Changelog

Highlights of user-facing changes, new features, and notable fixes in past releases. Every published package also ships its own CHANGELOG.md with the complete record for each version — for example @robota-sdk/agent-cli and @robota-sdk/agent-core. The commit history is on GitHub.

2026-11-10 — 3.0 to 3.3

New features

Read-only observer for a served session (3.3.0) — robota --serve --observer-port <port> opens a second loopback listener with its own token (ROBOTA_OBSERVER_TOKEN, at least 32 characters, different from the driver and HTTP tokens). An observer reads the conversation and state; it cannot submit, answer prompts, run commands or switch sessions, and its disconnect does not end the session. A secure runtime and a hosted task worker refuse the flag.

Secure conversations, opt-in (3.2.0) — The desktop shell opens a certified runtime for secure conversations, and robota secure setup, status, enable, disable, lock, verify and change-passphrase manage the keyring they are sealed to. Secure mode stays off until you enable it.

Fork and daemon commands (3.2.0) — Follow a running fork without stopping its parent, steer its job and continue its record after it ends. robota daemon clear-start-lock clears an abandoned daemon start lock.

Fixes

• Edit checkpoints are serialized per session and claimed across processes, and a session that closes normally keeps its record claim until it is down (3.3.0).
• A secure runtime launch that fails after the runtime was proved returns only once that runtime is gone (3.2.2).
• The workspace root broker exits once the user-state directory it serves is removed or replaced (3.2.2).

Breaking (SDK)

Explicit project access (3.0.0) — Project filesystem access is a host-issued authority, not an ambient consequence of cwd. A caller that supplies no projectAccess runs Restricted: it gets no project filesystem capability and does not load the project's instruction files or settings.

Merged SDK packages (3.2.2) — The CLI and the desktop app import providers, host primitives, transport carriers, the default tool set, product assembly and the shared contracts from the subpaths of @robota-sdk/agent-provider, @robota-sdk/agent-os, @robota-sdk/agent-transport, @robota-sdk/agent-tools, @robota-sdk/agent-product and @robota-sdk/agent-interface. The command line, the settings and the behavior are unchanged. The migration guide lists the import paths.

2026-07-06 — Beta 77–79

New features

/workflows create — natural-language workflow authoring — The /workflows CLI command can now author a workflow from a plain-English description. robota → /workflows create "fetch the issues, summarize each, then post a digest" asks the active provider to design the workflow, saves it as a reusable .workflows/<name>.json artifact (with any prompt-backed nodes under .workflows/nodes/), and runs it immediately. At this release, create was model-invocable. The current CLI requires the user to invoke the command and confirm the write. The command surface at this release was create "<desc>" [--input key=value] [--name <name>] | list | catalog | validate <file> | run <file> — list shows available workflow node kinds, catalog lists saved workflow files in .workflows/, and validate / run operate on a workflow file.

Internal

Self-contained agent-cli bundle — @robota-sdk/agent-cli publishes as a self-contained bundle: the @robota-sdk code it uses, including the /workflows command, is bundled into dist, and the published dependencies contain only third-party npm packages.

Instant-node provider and persistence round-trip — Instant (prompt-backed) workflow nodes resolve their provider from a single source of truth and survive a symmetric save/load round-trip, so an authored workflow reloads with the same provider wiring it was created with.

Single-OTP publish — The release/publish flow detects already-published versions per package version and completes a multi-package publish with a single OTP entry instead of prompting repeatedly.

2026-06-14 — Beta 68–76

New features

@robota-sdk/agent-session-analytics — A package that analyzes session logs to report timing breakdowns (LLM wait vs. tool/code time, slow intervals). Session-log analysis lives in this dedicated package instead of the agent-cli thin shell.

Internal

Design-quality audit remediation (SSOT consolidation, package cohesion, error/observability hygiene) and a round of harness/rules hardening. No runtime behavior change for existing apps.

2026-05-23 — Beta 67

New features

Plugin development guide and directory — A new Building Plugins guide covers the full lifecycle for writing, testing, and publishing community plugins. The Plugin Directory page lists official plugins with a process for submitting community plugins.

robota init — project initialization — Running robota init in any directory creates a starter AGENTS.md and <project-state>/settings.json. If a .claude/ directory is detected, it offers to migrate Claude Code settings automatically.

Local LLM support guide — A dedicated Local LLM Setup guide covers Ollama, LM Studio, and llama.cpp with per-model recommendations. No API key needed for any local inference server.

Context warning banner — The TUI now shows a yellow warning at 70% context usage and a red alert at 90%, with a prompt to run /compact.

Improved /compact output — Running /compact now reports how many messages were removed, the percentage reduction, and before/after context window usage.

3-level permission memory — The permission prompt now has a third option: "Allow always (this project)." Choosing it writes the tool pattern to <project-state>/settings.local.json so the permission persists across sessions in that project.

2026-05-10 — Beta 60–66

New features

maxSameToolInputs safety limit — Set maxSameToolInputs on a ConversationAgent to automatically abort a run if the same tool is called with identical inputs more than N times. Prevents infinite retry loops.

Fixes

Session log correlation — Tool calls in session logs now carry consistent correlation keys, making replay-grade reconstruction reliable.

2026-05-02 — Beta 59

See the full release notes for Beta 59 for a detailed breakdown of all changes in this release window, including:

• Subagent background jobs with transcripts, watchdogs, and command surfaces
• Provider composition: Qwen, Gemma, Gemini, Anthropic, OpenAI, and OpenAI-compatible in one config
• Session log events for replay-grade provenance
• Deterministic parallel Agent tool calls via jobs
• CI/deploy build reuse — single monorepo root build, no per-package rebuild loops

Version history

Stay updated

• Watch GitHub releases for tagged versions
• Subscribe to npm for package updates