Skip to content

v0.6.0

Choose a tag to compare

@beonautobot beonautobot released this 12 Oct 23:26
Immutable release. Only release title and notes can be modified.
2bdbe07

0.6.0 (2026-10-12)

⚠ BREAKING CHANGES

  • server: npx @on.auto/runtime starts the runtime, on port 5555 by default (#163)
  • global: the runtime is @on.auto/runtime and every package is under @on.auto (#155)
  • global: the runtime is named auto-runtime (#153)
  • global: playbooks are the one name in storage, configuration and code (#151)
  • global: align runtime APIs and documentation around playbooks

Features

  • api: a browser opening the server finds a page that opens the console (#92) (7dc46cb)
  • api: authenticate callers and serve the brain operations over HTTP (#66) (4d744b0)
  • api: instructions, tool descriptions, guides and prompts for agents over MCP (#126) (f4a7d7f)
  • api: lead MCP tool results and errors with plain words (#83) (1021942)
  • api: serve every operation on one MCP endpoint (#74) (7b1d79f)
  • api: serve operations as MCP tools (#69) (266df06)
  • brains: add the brain operations and the ledger that stores them (#64) (3d0cfe5)
  • computation: a brain computes with a program, exactly and the same way every time (#106) (99b59df)
  • definitions: one answer fact, facts composed from every capability, an input-aware contract (#147) (db609dc)
  • global: add README quickstart and Studio invite links (#95) (68b0f3b)
  • global: align runtime APIs and documentation around playbooks (4e57fe2)
  • global: rebuild the runtime as a Node 26 TypeScript monorepo (#46) (405d04b)
  • inference: call what a brain does by its functions (#84) (7eaf4dc)
  • inference: execute prompts as specs (#71) (fda92e7)
  • inference: list the models a server can call (#87) (90f3cde)
  • inference: tell callers which model providers a server has (#79) (4d522f6)
  • interaction: a brain asks a person or a system and takes the answer later (#123) (b4512c0)
  • interaction: an interaction function calls a tool and answers with its result (#139) (d6998f1)
  • interaction: an interaction function names the tool it sends through and reads its replies (#135) (af87111)
  • interaction: an open request lists the shape its answer takes (#129) (04201ab)
  • ledger: content kept as streamed bytes, and operations served over HTTP alone (#146) (f78ab74)
  • ledger: keep the ledger in PostgreSQL (#86) (cbf6c4a)
  • ledger: named facts, a context typed by each fact, and content kept whole under its digest (#143) (12f5816)
  • ledger: open the ledger on any Emmett SQLite driver, without sqlite3 (#82) (a4c766a)
  • ledger: read what a brain recorded, bound to the brain (#89) (618517a)
  • mcp: a server's tools are allowed and marked testable on its own entry (#134) (9203df9)
  • mcp: an agent tests a tool call through the brain before writing a function (#131) (498219d)
  • mcp: list the tool servers a brain may use, with the tools each offers (#119) (d71c0cd)
  • mcp: reason functions use tools from the MCP servers the operator configures (#98) (f86bd00)
  • object-storage: content kept on an S3 object store behind CONTENT_STORE_URL (#150) (a400df9)
  • operations: add the application layer (#61) (c8c4a4a)
  • orchestration: a workflow starts on a schedule and on events, with up to three triggers (#130) (a06e969)
  • orchestration: a workflow waits for runs that finish later, with deadlines and cancel (#115) (48d90dd)
  • orchestration: run workflow specs on Temporal (#70) (b0faa0f)
  • prediction: the prediction function's document, its schemas and its dataset from history (#157) (4f4c3b3)
  • recollection: a brain keeps views of its own history and answers from them (#112) (bd334af)
  • server: npx @on.auto/runtime starts the runtime, on port 5555 by default (#163) (0205f63)
  • server: read structured settings from a YAML configuration file (#80) (1460def)
  • server: readable start-up and a quick start for AI assistants (#77) (a9fcb65)
  • server: run a local Temporal with pnpm dev (#75) (be0b02a)
  • server: serve workflows (#72) (aecb3f9)
  • server: the database password signed from the cloud identity under DATABASE_CREDENTIALS (#160) (1b17197)
  • specs: add the spec operations (#67) (7651e40)
  • specs: answer a brain's analytics from a table of run outcomes kept in the ledger (#105) (f71f442)
  • specs: every event names itself, its cause and its run, and a workflow's steps are events (#107) (adf527b)
  • specs: list a brain's runs, read a run's history and follow a brain's events (#93) (2019e5b)
  • specs: publish events into a brain and present its own facts as events (#104) (77e8f8f)
  • workflow-engine: the brain's one language is TypeScript, run in a QuickJS sandbox (#141) (5791aff)
  • workflow-engine: the contract of a workflow engine on the ledger (#85) (a862223)
  • workflow-engine: the workflow machine, its engine and an in-memory driver (#88) (61d3ade)
  • workflow-host: a projector folder per definition type, and a pool of model processes (#148) (3281b17)
  • workflow-host: run every workflow in the server, on the workflow engine, without Temporal (#96) (13f4a7e)
  • workflow-host: workflows react to what their brain records, on events and on a schedule (#113) (41baa13)

Bug Fixes

  • api: judge the arguments of an MCP tool call as HTTP does (#76) (b07642d)
  • api: the served tools advertise no output schema (#136) (b140126)
  • api: the server's page opens the studio on this server again (#128) (fad12d6)
  • api: the server's page opens the studio, whose origin is the one always allowed (#94) (5765599)
  • global: remove the duplicate recollation folder and third-party references (#57) (28f3184)
  • global: settle three tests, show a rejected run's record, and instruct connecting agents (#118) (4f99a5e)
  • global: settle two flaky tests and bound what tool servers log (#103) (55f1360)
  • interaction: answers survive a restart, timers run while receivers hang (#125) (c4d01f9)
  • ledger: the run list fills every page when filtered by primitive or name (#127) (6c9fd7e)
  • mcp: list only reachable tool servers, refuse an unknown server, and decode MCP inputs once (#120) (73cb2ab)
  • mcp: the tool-server listing answers for the org when no brain is named (#132) (50fb66a)
  • operations: a key limited to some brains is told to name one (#133) (c437767)
  • operations: make UUIDv7 ids portable to Cloudflare Workers (#81) (c5a48c4)
  • prediction: the data test in one session a page, the brief checked at save, reads bounded (#159) (2e09885)
  • server: computation and recall run under the development runner again (#121) (40e4f82)
  • specs: bound nested event data and close the gaps a published event could use (#108) (5584e7f)
  • specs: keep the deferral out of the history, and cap how deep an expression's value may nest (#110) (14c9774)
  • workflow-engine: bound regular expressions and object keys, and give expressions a deadline (#99) (749bda9)
  • workflow-engine: workers inherit no flag from the server (#122) (4441b3e)
  • workflow-host: a lone recall function is never shown waiting behind others (#137) (3df2373)

Performance Improvements

  • ledger: size each fill listing by what the batch before showed fits in 16 MiB (#111) (673b4ab)
  • workflow-engine: keep workers warm between jobs, so a run costs only its own work (#117) (3a47712)

Code Refactoring

  • global: playbooks are the one name in storage, configuration and code (#151) (79fb090)
  • global: the runtime is @on.auto/runtime and every package is under @on.auto (#155) (6e04f07)
  • global: the runtime is named auto-runtime (#153) (54b7ae6)