Skip to content

Release v0.6.0

Choose a tag to compare

View all tags
@github-actions github-actions released this 18 Aug 23:06
· 104 commits to main since this release
v0.6.0
935513e

⚡ Local Coding Agent v0.6.0

🚀 Headline Features

  • Multi-Dimensional Capability Ladder (local_coding_agent.capability) (R15):

    • 5-tier difficulty taxonomy: Tier 0 (Syntax & Typo Repair), Tier 1 (Atomic Pure Functions), Tier 2 (Single-File Multi-Hunk), Tier 3 (Cross-File Invariants), Tier 4 (Algorithmic & Strict Constraints).
    • Adaptive Early-Exit ladder benchmark (local-agent benchmark --ladder).
    • Formal CapabilityVector tracking 95% Wilson confidence intervals, tokens/second generation speed, and maximum digestible chunk size (granularity_tolerance).

  • MCP Capability Discovery & Pre-Flight Gatekeeper (R16):

    • Exposes official MCP resource model://profile providing host agents (Codex, Claude, Antigravity) with dynamic insight into local model intelligence tiers and latest benchmark scores.
    • Pre-flight complexity gatekeeper rejecting out-of-tier task envelopes with CAPABILITY_OVERLOAD and pinpointed decomposition advice.

  • Multi-Backend Resilience (llama-server / OpenAI API):

    • Graceful fallback in ModelMemoryManager.snapshot() when VRAM introspection is not exposed by OpenAI-compatible backends.
    • Controller and tool context headroom optimizations (128 KB limit support).
    • Human-friendly ASCII capability ladder rendering in console CLI (local-agent benchmark --ladder).

🔬 Real-World Dogfooding & Field Report (Built with Local Coding Agent)

Important

100% Dogfooding Proof: This entire release (R15 & R16) was engineered and developed using our own Local Coding Agent service, delegating real implementation and refactoring tasks directly to the local Ling-3.0-tiny model!

Model Hardware & Hugging Face Specifications:

  • Model: inclusionAI/Ling-3.0-tiny (Ant Group / inclusionAI).
  • Architecture: BailingMoeV3 hybrid MoE with hybrid linear-attention (3:1 alternating stack of KDA: Kimi Delta Attention and MLA: Multi-Head Latent Attention).
  • Parameters: 7.9B total parameters, with only ~1.3B–1.4B active parameters per token (128 routed experts, 8 active + 1 shared).
  • Format & Runtime: Ling-3.0-tiny-Q6_K.gguf (6.5 GB) served locally via llama-server daemon.

Key Empirical Field Findings:

  1. Ultra-Fast Local Generation: Because only ~1.3B parameters are activated per token, the model sustains 118.8 tokens/sec on local hardware, completing atomic subtasks in 2–3 seconds without cloud latency or API billing.
  2. The Context Window Invariant: Small models fail when dumped with wide context (>300 lines of source code triggers context limits). Restricting tasks to strict 1-file / 1-function envelopes with character-exact SEARCH/REPLACE blocks yielded 100% first-turn success.
  3. SEARCH/REPLACE vs. Unified Diff: MoE models struggle with line-offset arithmetic in raw @@ -x,y +x,y @@ unified diffs, but generate structured JSON edits (file, search, replace) with surgical accuracy. The controller's internal translation to git apply --check eliminates all friction.
  4. Verified Intelligence Ceiling: The capability ladder proved Ling-3.0-tiny is a solid Tier 2 (Single-File Multi-Hunk) engine (100% Tier 0, 100% Tier 2 single-file edits, 95% on 20-case baseline benchmark), while multi-file coordination (Tier 3) requires decomposition by the host agent.
  5. Zero-Distillation Mediated Apply: Automatic test verification and git auto-rollback gave 100% confidence during development — zero regressions across the 290-test test suite.
Assets 4
Loading