Milestones
List view
Maintainer decision 2026-11-04: answers from the natural-language database query and the knowledge base are measured, grounded, and say 'cannot answer' instead of guessing. Phase M9.1: measurement, grounding and repair for NL→SQL and KB answers.
No due date•0/19 issues closedGoal (owner 2026-11-02, moved out of M3): the Company OS can receive email: inbound mail, autonomy-gated, with credentials in the vault. Umbrella #13707. The mailbox screen #13714 stays in M3 with an honest empty state. Phases: - M8.1 (label phase:M8.1) Mailbox backend and integration.
No due date•0/7 issues closedVoice and speech work. Backlog by maintainer decision 2026-10-31: low priority, scheduled after M1-M6. Voice API security keeps its own priority.
No due date•2/59 issues closedOwner 2026-10-25: the backend stays responsive under load without a language rewrite. A slow path is fixed with a better algorithm first, then by moving CPU-bound work out of the single API process, and only then with native code once a profile shows it is needed. Phases (labels): M6.1 slow algorithms replaced, M6.2 CPU-bound work out of the API process, M6.3 runtime and language options (free-threaded CPython, native extensions, more indexed languages). Research: docs/research/RES_rust-usage.md
No due date•2/14 issues closedOwner 2026-10-25: adopt open-source and OSS-plan tooling for CI/CD and debugging. Phases (labels): M5.1 supply-chain security, M5.2 error tracking and debugging, M5.3 dependency management, M5.4 test analytics, M5.5 security review and audit tooling.
No due date•60/132 issues closedGoal (owner 2026-10-22, runs in parallel with M1-M3): one canonical implementation per concept across the repo: no symbol forks, no copy-paste clones, no duplicated config defaults, no parallel mirrors of deployed code. Covers backend, shared, infrastructure and tooling; frontend kit/cross-GUI dedup lives in M3. Exit: the symbol-fork baseline is empty or every remaining cluster is a recorded homonym/exception; the jscpd guard covers both GUIs and every tree with pins that only go down; every #13916, #18717 and #17312 child is closed with code evidence. Phases, in order (filter issues with label phase:M4.N): - M4.1 (label phase:M4.1) Land the in-flight dedup and KB quarantine fixes. Done when: Open PRs #19142 (in train #19191) and #19169 are merged, an MCP search with caller filters still applies the research quarantine (test), and one optimize_agents module remains - M4.2 (label phase:M4.2) Retire the deploy-side agent forks. Done when: Every node type has pasted host evidence of one agent deploy path with matching module sha256, and the legacy branch mirror is removed without breaking the live updater - M4.3 (label phase:M4.3) Guards see cross-tree forks. Done when: duplication-guard.yml has one cross-tree scope over both GUIs, plugins, backend, SLM and shared with re-measured pins, pins auto-lower on every drop, and a planted copy in any scope or an unguarded extraction fails CI - M4.4 (label phase:M4.4) One config source. Done when: A test finds 0 conflicting AUTOBOT_* defaults between planes, paths and Redis DB numbers each have one home, and the hardcoded-value baseline is content-keyed, keeps its comments and only shrinks - M4.5 (label phase:M4.5) Delete dead copies and facades. Done when: Every listed facade, dead unit file, zero-coverage module and duplicate workflow job is deleted or wired, with a recorded verdict each and no duplication pin raised - M4.6 (label phase:M4.6) Shared primitives in one place. Done when: Redis, KnowledgeBase, chunk identity, rate limiting, circuit breaker, LLM formatting and RBAC middleware each have one implementation in autobot_shared or a current row in ARCHITECTURE_EXCEPTIONS.md - M4.7 (label phase:M4.7) Central model, provider and HTTP constants, no fleet addresses. Done when: Model maps, provider default models and the Ollama host each resolve from one SSOT entry, raw HTTP sessions are under a guarded ceiling, and no fleet address ships in code, docs or the JS bundle - M4.8 (label phase:M4.8) Shrink the symbol-fork baseline. Done when: repo_tests/symbol_fork_baseline.json shrinks from 93 clusters, every listed cluster is merged or recorded with a reason, and the ratchet fails on any new cluster - M4.9 (label phase:M4.9) One definition per concept. Done when: Each concept in the canonical-source inventory has one definition or a recorded exception, the other copies import it, and each consolidation lands with a guard test - M4.10 (label phase:M4.10) One API surface. Done when: Every high-confidence duplicate route pair has a recorded verdict, the duplicate-route guard resolves include_router, every route is gated or recorded as public, and every uncalled client function is retired or given a route - M4.11 (label phase:M4.11) Safety net: tests and analysers reach the whole tree. Done when: Every test directory is collected by CI, the coverage step enforces the number it states, and the import-cycle analyser resolves relative imports and reports a measured count - M4.12 (label phase:M4.12) Single-source frontend dependencies. Done when: One workspace root with one lockfile declares each frontend dependency version once, and no package.json carries overrides or resolutions
No due date•391/871 issues closedGoal (owner 2026-10-22, runs in parallel with M1 and M2): the Refined UI redesign (#18237) is fully implemented and working in both apps: tokens and themes, shell, one shared kit and one ApiClient, dashboards, and every step-5 section, with no forked components between the two GUIs. The backend behind a screen may be partial, but the screen must handle a missing backend honestly (an empty or not-available state, never a broken page). Exit: every #18237 build step is closed with code evidence; the cross-GUI duplication guard is green; i18n has no hardcoded strings on the new screens. Phases, in order (filter issues with label phase:M3.N): - M3.1 (label phase:M3.1) Design docs, tokens and themes. Done when: #18238's acceptance boxes are ticked, all four themes switch in both apps from one token source, and the duplicate-token guard fails on a second base declaration (#17566 SLM tokenisation is done per screen in M3.6, owner decision 2026-10-22) - M3.2 (label phase:M3.2) Shell: top nav, UI mode, appearance, brand. Done when: Both apps use the new top nav, UI mode and profile appearance, and the brand mark replaces the stock logo, the social preview and the README hero - M3.3 (label phase:M3.3) Shared kit primitives. Done when: No component in either app re-implements a kit primitive (KIT_FORK_GRANDFATHER is empty) and the design-system registry in code names each kit owner - M3.4 (label phase:M3.4) One ApiClient and generated types. Done when: Each app has one HTTP client module with generated OpenAPI types, no production file calls fetch directly, and the codegen check is required - M3.5 (label phase:M3.5) Dashboards with movable widgets. Done when: Dashboard widgets move and resize by mouse and keyboard, each shows one header, and every figure is measured or shown as not reported - M3.6 (label phase:M3.6) SLM platform screens. Done when: Every server-wide screen exists only in the SLM app in the new design, a guard holds the boundary, the user app's Platform admin opens it, and screens without a backend show an honest empty state - M3.7 (label phase:M3.7) Company OS, Automation and BI screens. Done when: Every Company OS, Automation and BI screen uses the kit and the new design, approvals and workflow runs each have one component, approvals enforce expiry and ownership, and screens without a backend show an empty state - M3.8 (label phase:M3.8) Knowledge screens. Done when: Knowledge screens use one graph component and the new design, with upload and transcriber intake reachable from the UI - M3.9 (label phase:M3.9) Chat, Agents, Analytics and Settings screens. Done when: Chat, Agents, Analytics, Settings and Account screens use the new design, chat has a stop control and one tool-approval gate, and Usage is gated by role with an empty state - M3.10 (label phase:M3.10) New screens and cross-cutting close-out. Done when: Each step-6 screen ships with its backend or an honest empty state, every new-design view uses t() keys in all locales, the SFC clone pin is lowered, and every #18237 child is closed with code evidence
No due date•343/530 issues closedGoal (owner 2026-10-22, runs in parallel with M1 and M3): Company OS matches or exceeds the leading open-source agent-company orchestrators, so real company workflows can run on it end to end: create a company, hire agents into roles, set goals, assign tasks, heartbeats run on schedule, budgets are enforced, approvals gate actions, and every step is in the audit trail. Exit: that workflow runs end to end on a live install with host evidence; every capability gap from the parity review is closed or explicitly ruled out by the owner. Phases, in order (filter issues with label phase:M2.N): - M2.1 (label phase:M2.1) Company and company boundaries. Done when: A company can be created, archived and restored from the UI, and a non-member never sees another company's agents, org chart, work items, work products or secrets (tests) - M2.2 (label phase:M2.2) Hire agents onto a working org chart. Done when: An agent hired from the UI can be heartbeat-enabled in the hire flow or agent settings, migrated CEO agents are schedulable, and every person and agent has a manager or is a deliberate root - M2.3 (label phase:M2.3) Goals and work hierarchy. Done when: Goals, portfolios, programs and boards can be edited and archived from the UI, and the work-item hierarchy rules (parent type, same company, no cycles) are enforced (tests) - M2.4 (label phase:M2.4) Tasks and company workflows. Done when: An agent checks out a work item and its comments can be listed, and a company workflow runs on the workflow engine against Company OS work items - M2.5 (label phase:M2.5) Heartbeats fire reliably. Done when: Every heartbeat-enabled agent stays on the heartbeat schedule after a backend restart, each dispatch writes a backend log line, and the run token reaches the agent - M2.6 (label phase:M2.6) Budgets enforced and costs recorded. Done when: An over-budget, paused or terminated agent is not dispatched (tests and host evidence), every LLM call records a cost event tied to its work item, and cost rollups never mix currencies - M2.7 (label phase:M2.7) Approvals and board controls take effect. Done when: A hire, strategy or budget-override decision takes effect only through its post-decision hook (tests), there is one approval system, and pause and terminate survive a reload on a host - M2.8 (label phase:M2.8) Audit trail and run supervision. Done when: Every activity record carries an actor type, a timed-out run shows in scorecards without error, a stalled run releases its claims, and the activity feed updates live - M2.9 (label phase:M2.9) External agents: protocol and adapters. Done when: An external agent gets a protocol skill, passes a sign-in and version check before dispatch, and its declared work scope is enforced on every path - M2.10 (label phase:M2.10) Portability, API docs, plugins and end-to-end proof. Done when: Export then import keeps the org chart and seed work, the Company OS routes have API reference docs, and a scripted run on a clean install passes company, hire, goals, tasks, heartbeat, budget stop, approval and audit
No due date•235/504 issues closedGoal (owner 2026-10-22, runs in parallel with M2 and M3): every supported install method - install script, deployment wizard / first-run setup, Docker, and the builtin updater - completes on a clean host with no manual fix-ups, ends secure (no default credentials, secrets in the vault), and reaches a working first login. User and operator docs describe exactly what the code does. Exit: each method passes on a clean host with host evidence; no open CRITICAL or HIGH install defect; the docs-drift checks are green. Phases, in order (filter issues with label phase:M1.N): - M1.1 (label phase:M1.1) install.sh completes, fails loudly and is safe. Done when: On a clean supported host, `curl -fsSL <raw install.sh> | sudo bash` completes with a terminal, with none, and on a multi-interface host; any failed health, HTTPS, login or registration step exits non-zero with no install marker; admin login and node registrat - M1.2 (label phase:M1.2) install.sh docs. Done when: Every native install step in README.md and INSTALL.md works verbatim, there is one requirements table pinned to install.sh's constants by a doc test, the pinned verified install matches what releases publish, the README anchor and version come from one source, - M1.3 (label phase:M1.3) One model map and the Local / Hybrid / Cloud resolver. Done when: The tier map has exactly one definition (guard test), every hosted provider resolves its default from one SSOT table, each node reports a hardware profile, and the resolver maps Local, Hybrid and Cloud to a full tier map (unit tests: no GPU, small VRAM, large - M1.4 (label phase:M1.4) Wizard and SLM: choose a mode, apply it, default models work. Done when: On a clean install.sh install, the mode picked by install.sh or in the wizard's configure_llm step is what provision_fleet applies (asserted on the generated extra_vars); a Local install on the stated minimum hardware pulls the default set and the first chat a - M1.5 (label phase:M1.5) Wizard docs. Done when: After install the docs give the chat UI URL and how to get its generated admin credential, document every wizard step (including configure_llm) with the real UI labels and the setup-reset token, and the user guide's quickstart and configuration pages describe - M1.6 (label phase:M1.6) Docker: secure defaults, honest startup and three modes. Done when: `docker compose --env-file docker/.env.docker --env-file docker/.env.secrets up -d --build` on a clean host pulls no autobot/* image, refuses to start while any credential is missing or default, Redis is loopback-only with a password, a failed migration or imp - M1.7 (label phase:M1.7) Docker docs and the docs close-out. Done when: Every documented install command for install.sh, the wizard and Docker works verbatim on a clean host (doc test), a first chat after following README alone gets an answer on Docker (host evidence), the port and health tables are per install mode, user-facing d
No due date•274/723 issues closedKB ingestion, watch folders, redaction chokepoints, retrieval, ChromaDB and vectorisation.
No due date•7/43 issues closedThe CI plane, hooks, ratchets, guards and the test harness. Largest chunk by count; no critical-tier work, so it runs behind the four above. SEVEN FIXES SPECIFIED FROM SOURCE, AWAITING A PUSHER (2026-09-28). Batch A, hook discipline, four files: #14144, #17575, #17614 (folds in #17120), #17071. Batch B, workflow exit-code discipline, .github is hub-exempt: #17674 (two instances) and #17700. Each specification is a comment on its issue and is written to be landed whole. Behaviour unverified on all of them — CI on the carrying PR is where they first execute, so none merges on the specification alone. NO commit trailer and no PR footer: a squash composes the body onto the commit (#16268). #17700 makes the weekly audit red on purpose; re-derive its 46 findings and decide the sequencing first. #17175 is an adjudication, not a fix.
No due date•16/88 issues closedRouters, models, startup, LLM providers and the API surface.
No due date•10/38 issues closedSecrets, RBAC, unauthenticated surfaces, firewall exposure, desktop and VNC access, approval gates.
No due date•9/58 issues closedSLM node lifecycle, provisioning, code_source delegation, ansible roles and the deploy path. Second-most critical, and the largest single seam.
No due date•4/44 issues closedAgent coordination, peer messaging, authority on relay, resource contention, and the LLC scheduler. Ordered first: carries the most critical-tier work in the milestone.
No due date•8/40 issues closedContainers, not work. An umbrella is delivered by its children; it can never close on its own, so it must not occupy a slot in a numbered chunk. Three of these do not declare themselves in the title.
No due date•0/13 issues closedWaiting on host evidence. Acceptance criteria that need a measurement against the running deployment, not the codebase.
No due date•0/4 issues closedBlocked by another issue. Carries a blocked_by edge or names its blocker. Not workable until that clears.
No due date•0/2 issues closedWaiting on an owner ruling. Not workable by a session. An issue leaves here the moment its decision lands.
No due date•3/19 issues closedRelease v0.16.0: the tail. Entry: the last consolidation defects, and capabilities whose blocker sits in an overflow version. Never ahead of an open blocker.
No due date•0/12 issues closedRelease v0.15.0: consolidation second pass. Entry: overflow from v0.12.0: defects in CI gates, the test suite and the deploy lifecycle, documentation gaps, remaining tech-debt. Never ahead of an open blocker.
No due date•12/133 issues closedRelease v0.14.0: wiring second pass, then consolidation begins. Entry: overflow from v0.11.0 (lower-ranked unwired, partially-delivered and wiring items), then the top of the v0.12.0 tech-debt overflow. Never ahead of an open blocker.
No due date•39/131 issues closedRelease v0.13.0: new capabilities, by umbrella. Entry: feature issues and umbrella epics with their children, and bugs in capability areas (company OS, knowledge retrieval, memory, code intelligence, workflow, voice). Never ahead of an open blocker.
No due date•13/116 issues closedRelease v0.12.0: consolidation, tech-debt and documentation gaps. Entry: tech-debt, canonical, refactor and guards issues, documentation gaps, and bugs in infra, deploy, CI and test components. Never ahead of an open blocker.
No due date•24/135 issues closedRelease v0.11.0: wire what's built, knowledge-access security, coordination by default. Entry: every actionable security-labelled issue and its open blockers; built-but-unwired or partially-delivered issues (ticked ACs, or a merged PR claiming to close them); bugs in security, access, coordination and wiring components; the knowledge-access (#16654, #16656, #16642, #16643), work-claim coordination (#15946's children) and offloaded-context (#16617) sets. Never ahead of an open blocker.
No due date•29/161 issues closedNot yet scheduled into a release (low-priority and unprioritized issues)
No due date•34/101 issues closedRelease v0.10.0: the next release (medium-priority work)
No due date•58/270 issues closed