Tags give the ability to mark specific points in history as being important
-
-
v3.17.0
0a31ef85 · ·v3.17.0 — talkies v0.13.2 + five new streaming ASR models (Sherpa-ONNX Zipformer, Vosk) on CPU and CUDA
-
-
-
-
-
v3.15.7
08a51c70 · ·v3.15.7 — Claude Code + Codex plugin manifests, Agent integrations README section
-
-
-
-
v3.15.3
ca8f882d · ·v3.15.3 — sdcpp image-generation fixes (juggernaut-xi load, VAE tiling, fallback thrash)
-
-
-
-
v3.14.9
e5a83cfe · ·v3.14.9 — bump claudebox -> v2.0.13, pibox-zai -> v0.14.0 Restores the claudebox v2 API-mode migration that had drifted out of docker-compose.yml (CLAUDEBOX_API_MODE env vars, /home/aicode/.aicodebox + /workspace volume paths, /healthz healthcheck), enables claudebox's MCP server (CLAUDEBOX_MCP_MODE), and fixes a file-ops 500 caused by a root-owned workspace host dir. Bumps both images to their current tags: psyb0t/claudebox:v1.14.0-minimal -> v2.0.13 psyb0t/pibox:v0.12.0 -> v0.14.0 Full tests/test_claudebox.sh suite verified 4/4 passing against the recreated containers.
-
v3.14.8
4dfd97cd · ·v3.14.8 — cap LiteLLM container memory The litellm service in docker-compose.yml had no mem_limit, memswap_limit, or pids_limit — every other long-running service in the stack had caps, but the LLM gateway itself was the one gap. Container was drifting from ~800 MiB idle to 3+ GiB under sustained load with no ceiling, eventually starving the host. Added: mem_limit: ${LITELLM_MEM_LIMIT:-4g} memswap_limit: ${LITELLM_MEMSWAP_LIMIT:-8g} pids_limit: ${LITELLM_PIDS_LIMIT:-512} Sizing: 4 GiB matches 4 workers at ~1 GiB each plus master. Swap at 2x mem matches convention. PIDs at 512 covers async workers + subprocesses. Outcome: LiteLLM OOM-kills only itself under load and auto-restarts. Host stays alive. Contained failure. Full-stack sweep: every other long-running service has caps. postgres-init (one-shot) is the sole remaining unbounded container — negligible blast radius. -
v3.14.7
60035479 · ·v3.14.7 — bump pibox v0.11.3 → v0.12.0 Image-tag bump only. Base tracked to psyb0t/aicodebox:v0.11.0. Upstream behavior change: /openai/v1/chat/completions returns HTTP 400 with the provider's error message on provider rejection (content-safety filter / rate limit / auth failure) instead of HTTP 200 with empty completion body. Pi reports the underlying detail as stopReason=error + errorMessage on the assistant turn. Aigate impact: LiteLLM's proxy sees a real 4xx and retries per its num_retries + fallback chain instead of ambiguous phantom success reaching the caller. No API / env / config change on the aigate side. HEADS-UP: psyb0t/pibox:v0.12.0 + psyb0t/aicodebox:v0.11.0 may not be on Docker Hub at this tag's cut moment.
-
v3.14.6
00e67f01 · ·v3.14.6 — talkies v0.10.0 + flickies v0.3.1 (CPU + CUDA) Image-tag bump only. Both upstream releases are wire-compatible with their prior versions — no aigate-side wiring changes. TALKIES v0.9.0 → v0.10.0: - New TALKIES_LOG_LEVEL env var (debug/info/warn/error/fatal). - Opt-in full-request DEBUG body logging (PII — TTS input, cloned-voice transcripts, ASR transcripts). Startup WARNING when debug active; body content never logged at info+. - Housekeeping — black + isort + committed flake8 / mypy config. FLICKIES v0.3.0 → v0.3.1: - Engine-unload memory-leak fix: gc.collect() before torch.cuda.empty_cache() so the cyclic model graph is reclaimed. Applies to idle sweep, DELETE /v1/engines/{slug}, hot-swap. Improves the v3.14.5 competing-group eviction path. - Opt-in DEBUG observability (FLICKIES_LOG_LEVEL=DEBUG): ffmpeg commands, transform decisions, inference wall_secs, URL byte counts, job lifecycle, HTTP 4xx/5xx rejections. - Security: logged URLs query-stripped (no presigned creds in logs). HEADS-UP: psyb0t/talkies:v0.10.0(-cuda) and psyb0t/flickies:v0.3.1 (-cuda) may not be on Docker Hub at this tag's cut moment. -
v3.14.5
57669601 · ·v3.14.5 — cross-service GPU eviction hardening + /v1/unload endpoints Closes the resource_manager gap that let audiolla / flickies OOM-kill talkies / ollama by hoarding VRAM. Ships new POST /v1/unload/{cuda,cpu} endpoints so operators can free VRAM without piggybacking on a fake LLM call. Bundles the flickies v0.2.0 → v0.3.0 upgrade. RESOURCE_MANAGER: Added cuda-audiolla + cuda-flickies (+ CPU siblings) to _ALL_CUDA_GROUPS / _ALL_CPU_GROUPS. Matching _unload_* funcs hit audiolla's POST /v1/unload (bulk evict) and flickies' DELETE /v1/engines/{slug} (per loaded engine). Downstream calls carry AIGATE_TOKEN as bearer. Before: LatentSync holding 7 GiB VRAM → next talkies Whisper load → torch.OutOfMemoryError → container restart. After: pre_call unload includes audiolla + flickies → their engines released → next call has VRAM. NEW ENDPOINTS: POST /v1/unload/cuda — fan out to all 7 CUDA services POST /v1/unload/cpu — CPU counterpart POST /v1/unload — both classes Response is per-service outcome map. Per-service errors do not fail the whole call. Implemented in mcp/server.py via @mcp.custom_route; nginx routes /v1/unload/{cuda,cpu,''} → mcp:8000 with 120s timeout. FLICKIES v0.2.0 → v0.3.0 (CPU + CUDA): Image-tag bump. Adds `precise: bool` (default false) on POST /v1/video/trim + POST /v1/video/concat for frame-accurate boundaries (libx264 CRF 18 / AAC 192k). Default preserves prior stream-copy behavior. Pure additive. Minor bump — additive endpoints + new groups + image bumps. No breaking wire changes. HEADS-UP: psyb0t/flickies:v0.3.0 + psyb0t/flickies:v0.3.0-cuda may not be on Docker Hub at this tag's cut moment. -
v3.14.4
cd77e76f · ·v3.14.4 — expose all 8 z.ai GLM models z.ai serves 8 GLM models; aigate was only exposing 3. This release adds the missing 5 routes so the full z.ai catalog is reachable via pibox-zai aliases. No removals, no breaking changes. ADDED: - pibox-zai-glm-5.2 (newest flagship, released 2026-06-17) - pibox-zai-glm-5-turbo (fast tier) - pibox-zai-glm-5 (earlier 5-series) - pibox-zai-glm-4.6 - pibox-zai-glm-4.5 PIBOX_AVAILABLE_MODELS default expanded to all 8. QUOTA NOTE: z.ai bills the 5-series at 3x during Beijing peak (14:00-18:00) and 2x off-peak (1x under the current promo through end of September 2026). The 4-series bills at the flat 1x rate. Defaulting to glm-5.2 burns subscription quota ~3x faster than glm-4.7 during peak hours. docs/providers.md has the breakdown. Patch bump — additive provider audit. Smoke clean for all 5 new routes via litellm → pibox → z.ai.