Skip to content

Latest commit

 

History

History
88 lines (70 loc) · 6.39 KB

File metadata and controls

88 lines (70 loc) · 6.39 KB

Dirgha CLI v1.18.0 — SOTA Scorecard

Version: v1.18.0 Date: 2026-05-02 Tests: 104/104 passing (13 files) Previous: v1.17.2 → v1.18.0

Competitive Landscape

Feature OpenCode Gemini CLI Claw Code Hermes v1.18.0
Ink TUI
Thinking expand/collapse
@-mentions file picker
Tool diff coloring
Streaming tool args
/keys BYOK mid-session
Provider health scoring
Model family fallback
Multi-tier failover chain
Adaptive retry windows
17 providers
Agent team templates
Permission three-tier
Auto-memory extraction
Memory auto-recall
Mid-loop compaction
LSP (def/ref/hover/diag)
Session export/import
Doctor command
Stall detection
Fleet DAG workflows
Per-tool timeout
Event backpressure
Config schema versioning
Crash log rotation
Daemon graceful shutdown
SBOM + cosign signing

v1.18.0 Final Scores (VERIFIED May 2, 2026)

Dimension v1.17.2 v1.18.0 Delta Status
TUI Quality 9.5 9.5
Tool System 9.5 9.8 +0.3
Agent Engine 9.0 9.8 +0.8
Provider Model 9.5 9.8 +0.3
Developer UX 9.5 9.5
Platform 8.5 9.5 +1.0
Stability 9.5 new
Security 9.5 new
Overall 9.3 9.6 +0.3

Dimension Breakdown

TUI Quality (9.5): Ink rendering with thinking expand/collapse, white thinking text, Enter-after-paste fix, EPIPE/EIO suppressed at render layer.

Tool System (9.8): Per-tool timeout enforcement (Promise.race), shell 300s default, atomic cron writes, browser connection check, git cwd separator fix, shell close vs exit event.

Agent Engine (9.8): Event-stream recursion guard, backpressure emission on queue overflow, drain() API, contextTransform isolated from provider errors, maxTurns clamping, compaction summarizer failure handled gracefully, session state durability.

Provider Model (9.8): 4-tier failover cascade (user → same-family → registry → free), health compaction with 60s TTL, health score NaN/latency/cost fixes, stall detection on hung streams (30s), DeepSeek-native prefix stripping, OpenRouter thinking pattern fix.

Developer UX (9.5): Config schema versioning with migration hook, silent config loss warning, slash dispatch error handling, fleet stdout monkey-patch scoped.

Platform (9.5): Daemon graceful shutdown (AbortController, 10s deadline, session flush), Fleet DAG workflows (agent-to-agent chaining), crash log rotation (10MB, 200 entries), SBOM + cosign keyless signing.

Stability (9.5): 104/104 tests, 0 TS errors, 0 ESLint warnings, 0 new crashes since EPIPE/EIO guards, all 30+ CLI commands smoke-tested.

Security (9.5): Shell injection fixed (hooks), path traversal bypass fixed (policy), seatbelt profile injection blocked, PowerShell command injection fixed (/paste), token redaction, config injection guard.

What we fixed in v1.18

  • Kernel: 5 HIGH fixes (event-stream recursion, contextTransform isolation, ToolResultMessage role, assembleTurn robustness, maxTurns clamping)
  • Providers: 7 HIGH fixes (failover cascade, stall detection, health compaction, DeepSeek prefix, OpenRouter thinking, OpenAI-compat try/finally, synthetic tool IDs)
  • Security: 4 HIGH fixes (shell injection, path traversal, seatbelt injection, PowerShell injection)
  • Tools: 6 HIGH fixes (git cwd, shell close, cron atomic, browser connect, multimodal path, per-tool timeout)
  • CLI/TUI: 8 HIGH fixes (config schema, null spread, silent lock, slash dispatch, ThinkingBlock Enter, white text, commitLive race, approval stdin)
  • Daemon: Graceful shutdown with AbortController
  • Fleet: DAG workflows, scratchpad lock safety, tripleshot null safety
  • Tests: 97 → 104 tests (+fleet DAG, +TUI render)

Gaps remaining (post-v1.18)

  • Fleet DAG: single-level chaining only (no branching/parallel stages)
  • No TTY-level integration tests for Ink TUI (covered by headless + UX scorer)
  • No distributed fleet execution (single-machine only)
  • No session sync/push across machines