mini) instead of the primary model. Saves $0.10-0.50 per compaction.
- Context Budget Allocator: Formal token allocation across system prompt, tool defs, repo map, memory, workspace, pre-loaded files, and conversation. Adaptive: shrinks file budget as conversation grows. Triggers compaction at threshold.
- LLM Reflection Engine: Verbal self-reflection after failed attempts (Reflexion pattern). Asks "what failed, why, what to do differently" instead of mechanical summaries. Accumulates episodic memory buffer.
- Self-Review Before Write: Rubber duck debugging step between code generation and file write. Model explains its code and checks for bugs/regressions before applying.
- Session Lifecycle (Self-Improvement Loop): Closed loop wiring OnSessionStart (retrieve guidelines + skills) and OnSessionEnd (learn guidelines, distill skills, record cost).
- Import/Dependency Graph: Parses import statements for Go, Python, TypeScript. Builds forward/reverse edges. DependenciesOf, DependentsOf, ImpactSet with BFS depth control.
- Change-Set Aware Context: Loads only code relevant to current
git diff. 70-90% context reduction for focused tasks. FormatContext with token budgeting. - Landlock Sandbox (Linux): Zero-dependency, zero-overhead, unprivileged filesystem isolation. Restricts agent to project dir + /tmp. Default Linux sandbox.
- seccomp-bpf Syscall Filtering (Linux): Blocks 21 dangerous syscalls (mount, ptrace, reboot, kexec_load, init_module, bpf, etc.). Applied via SysProcAttr.
generateSummary()now uses cheapest available model per provider instead of primary model- Ecosystem roadmap added:
ECOSYSTEM-ROADMAP.mdwith 30-feature prioritized implementation plan
- Bash Security: zsh bypass protection, process substitution blocking, IFS injection detection, carriage return prevention, ANSI-C quoting detection, git commit safety
- Hook System: 8 event types with priority-based execution (pre_query, post_query, pre_tool, post_tool, session_start, session_end, permission_ask, error)
- Plugin System: manifest validation, install/list/uninstall, hook registration, command execution
- Advanced Permissions: auto-mode learning, command classifier, bypass killswitch, shadowed rule detection
- Model Catalog: 25+ models across 7 providers with pricing and context sizes
- Session Memory: extraction, search, and consolidation
- Analytics: event logging, session traces, cost tracking
- Auth System: secure token storage with OS keychain integration
- Auto-Update: GitHub release checking
- LSP Integration: JSON-RPC client and server manager
- Voice Mode: Whisper.cpp integration
- Magic Docs: Go AST parsing and automatic markdown generation
- Worktree Tools: EnterWorktree/ExitWorktree with validation
- Retry Package: exponential backoff for API resilience
- Circuit Breaker: three-state circuit breaker for fault tolerance
- Rate Limiter: token bucket algorithm
- Logger: structured logging with levels
- Health Checks: registry with status aggregation
- Metrics: counters, gauges, timers with atomic operations
- Graceful Shutdown: signal-based shutdown with hooks
- Profiling: CPU, memory, goroutine profiling
- Tracing: distributed tracing spans
- Config Validation: field-level validation errors
- Benchmarks & Fuzz Tests: bash security parsing
- Shell Completion: bash, zsh, fish, powershell
- Docker: multi-stage build with non-root user
- Nix Flake: reproducible builds
- GitHub Actions: CI with test, build, lint, coverage; release with GoReleaser
- Improved error messages with context wrapping
- JSONL session storage with legacy JSON fallback
- Stream-JSON usage events with token tracking
- Pre-compiled regexes for performance
- Project scaffold with cobra CLI and Bubbletea TUI
- Interactive chat REPL with textarea input, spinner, lipgloss styling
- eyrie wired as LLM provider dependency
- GitHub Actions CI