v0.9.27

0.9.27 (2026-05-30)

• fix(tools): non-null provider assertion in streamText shim loop (09d031ace322b4868f933b77621f31829f1885e4)
• feat(tools): buffered prompt-emulated tool loop (49184f6fc60219e6f34b8858c7f42c56d9761e42)
• feat(tools): prompt-emulated tool schema renderer (590f74f8ebdebf1e1a0e32314b8875d036a46082)
• feat(tools): prompt-emulated tool-call parser (fd030fac877c09ac838933b986aab3e2dac5fa3a)
• feat(tools): streamText auto reactive fallback for tool-incapable models (6dd2c58b1234e4133737f73f79a7b0a2424d428b)
• feat(tools): tool-emulation activation + response formatter (db2fb17a87cd0bbdea6ec309f88b26ac3c3f7de1)
• feat(tools): toolMode prompt shim in streamText (buffer hops, replay final) (6c96ff689f4bec0fa1d59b7cf4415070a82be4c4)
• feat(tools): toolMode shim path in generateText (prompt + auto reactive fallback) (2c7935876a54b5653ac1badc597a87defd28210d)

v0.9.26

0.9.26 (2026-05-30)

• fix(api): re-export setGlobalLlmObserver from package main (9036539f19dffa37154cb71cf91acd91b8ced3ad)
• test: repair relative imports for relocated colocated tests; drop dead tsconfig test globs (01b4e276e7b3668feec10e5c2cb564b44fb94e19)

v0.9.25

0.9.25 (2026-05-29)

• feat(api): global LLM usage observer hook (53fe0fcfdcdbc4fb9d424a805db86ffaba602115)

v0.9.24

0.9.24 (2026-05-29)

• feat(voice-graph): speak-only TTS delivery + injectable orchestrator/session (3bf2831854ecbffb6c15aade115965501520e6bb)
• Correct LLM provider count to 11 across README, docs, corpus, and tests (1e0c2f82cd2e479972444ecd39ca68e8f482ad91)
• Correct LLM provider count to 11 in architecture alt text (63ed327fe991cbf5fe1e01bca76416a3aaa76167)
• Rewrite README hero and fix LLM provider count to 11 (5024065f1be17885edd02d2003cff15d51485535)

v0.9.23

0.9.23 (2026-05-28)

• fix(generateObject): route through buildResponseFormat for anthropic + gemini (143d5117cc550c058eef4244b8e8c077c771a52c)

v0.9.22

0.9.22 (2026-05-28)

• fix(graphrag): use ON CONFLICT instead of INSERT OR REPLACE so Postgres adapters can persist (4bb7186dc385fcd9671769ef4e0e7a2be1a060ee)
• docs(memory): sync multimodal-rag + working-memory updates (b847096d91ff28d174bbf4c172fe565834257236)

v0.9.21

0.9.21 (2026-05-27)

• feat(rag): export PostgresVectorStore from cognition/rag public surface (f55db21ccb2fc05134c086d12b15af8de4f7ae24)
• docs(readme): add forge demo hero (gif + webp fallback) (2146cfe985426cfb1997b8b47707109422527c1c)
• docs(readme): trim filler tail on forge demo caption (7273b59cde46a46d66151ea840fbf3c5aee1b1c0)
• rag: add fetchByIds to IVectorStore + PostgresVectorStore; hydrate sparse-only HybridSearcher winners (1ab0da74bfd3c71b3763b41615e394dd10af9adb)
• Merge branch 'master' of https://github.com/framersai/agentos (775e73921dfe04efbbbd8601c04d3ac70d542edd)
• rotate Discord invite and route hardcoded URLs through wilds.ai/discord (7b13f2b1cb235f76c991e04b4fa593ea71acd331)

v0.9.20

0.9.20 (2026-05-22)

• fix(generateText): match 'credit balance' billing message in isRetryableError (6060864a393103bf4d6709ab5a0df9467b3bb293)

v0.9.19

0.9.19 (2026-05-21)

• test(voice): use AudioFrame.timestamp (not timestampMs) in STT regression test (dd7752f35df767d3f7287f46b941cda7d4c8d0b5)
• fix(voice): elevenlabs chunked STT multipart field name=audio → name=file (4061bfb53c4b06e48a2887acea6e044e8a25f6a2)

v0.9.18

0.9.18 (2026-05-18)

• fix(voice): bound batch TTS synthesize fetches with AbortSignal (80b51fb4cbba7e39fb07c415488688fd2d193cac)