v0.2.0

[0.2.0] - 2026-09-06

The reproduction + benchmark release (unswallow 0.2.0, unswallow-matrix
0.2.0, Python unswallow 0.2.0). The headline: the first verified live
reproduction of a swallowed tool call, closing the evidence gap between
"reported upstream" and "reproduced and recovered here".

Added

• Verified live reproduction (Pattern A, llama.cpp b8461): drove the
  exact bug-era build from ggml-org/llama.cpp #20837
  (b8461, cea560f) with unsloth Qwen3.5-9B-UD-Q4_K_XL and thinking
  enabled through a multi-turn agent loop. After one successful tool call,
  the model emitted the XML envelope
  (<tool_call><function=read_file>...) inside reasoning_content and
  stopped with finish_reason: stop and no tool_calls — in both
  non-streaming and streaming modes. Raw responses were captured
  pre-unswallow and are pinned as fixtures
  llamacpp-b8461-qwen3.5-9b-multiturn-pattern-a /
  -streaming-multiturn-pattern-a; unswallow recovers
  read_file({"path": "package.json"}) at confidence 0.95. The engine
  matrix llama.cpp row is now verified: true with the live evidence
  recorded (matrix v1.3.0), and docs/reproduction.md documents the exact
  repro commands. Also recorded as dated negative live runs: llama.cpp
  b10819 + Qwen3-0.6B (healthy), Pattern B single-turn (9/9 healthy),
  Pattern C mid-think truncation (clean channel split).
• Live-reproduction harness (packages/bench/live-probe, TS +
  Python): runs recorded synthetic cases or any OpenAI-compatible endpoint,
  captures the raw provider response before unswallow processing, records
  provider/engine/version/model/pattern/stream/detection/recovery/latency/
  errors, and writes machine-readable JSON + human output. Live cases skip
  cleanly when no endpoint is configured.
• Real agent integrations verified (LangChain + LlamaIndex): installed
  langchain-openai / llama-index / openai and drove real agents
  through the unswallow proxy against a mock that serves the live-captured
  b8461 swallow — both frameworks execute the recovered read_file
  end-to-end (packages/examples/integration_langchain.py,
  integration_llama_index.py, run_framework_integrations.py, plus the
  manual-dispatch framework-integrations workflow).
• Formal fp-eval metrics: TP/FP/TN/FN + precision/recall/specificity/F1
  reported in both languages (pinned corpus now 12 positives + 8 negatives
  • 200 synthetic negatives; 0 FP / 0 FN, precision/recall/F1 1.0), with
    the evaluated engine-matrix version recorded in the results.
• Edge-case + regression sweep: 14 new tests per language (parsing
  escapes/unicode/whitespace/missing keys, streaming robustness against
  empty and non-string deltas, history corruption survival). TS 99, Python
  94.
• CI completeness: ruff lint on Python, fp-eval gates (TS + Python),
  bench:smoke aggregate on the test matrix, docs link checker over all
  markdown, and a package artifact validation job (npm pack + wheel
  install smoke). All green on GitHub Actions for main.

Changed

• Engine matrix matrixVersion 1.2.0 → 1.3.0 (llama.cpp row verified).
• docs/reproduction.md, docs/compatibility.md,
  docs/agent-integrations.md: Verified row, live-run records, exact repro
  commands, and real-install integration notes.

Migration notes

• No API or behavior changes; confidence/recovery for the affected
  llama.cpp + Qwen3.5 combination is unchanged (still 0.95 on a matrix
  hit). The matrix verified flag and the new fixtures are additive.