v0.2.0

Antfly version v0.2.0

Documentation (2026-08-11)

The documentation is available here:

https://docs.antfly.io

What's Changed

• fix: reject writes outside byte range during split transition by @ajroetker in #35
• Add serverless operator, proxy, and TLA+ trace validation by @ajroetker in #36
• Fix large linear merge request handling by @ajroetker in #39
• Fix divide-by-zero panic in Pebble FileCache during shutdown by @ajroetker in #40
• Fix TLA trace bridge and CI validation by @dovinmu in #41
• Add AFB portable backup format for cross-backend backup/restore by @ajroetker in #42
• fix antflycluster phase calculation by @timkaye11 in #49
• operator: add single-node swarm mode support by @timkaye11 in #50
• Consolidate proxy and operator modules into the monorepo by @ajroetker in #51
• Consolidate pkg/termite into monorepo and move API to /ml/v1/* by @ajroetker in #52
• Propagate AntflyCluster labels to pods by @ajroetker in #54
• Propagate termite pool labels to pods by @ajroetker in #55
• Add index query latency metrics by @ajroetker in #56
• Consolidate Antfly and Termite operators by @ajroetker in #57
• Add operator runtime recovery diagnostics by @ajroetker in #58
• Add Antfly TypeScript design system workspace by @ajroetker in #60
• Add operator autoscaling and storage scale-down safety by @ajroetker in #59
• Apply TypeScript private deps patch by @ajroetker in #61
• Add canonical design system exports by @ajroetker in #62
• Fix operator PVC ownership for non-root Antfly runtime by @ajroetker in #63
• Adopt shared design system in Antfarm by @ajroetker in #64
• Add structured record mapping utilities by @ajroetker in #65
• Fix Termite embedded health routes by @ajroetker in #68
• Use Cloud Build for container publishing by @ajroetker in #67
• Consolidate Termite backend CI by @ajroetker in #70
• Add external Termite pool discovery by @ajroetker in #69
• CI arc revamp by @dovinmu in #53
• ci: run omni on Ubuntu 24 ARC runner by @ajroetker in #71
• Migrate Zig runtime into monorepo by @ajroetker in #72
• Add GLiNER2 relation extraction direct quant support by @timkaye11 in #74
• zig: rotate remaining secret consumers live by @ajroetker in #73
• Document and guard join-aware auth by @ajroetker in #76
• Remove Codecov from Python CI by @ajroetker in #78
• lora policy alignment across finetuning recipe layer by @timkaye11 in #75
• Fix SearchAF Zig runtime failures by @ajroetker in #83
• zig: add secrets reload status coverage by @ajroetker in #77
• Harden A2A and internal retrieval routes with admin-only requirement by @dovinmu in #82
• Require authentication for MCP public endpoint by @dovinmu in #79
• Fix metadata admin bootstrap ordering by @ajroetker in #84
• Fix remote content credential resolution for non-S3 URLs by @ajroetker in #86
• Fix SearchAF swarm query and health binding by @ajroetker in #93
• Bump vulnerable TypeScript dependencies by @dovinmu in #92
• Restrict Termite Electron file serving by @dovinmu in #89
• Use Zig runtime contract in operators by @ajroetker in #87
• Fix SearchAF Zig runtime response contracts by @ajroetker in #96
• Fix Zig OOM retirement lifecycle cleanup by @dovinmu in #80
• Fix Zig partial config defaults by @ajroetker in #94
• Fallback algebraic bucket aggregations by @ajroetker in #98
• Reduce operator RBAC permissions by @dovinmu in #90
• Repair stale swarm StatefulSet rollouts by @ajroetker in #99
• Harden store backup path handling by @dovinmu in #88
• Bump vulnerable Go dependencies by @dovinmu in #91
• Jina v5 direct metal embeddings by @timkaye11 in #95
• Antfarm Zig Updates by @timkaye11 in #85
• Make Zig runtime the default release by @ajroetker in #81
• Add native Gemini multimodal embeddings by @ajroetker in #102
• Make Antfarm runtime configurable for hosted deployments by @ajroetker in #103
• Build Antfarm dashboard assets by @ajroetker in #105
• Hide local Antfarm admin routes in hosted mode by @ajroetker in #106
• Add Bedrock multimodal embedding support by @ajroetker in #104
• Bedrock Titan Multi-modal Support by @excenter in #107
• Fix Zig container root-context build by @ajroetker in #108
• Avoid creating legacy storage dirs for Zig swarm by @ajroetker in #110
• Rebuild Antfarm assets without conflict markers by @ajroetker in #111
• Fix composed vector query correctness and speed hybrid filtering by @ajroetker in #100
• Consolidate SDK contracts and Go layout by @ajroetker in #112
• Support CloudAF data-plane client URLs by @ajroetker in #109
• Fix CI failures from split lifecycle and SDK/runtime drift by @ajroetker in #114
• fix(sdk): bundle openapi-fetch to fix CJS interop crash by @markhayden in #45
• fix(termite-sdk): bundle openapi-fetch to fix CJS interop crash by @markhayden in #46
• Unify managed antfly and termite embedders by @ajroetker in #116
• Fix remote content security config handling by @ajroetker in #115
• Update Epstein example for Zig Termite runtime by @ajroetker in #113
• Fix dependency-free ONNX model loading by @ajroetker in #118
• Fix TTL metadata for local document writes by @ajroetker in #121
• Fix quantized GGUF embedding lookup shapes by @timkaye11 in #122
• Productionize GLiNER2 resident fine-tuning on Metal and MLX by @timkaye11 in #123
• Fix query aggregation response label lifetime by @ajroetker in #125
• Speed up Zig Termite models listing by @ajroetker in #124
• Implement DOCID document identity and query APIs by @ajroetker in #97
• Make Termite readyz independent of model inventory by @ajroetker in #127
• Add cluster topology status endpoint by @ajroetker in #128
• Move container publishing to us-central1 GAR by @ajroetker in #126
• Consolidate Go unit and omni CI by @ajroetker in #132
• Support shared Termite API keys by @ajroetker in #130
• Fix bool must+filter clause handling in structured filter parser by @dovinmu in #131
• Run Playwright CI on ARC with local browser installs by @ajroetker in #133
• @antfly/design-system visual language rebuild by @stinkbugaf in #135
• antfarm: visual sweep to design-system 0.1.0 (square, flat, amber, mono) by @stinkbugaf in #136
• Design asset artifacts API by @ajroetker in #117
• Consolidate inference CLI under antfly by @ajroetker in #129
• Accelerate GLiNER2 with Zig Metal, CPU, and ONNX Runtime support by @timkaye11 in #137
• Fix macOS Zig native Metal builds without MPSGraph by @timkaye11 in #134
• Add Antfly Cloud CLI shim by @dovinmu in #139
• Add extraction inference helpers by @ajroetker in #147
• Bound SDK batch request and response bodies by @ajroetker in #146
• Harden autoscaling split e2e wait by @ajroetker in #149
• Add TypeScript SDK runtime guardrails by @timkaye11 in #148
• design-system v0.1.1: Heading/Text/Link primitives + highlight fix by @stinkbugaf in #150
• Reduce load memory pressure by @ajroetker in #138
• Add trusted-principal MCP auth by @ajroetker in #151
• Fix CLIP image preprocessing contract by @ajroetker in #152
• Add Antfarm Graph Index Explorer by @timkaye11 in #154
• Add pip and npm CLI packaging by @ajroetker in #156
• Fix CLI package bin path by @ajroetker in #157
• Namespace Python SDK release tags by @ajroetker in #158
• Wire release container publishing to Zig runtime by @ajroetker in #159
• Fix TypeScript CI runner temp env by @ajroetker in #160
• Run Python SDK generation check on canonical Python by @ajroetker in #161
• Drop Python 3.9 support for Python packages by @ajroetker in #162
• Fix Antfly release Zig setup by @ajroetker in #163
• Add entity-resolution design and comparison-levels scorer by @ajroetker in #142
• Stop building Go fallback artifacts during release by @ajroetker in #167
• Harden CLI NDJSON loading by @ajroetker in #165
• include Antfly Inference unit tests in ci pipeline by @timkaye11 in #166
• Block graph replay on missing resolver contracts by @ajroetker in #169
• Harden metadata runtime readiness by @ajroetker in #168
• Split full-default Zig CI phases by @ajroetker in #170
• Set up Go for GoReleaser metadata by @ajroetker in #171
• Stabilize scheduled Zig CI followups by @ajroetker in #172
• Replace GoReleaser release publishing by @ajroetker in #173
• Fan out container release publishing by @ajroetker in #174
• Add recall harness CI progress output by @ajroetker in #175
• Add Gemma4 12B unified image/audio inference support by @timkaye11 in #176
• Add resolution graph flake diagnostics by @ajroetker in #178
• Tune bulk ingest and inference smoke flows by @ajroetker in #177
• Support GLiNER2 split GGUF bundles by @ajroetker in #185
• Add Epstein graph visualization by @ajroetker in #186
• Fix public index and inference pull regressions by @ajroetker in #184
• Improve recall harness failure diagnostics by @ajroetker in #187
• Make recall harness heartbeat optional by @ajroetker in #188
• ci: trigger Colony docs builds by @ajroetker in #189
• Fix rc2 query and inference regressions by @ajroetker in #198
• Segment transaction TLA traces for CI by @ajroetker in #199
• Fix inference and enrichment issue regressions by @ajroetker in #209
• Store LSM bloom filters only in SSTables by @ajroetker in #211
• Support GLiNER2 ONNX Exports and Symbolic Metal Fallbacks by @timkaye11 in #212
• Add native Traditional ML predictors with CLI model pull and /ml/v1/predict API by @timkaye11 in #155
• Tune LSM online write and bulk ingest behavior by @ajroetker in #200
• Make Gemma4 Metal tool calling production-ready by @timkaye11 in #218
• Add resolver replay diagnostics to index status by @ajroetker in #213
• Zig query perf: live-doc visibility, search effort policy, stale payload repair by @dovinmu in #217
• Restore Go portable AFB backups in Zig by @dovinmu in #215
• Fix e2e flakes: LSM reclaim read race and startup catch-up livelock by @ajroetker in #221
• Read-path scalability, quarantine self-heal, and HBC link repair by @ajroetker in #220
• Design system 0.2: quiet chrome, restraint, and the ink primary by @stinkbugaf in #222
• Add derived document hierarchy support by @ajroetker in #216
• Fix raft apply deadlock behind nightly e2e hangs; harden CI diagnostics by @ajroetker in #224
• Implement docsaf derived document hierarchy by @ajroetker in #225
• Fix swarm MCP route registration by @ajroetker in #227
• Bound metadata raft memory retention by @ajroetker in #226
• Fix linalg futex wait syscall by @ajroetker in #229
• Add GET /connections: list connected providers, models, and stores by @ajroetker in #228
• Clean up Antfarm connections rendering by @ajroetker in #231
• Simplify Antfarm navigation by @ajroetker in #230
• Implement Antfarm experience follow-ups by @ajroetker in #232
• Recover malformed PDF content parsing by @ajroetker in #234
• Fix document extraction failure recovery by @ajroetker in #235
• Align dense stress test with stable vector ids by @dovinmu in #219
• Fix runtime shapes for symbolic inference ops by @ajroetker in #236
• Fix ONNX symbolic projection embeddings by @ajroetker in #237
• Extension system v1 by @ajroetker in #233
• Remove MLX backend support by @ajroetker in #119
• Update TypeScript workspace dependencies by @ajroetker in #239
• Add data-plane auth CLI commands by @ajroetker in #238
• Bound document extraction memory by @ajroetker in #240
• Drop corrupt indexes without blocking shard startup by @ajroetker in #241
• Add Google Drive auth to docsaf CLI by @ajroetker in #242
• Fix Antfarm table detail crash on sparse schemas by @timkaye11 in #245
• CUDA 13.2 production readiness for ClipClap, GLiNER2, and mxbai-rerank by @timkaye11 in #244
• Filter extension MCP tools by permissions by @ajroetker in #246
• Enable Tensor Core CUDA dispatch for mxbai reranker, CLIP/CLAP, and GLiNER2 by @timkaye11 in #249
• Stabilize scaling drain e2e seed writes by @ajroetker in #251
• Fix backup job URL configuration by @ajroetker in #247
• support fielded MCP full text search by @ajroetker in #254
• Speed up Gemma4 E2B on Metal by @timkaye11 in #252
• add SDK model pull helpers by @ajroetker in #248
• Fix full text enrichment index provisioning by @ajroetker in #250
• Add tenant-scoped ARD discovery surfaces by @ajroetker in #253
• support raw MCP query requests by @ajroetker in #256
• Add hot-standby HA design and runtime surface by @ajroetker in #255
• configure retrieval agent tools by request step by @ajroetker in #259
• Speed up Gemma4 CUDA inference by @timkaye11 in #258
• Mount public API on metadata server by @ajroetker in #262
• Florence2 and the Machine: Metal Version by @timkaye11 in #263
• Stabilize TTL full-text search test by @ajroetker in #264
• implement AntflyLite by @ajroetker in #261
• CUDA support for Florence-2 GGUF by @timkaye11 in #265
• Fix lite index benchmark reporting by @ajroetker in #268
• Fix startup catch-up busy retry loop by @ajroetker in #267
• Gemma4 CUDA production readiness: E4B Q4_K, MTP, and 12B encoder-free multimodal by @timkaye11 in #266
• Improve Gemma 4 E2B Metal inference reuse and fix unbounded RAG generation by @timkaye11 in #260
• Fix operator container build with local module replaces by @ajroetker in #269
• Flatten extension package source layout by @ajroetker in #270
• Shrink generated operator CRDs by @ajroetker in #271
• Use read-index barrier for metadata backups by @ajroetker in #273
• Support Antfly inference defaults by @ajroetker in #275
• Add artifact CLI and artifact enrichment listing by @ajroetker in #274
• Fix HTTP raft snapshot catch-up by @ajroetker in #276
• Fix clipclap ONNX embedding correctness by @ajroetker in #280
• Fix backup forwarding and runRound phase logs by @ajroetker in #282
• Fail loud on remote media render failures by @ajroetker in #283
• Use doc_type for docsAF rows and harden readiness tests by @ajroetker in #285
• Fix dense replay catch-up convergence by @ajroetker in #290
• Harden LSM root and WAL locking by @ajroetker in #284
• Fix dense replay progress and alias-safe startup drains by @ajroetker in #292
• Diagnose backup raft stalls and reconcile backup jobs by @ajroetker in #294
• Fix data raft leader readiness publication by @ajroetker in #295
• Fix Metal SDPA seq-major layout by @ajroetker in #297
• Clean recovered LSM table temps and detail raft persist latency by @ajroetker in #296
• Fix backup restore writer conflict by @ajroetker in #298
• Clean recovered LSM temp files before WAL replay by @ajroetker in #299
• Harden recovery and backup operations by @ajroetker in #300
• Route enrichment assets into full text by @ajroetker in #301
• Fix metadata bootstrap churn and raft no-op persistence by @ajroetker in #311
• Default imported ONNX graphs without enabling ORT by default by @timkaye11 in #315
• Gemma4 E4B QAT CUDA by @timkaye11 in #320
• Fix artifact-backed chunk embeddings by @ajroetker in #318
• Fix CLIP image preprocessing and Metal parity by @timkaye11 in #323
• Handle unsupported OCR backend paths by @stinkbugaf in #324
• Productionize Gemma4 E4B QAT Q4_0 on Metal: kernels, frame safety, sampling, speculative groundwork by @timkaye11 in #321
• Repair corrupt embedding artifact replay by @ajroetker in #316
• Fix thin replay delete hints for managed indexes by @stinkbugaf in #327
• Fix document extraction state recovery by @stinkbugaf in #329
• Harden index repair followups by @ajroetker in #326
• Antfly Lite: cache native pages, chain links, and docstore snapshots by @dovinmu in #330
• Include OpenAPI specs in Docker build context by @bpopadiuk in #333
• Skip latest aliases for v0.1 releases by @ajroetker in #332
• Wire repair regression tests into focused bucket by @ajroetker in #331
• Harden std HTTP listener threading by @ajroetker in #312
• Fix HA admin token handling by @bpopadiuk in #336
• Add per-item embedding error policy by @ajroetker in #338
• Pin release publishing to npm 11 by @ajroetker in #341
• Batch native Florence reader inference by @ajroetker in #340
• Fix provisioned write coalescer losing later writes to deleted keys by @dovinmu in #314
• Fix HA failover recovery from live Kind validation by @bpopadiuk in #339
• Exclude-set live-doc filtering and cached WAL retention snapshots by @dovinmu in #342
• Fix HA promotion and replication state races by @bpopadiuk in #343
• Cache query embeddings with bounded singleflight by @ajroetker in #348
• Add prompt-prefix KV caching for native, Metal, and CUDA inference by @timkaye11 in #337
• [pdfs] Preserve remote PDF MIME across large downloads by @bpopadiuk in #357
• [bug] Fix HTTP listener timeout starvation by @bpopadiuk in #358
• Productionize Florence-2 Metal KV-cache decoding by @timkaye11 in #349
• Fix MCP index list structured content by @jmcdermott45 in #361
• Harden derived coverage, restore generations, and full CI gates by @ajroetker in #335
• Add standalone Lite storage and harden Zig storage lifecycle by @ajroetker in #344
• Bound zig-base unit test concurrency by @ajroetker in #362
• Harden dense index publication and restart recovery by @ajroetker in #363
• Productionize bounded multi-model inference by @timkaye11 in #359
• Add multi-source artifact indexes across API and SDKs by @ajroetker in #367
• Optimize full-text index performance by @ajroetker in #365
• Harden storage generations, Raft snapshots, and split publication by @ajroetker in #364
• gate hot-standby HA pending paired release by @bpopadiuk in #373
• Make Antfarm inference connection-first by @timkaye11 in #369
• Add PDF OCR fallback to document extraction by @dovinmu in #374
• Fix Linux musl statvfs builds by @bpopadiuk in #375
• Fix S3 document fetch hangs by @dovinmu in #379
• Route selective dynamic filters through exact vector search by @dovinmu in #380
• Implement secret and remote-content reload lifecycle by @ajroetker in #387
• Use opaque secret publication acknowledgements by @ajroetker in #389
• Report layered secret acknowledgement capability truthfully by @ajroetker in #390
• Optimize Metal SDPA with bounded-barrier dispatch by @ajroetker in #391
• Fix Metal SDPA drop-under-load regression by @bpopadiuk in #388
• [Codex] July 24 release hardening: integrate #371 and triage smokeout blockers by @dovinmu in #393
• Back off recoverable derived-index retries by @bpopadiuk in #401
• Optimize Zig tokenizer toward Gigatoken throughput by @ajroetker in #402
• Optimize large-corpus Zig tokenization by @ajroetker in #403
• fix(zig): harden 0.2.0 release blockers by @dovinmu in #409
• inference: make the 0.2.0 model support surface true by @dovinmu in #404
• Fix MCP bridge leaks and gate A2A as experimental by @dovinmu in #413
• Fix restore, split, and HA lifecycle races by @ajroetker in #371
• Harden StreamVByte decoding and add native ARM64 gate by @bpopadiuk in #411
• Fix Zig full and inference CI regressions by @ajroetker in #415
• Protect rebuild cursors against silent corruption by @ajroetker in #414
• Strip Zig production release artifacts by @ajroetker in #417
• Use ReleaseFast for ARM64 release archives by @ajroetker in #421
• Fix scheduled full Zig and inference CI failures by @ajroetker in #419
• Harden autosplit median-key route fallback by @ajroetker in #425
• storage: bound LSM WAL retention and run files by @ajroetker in #418
• Fix full-text prefix and deletion query cliffs by @ajroetker in #420
• Harden Zig overload recovery and keyword recall by @bpopadiuk in #424
• Restore release-safe ARM64 archive settings by @ajroetker in #428
• Harden full-text index provenance and analyzer publication by @ajroetker in #429
• Fix projected graph pushes and index deletion races by @ajroetker in #430
• fix(inference): serialize GPU embedding execution by @ajroetker in #436
• Fence dense catalog lookup before fast-path preflight by @ajroetker in #432
• Fix full inference E2E model contention by @ajroetker in #427
• fix: automate schema generations and macOS inference admission by @ajroetker in #438
• Bound native FDs and prevent text merge livelock by @ajroetker in #433
• Fix full E2E promotion and inference flakes by @ajroetker in #439
• Harden storage recovery and index readiness by @ajroetker in #445
• Canonicalize extraction and harden inference E2Es by @ajroetker in #446
• GLiNER2 fine-tuning: Zig-native LoRA training at result parity with upstream by @timkaye11 in #377
• Retry transient Zig ARM64 release allocation failures by @ajroetker in #452
• Fix strict multimodal admission and retained HTTP workers by @ajroetker in #449
• Harden inference E2E admission and organize GLiNER2 fixtures by @timkaye11 in #447
• Fix rc.26 filtered ANN and inference quality gaps by @ajroetker in #456
• Harden Raft write routing during leader transitions by @ajroetker in #461
• Fix dynamic sparse ONNX batches and trim padded sequences by @timkaye11 in #457
• Fix enrichment retry lifecycle and repair UX by @ajroetker in #462
• Fix fused aggregations on truncated result pages by @ajroetker in #464
• fix(zig): make batch writes atomic across shards by @ajroetker in #458
• Fix Raft timeout diagnostics from wedging writes by @ajroetker in #465
• Fix metadata simulation transaction wire outcomes by @ajroetker in #466
• Fix inference run help starting the server by @dovinmu in #459
• Enforce read permission for document scans by @dovinmu in #467
• Batch and harden GLiNER2 and reranker inference by @timkaye11 in #470
• fix(api): handle missing tables in HTTPX reads by @ajroetker in #471

New Contributors

• @timkaye11 made their first contribution in #49
• @markhayden made their first contribution in #45
• @stinkbugaf made their first contribution in #135
• @bpopadiuk made their first contribution in #333
• @jmcdermott45 made their first contribution in #361

Full Changelog: v0.1.1...v0.2.0

---

Released by Antfly release scripts.