Issues

• [P2] Session log (JSONL): one shared recording path for OpenAI/Anthropic/Responses, shared meta/full record structure, full by default, session resume

  enhancement
  goodmai/basalt#106 · Filter by author goodmai opened on Oct 20, 2026

• [P2] Chat-template fidelity vs Jinja2 (comments, tojson) and reasoning_content alias

  goodmai/basalt#104 · Filter by author goodmai opened on Oct 19, 2026

• [P3] Return mlx-swift-lm from the 7e4024e revision pin to a release

  goodmai/basalt#102 · Filter by author goodmai opened on Oct 19, 2026

• [P2] Security: chat-template control tokens in user text forge role boundaries

  goodmai/basalt#99 · Filter by author goodmai opened on Oct 19, 2026

• [P2] Continue an assistant prefill instead of opening a new turn

  goodmai/basalt#98 · Filter by author goodmai opened on Oct 19, 2026

• [P3] ARC-AGI hybrid: numpy transforms get a Grid object and silently return the input

  goodmai/basalt#96 · Filter by author goodmai opened on Oct 19, 2026

• [P2] Hybrid prompt cache: store an SSM snapshot at the divergence point

  goodmai/basalt#95 · Filter by author goodmai opened on Oct 19, 2026

• [P2] ModelRouter: use currentAvailableRAM instead of physicalMemory, propagate cloud stream errors, and forward local modelId

  bug
  goodmai/basalt#89 · Filter by author goodmai opened on Oct 19, 2026

• [P2] Per-session prompt-cache namespace isolation and safe concurrent hybrid SSM snapshot retention

  enhancement
  goodmai/basalt#88 · Filter by author goodmai opened on Oct 19, 2026

• [P2] Wire RepetitionGuard into MLXInferenceEngine stream loop and bound CacheBlend prefill & token-text cache

  bug
  goodmai/basalt#87 · Filter by author goodmai opened on Oct 19, 2026

• [P2] max_tokens cap: memory budget should cap, not only above --max-tokens; weights counted twice

  bug
  goodmai/basalt#85 · Filter by author goodmai opened on Oct 19, 2026

• [P2] Prompt-cache tree: reuse RotatingKVCache layers (Gemma 3/4) before the window wraps

  enhancement
  goodmai/basalt#81 · Filter by author goodmai opened on Oct 19, 2026

• [P3] Error responses in each provider's shape: error type, Retry-After, failures before the stream starts

  enhancement
  goodmai/basalt#76 · Filter by author goodmai opened on Oct 19, 2026

• [P2] Orchestrator: bound concurrent generations, single-flight model switch, bounded shutdown drain

  bug
  goodmai/basalt#74 · Filter by author goodmai opened on Oct 19, 2026

• [P3] Keep AppKit out of the CLI binary (Splash, swift-huggingface)

  enhancement
  goodmai/basalt#65 · Filter by author goodmai opened on Oct 19, 2026

• [P3] Finish the Basalt* split: explicit product deps, drop Sources/mlx, remaining Gem names

  enhancement
  goodmai/basalt#64 · Filter by author goodmai opened on Oct 19, 2026
  1

• [P2] CI hardening: pin actions to SHAs, reuse main's test result on release, header-aware metallib freshness

  enhancement
  goodmai/basalt#63 · Filter by author goodmai opened on Oct 19, 2026
  1

• [P3] Rate limiter: per-client buckets and a separate burst size

  enhancement
  goodmai/basalt#62 · Filter by author goodmai opened on Oct 19, 2026

• [P3] Monitor tab: bring back the GUI client and show /api/v1/stats

  goodmai/basalt#55 · Filter by author goodmai opened on Oct 18, 2026
  1

• [P2] Roofline modeling and thermal-adaptive scheduling for Apple Silicon

  enhancement
  goodmai/basalt#45 · Filter by author goodmai opened on Oct 13, 2026
  1

• [P2] Sanitization of abliterated and low-bit quantized weights (QuaRot + refusal-subspace)

  enhancement
  goodmai/basalt#44 · Filter by author goodmai opened on Oct 13, 2026

• [P2] L2-norm key-based KV cache eviction compatible with fast attention

  enhancement
  goodmai/basalt#43 · Filter by author goodmai opened on Oct 13, 2026

• [P1] Grammar-constrained decoding with a byte-level pushdown automaton (XGrammar)

  enhancement
  goodmai/basalt#42 · Filter by author goodmai opened on Oct 13, 2026

• [P2] Optimal test-time compute allocation with KL-based early exit

  enhancement
  goodmai/basalt#41 · Filter by author goodmai opened on Oct 13, 2026

• [P2] Speculative cascades between local MLX and cloud models

  enhancement
  goodmai/basalt#39 · Filter by author goodmai opened on Oct 13, 2026