Tags give the ability to mark specific points in history as being important
-
-
-
-
-
-
-
-
-
-
-
-
-
-
v0.11.2
a287a7d7 · ·v0.11.2 — skill security-documentation hardening (destructive-delete + exfil guardrails)
-
-
-
v0.10.1
37d0234b · ·v0.10.1 — ClawHub skill publishing + .agents/skills layout + model-catalog doc sync (8 ASR / 7 TTS)
-
v0.10.0
d90400e1 · ·docker-talkies v0.10.0 — configurable log level + opt-in full-request DEBUG logging Minor release. Two threads. 1. Logging. New TALKIES_LOG_LEVEL env var (falls back to LOG_LEVEL), case-insensitive: debug / info / warn / error / fatal. Default info; an unrecognized value fails fast at startup. At debug, the HTTP boundary logs FULL request + response bodies as structured JSON — TTS input text / instructions, cloned-voice reference transcripts, and ASR transcripts. That's PII: a one-time WARNING fires at startup when debug is active, and info+ never logs body content. 2. Housekeeping. Repo-wide black + isort format pass plus committed flake8 (.flake8, line length 88) and mypy ([tool.mypy] ignore_missing_imports) config so `make lint` runs clean. Extracted the executor allowlist into a single VALID_EXECUTORS constant. README documents the new env var; CHANGELOG.md created + backfilled. Wire-compatible with v0.9.0 — no API or request-shape change.
-
v0.9.0
a1b7d2af · ·docker-talkies v0.9.0 — Nemotron-3.5-ASR (parakeet.cpp) + GPU drain barrier Minor release. Three user-visible threads. 1. New multilingual ASR via mudler/parakeet.cpp (C++17/ggml). First slug shipped: nemotron-3.5-asr-0.6b (NVIDIA Nemotron-3.5-ASR-Streaming-0.6B, OpenMDW-1.1, 40+ locales, WER-0 against NeMo). Runs CPU-only in both images at this stage; the C-API JSON path returns per-word timestamps + confidence and the backend synthesizes Whisper- style segments from them via silence-gap grouping so verbose_json matches OpenAI's shape. Operators can register more parakeet.cpp checkpoints (any Parakeet TDT/CTC/RNNT GGUF in mudler/parakeet-cpp-gguf) via a custom models.json. 2. GPU drain barrier (server.py). Latent v0.8.0 bug: sibling eviction returned before async CUDA dealloc finished, so the next backend's load could race the still-freeing pool and OOM on a tight GPU. Now we synchronize() between gather() and the next get_model(). Microseconds when idle. 3. Integration harness — per-test filter. Pass any positional args to an e2e_*.sh / test_*.sh file and they act as an exact-or-substring whitelist over the test functions. Lets you re-run one failing case without recycling the full 5-model harness boot. Plus build-context cleanup (.dockerignore additions drop the docker daemon's per-build transfer from 24 GB to 2.83 KB when the e2e cache is warm) and a CPU-image refresh that registers the same nemotron slug. Wire-compatible with v0.8.0. No request shape changes. Full regression sweep on the CUDA host: 62/62 e2e cases green.
-
v0.8.0
388b100d · ·docker-talkies v0.8.0 — Qwen3-TTS CustomVoice + VoiceDesign + 1.7B Base + sampling controls Minor release. Two user-visible threads. 1. Full Qwen3-TTS mode coverage. Four new model slugs covering all three upstream operational modes: qwen3-tts-1.7b Base 1.7B voice cloning qwen3-tts-0.6b-custom CustomVoice 0.6B (9 preset speakers) qwen3-tts-1.7b-custom CustomVoice 1.7B (preset speakers + emotion) qwen3-tts-1.7b-design VoiceDesign 1.7B (synth voice from NL description) Mode is implicit in the model slug; the OpenAI wire format stays pure. `voice` and `instructions` semantics shift per mode (preset name / reference-WAV path / sentinel; emotion / NL description / optional style hint). GET /v1/audio/voices returns the right shape per mode. 2. Per-request sampling controls as OpenAI extras. New Optional fields on POST /v1/audio/speech, sent via `extra_body` on official OpenAI SDKs: temperature, top_k, top_p, repetition_penalty, max_new_tokens, do_sample. Plus `language` for selecting the spoken language on custom_voice and voice_design modes. Out-of-range returns 422. Bundled in: a Dockerfile build fix (--no-config on the heavy --require-hashes install) that v0.7.1 missed. Without it, v0.7.1 builds fail when transitive deps land newer than the [tool.uv] exclude-newer gate. Apply v0.8.0 to actually run the v0.7.1 supply-chain pipeline. E2E suite: tests/integration/e2e_qwen3_modes.sh exercises every variant via ASR round-trip — 12 cases, all green. Canonical fixture pair (audio.mp3 + audio.mp3.txt) committed so the suite runs deterministically on a fresh clone. Backwards compatible with v0.7.1: every existing slug works identically; new request fields are all Optional.