Highlights

• Treat everruns coordination threads as yolop processes and close the thread follow-ups.
• Seed connector defaults from the upstream catalog and register E2B tools.
• Default to the latest model on every provider.
• Search with the built-in rg fallback on machines without ripgrep, and ask structured questions across hosts.

• Code search works on machines without ripgrep
• Coordinators run threads as separate processes
• New setups default to the latest model on every provider

See the GitHub release and the full release notes for every pull request in this version.

A model saved in your settings, or passed with -m, is unaffected.

See Providers and models and #834.

Enable coordination in a profile and start a session with it. The TUI opens a THREADS board in the activity rail; Enter on a row opens that thread, and quitting it returns to the coordinator. A thread’s sandbox escalations are asked in the coordinator’s approval prompt. yolop sessions live lists running sessions and threads.

See Session coordination, #835 and #837.

yolop utils rg -n hello .
yolop utils rg --files -g '*.txt' .

When no rg is found on PATH at startup, the agent also gets an rg tool on the same engine: read-only, scoped to the workspace and with bounded output. Machines with ripgrep installed keep using it.

See Tools and #839.

[[capabilities]]
ref = "native_edit_tools"

Every yolop subcommand’s help now has a description and a worked example, and the system prompt prefers well-known CLIs such as git and gh over improvised scripts.

See Tools and #827.

Highlights

• Prefer known CLIs for integrations, add runnable help examples, and offer opt-in model-native apply_patch.
• Report live program status in the TUI and keep modals aligned with the active theme.
• Recover unfinished work across hosts and resume interrupted ACP response bodies once.
• Improve Codex authentication refresh, provider failure messages, and recovery from unknown tools and progress-guard rejections.
• Upgrade the Everruns crate family to 0.44.0 and Tuika to 0.13.0.

• Opt-in apply_patch editing for OpenAI models

See the GitHub release and the full release notes for every pull request in this version.

Highlights

• Move Yolop to the published Everruns 0.41 crate family and remove direct dependencies on retired shims while preserving runtime capabilities.
• Add GPT-6.1 speed tiers, GPT-6 and Claude 5.x model options, and open-source Sign in with ChatGPT for the Codex provider.
• Strengthen approval, extension, filesystem, image, and workspace boundaries.
• Improve ACP, TUI, and runtime behavior, including retries for expired response chains.

• GPT-6.1 Sol, the GPT-6 family and Claude 5.5 models
• ChatGPT sign-in through the open-source route

See the GitHub release and the full release notes for every pull request in this version.

A browser sign-in for --provider codex now uses OpenAI’s open-source Sign in with ChatGPT by default. Yolop registers its own client for your account on the first sign-in and runs turns on the public Responses API, still billed to your ChatGPT plan. It is a preview route, so a few request options (output token cap, temperature, speed) do not apply.

To keep the Codex login, set the route and sign in again:

yolop config set chatgpt_sign_in codex

YOLOP_CHATGPT_SIGN_IN=codex does the same. A device login uses the Codex route unless open-source is set.

See Providers and models, #806 and #808.

A new speed setting picks OpenAI’s service tier for models that offer one, and a tier the model does not offer is skipped:

yolop config set speed fast

YOLOP_SPEED sets it for one run.

See Providers and models, #804 and #805.

Highlights

• ACP now advertises user-invocable skills to clients, and announces finished background work even when the wake turn stays silent.
• Action-guard gains a Muse-only idle-promise guard on the Jev classifier.
• MCP hardens workspace execution by blocking stdio process spawning and requiring explicit environment credential bindings.
• Observability and provider handling improve: Logfire forwards full LLM cost metadata and span duration, and Codex handles cross-provider reasoning.

See the GitHub release and the full release notes for every pull request in this version.

Highlights

• Coordinator staffing: yolop coordination spawn-workers prints validated per-worker launch commands, and yolop_spawn routes agent versus session spawns.
• Setup can log out: leave model setup without saving, and drivers log out on exit.
• Settings and connections hot-reload; integration tests use isolated spawns.
• TUI adopts tuika 0.12 ToolShell and drops legacy shell paths.

See the GitHub release and the full release notes for every pull request in this version.

Highlights

• Setup validates provider auth before completing model selection (#709)
• CLI attaches through shell descendants so managed sessions stay reachable (#711)
• Muse xhigh is accepted on cold startup (#712)
• README rewrite and control-plane docs keep the public surface accurate (#707, #715)

See the GitHub release and the full release notes for every pull request in this version.

Highlights

• Named profiles can now be managed from the CLI and control plane (#698)
• Configured MCP servers are listed in the agent system prompt (#700)
• OpenRouter 402 billing pauses now guide users to wait or add credits, and validation errors are actionable (#701, #705)
• Compaction checkpoints stamp rows and trace Codex compaction (#703)

See the GitHub release and the full release notes for every pull request in this version.

• search_sessions is replaced by yolop sessions search and yolop sessions list. The session_history capability ID is now sessions.
• list_mcp_servers, set_reasoning_effort, search_models, set_model, set_provider and the connector tools are removed. Use yolop mcp, the provider and model commands, and the new yolop connectors list|get|connect|disconnect.
• Capability IDs drop the yolop_ prefix: yolop_approval is now approval.

Upgrade note: settings overrides that use the old capability IDs fail validation. Rename them to the bare IDs.

See CLI essentials, #691 and #693.

Highlights

• yolop sessions CLI replaces the search_sessions model tool, with search and list subcommands.
• Capability names unify behind CLI routes: yolop prefix stripped to bare IDs, MCP/provider/model self-configuration moves through CLI control routes.
• First-class env-context entries with --env-context flag and Paseo detection.
• OpenRouter attestation-gate guidance, account gates surfaced as links, plus yolop mcp login for remote OAuth servers.
• Root-managed cancellation with automatic conflict choice, and upstream v0.25.0 cone batch adoption.

• Manage and log in to MCP servers from the CLI
• Session search and self-configuration move to CLI commands

Before upgrading, read the breaking changes: Session search and self-configuration move to CLI commands.

See the GitHub release and the full release notes for every pull request in this version.

yolop mcp login <name> runs the OAuth flow for a remote server: it prints the authorize URL, opens the browser, waits for the callback and saves the tokens. Before, that flow existed only as /mcp login inside the TUI.

yolop mcp login linear

See MCP servers, #661 and #677.

Highlights

• Fullscreen turns compact their activity by default. --no-compact-work keeps the expanded narration and tool rows for anyone who prefers the old projection.
• yolop mcp is a complete configuration surface: mcp show <name> inspects global, workspace, or effective scope, and disable is its own command instead of enable --disable.
• Fresh cargo install yolop works again. Every everruns-* requirement is pinned exactly, so a from-scratch resolve picks the set that was actually built and reviewed, and that set is now the everruns 0.20 facade release.
• Reasoning effort resolves from every source that knows about it, curated registry, then provider-advertised parameters, then the reasoning-required families, so a model the registry has never seen keeps its /effort control and repairs a mandated-reasoning rejection once on its own.
• The runtime stops doing work twice: a background task the foreground turn already reported no longer wakes a duplicate turn, and the environment context reports the shell’s real network access instead of a stale no-network claim.
• web_fetch reaches public URLs again, and MCP restart/reconnect/reload land on the typed mcp reload command.

• Turn activity compacts into one summary row by default

See the GitHub release and the full release notes for every pull request in this version.

To keep the expanded transcript, start Yolop with:

yolop --no-compact-work

The mode was opt-in with --compact-work in 0.17.0 and became the default in 0.17.3.

See CLI essentials, #612 and #658.

See the GitHub release and the full release notes for every pull request in this version.

Highlights

• Releases ship two binaries per target. yolop-<target>.tar.gz is the default-features build the Homebrew formula points at, and no longer carries the local-inference engine; yolop-<target>-metal.tar.gz and yolop-x86_64-unknown-linux-gnu-cuda.tar.gz carry the engine with the backend that target can use. The Linux tarball drops from 29.9 MB to 18.2 MB. Extension servers are now built for every target too, not just the last one the matrix collapsed onto.
• The runtime recovers from failure loops instead of repeating them: consecutive equivalent invocation, missing-command, wrong-path, and usage failures on an unchanged workspace now require a different action or a bounded progress checkpoint, and a checkpoint is a real trajectory transition rather than a hook-only backstop.
• File discovery is batch-native. read_many_files is advertised eagerly, so a model can read independent known paths in one call even when the provider emits one tool call per turn, and repo_map keeps its compact parameter schema in the first-turn profile.
• Input routing is one ordered ownership table, so a paste while the sandbox-approval prompt owns the screen no longer lands in the composer behind it.
• ACP prompts continue once after the provider exhausts its stream-stall retries, replaying from the durable transcript so settled tools are not re-run.
• The everruns family moves to core 0.19 and provider 0.20, where reasoning is ordered ContentPart::Reasoning artifacts rather than a flat message field, and tuika to 0.11.0.

• Run models on your machine with the local provider

See the GitHub release and the full release notes for every pull request in this version.

yolop weights pull
yolop --provider local

The provider is experimental. The engine ships in separate release archives, yolop-<target>-metal.tar.gz and yolop-x86_64-unknown-linux-gnu-cuda.tar.gz. The default archive, which Homebrew installs, leaves it out, so the Linux tarball dropped from 29.9 MB to 18.2 MB. From source, build with the metal or cuda feature.

See Providers and models, #581, #599 and #623.

Highlights

• Compiled extensions install end to end. A package may declare capabilityServer.binaries, and yolop extensions install fetches the prebuilt server for the host’s target triple, verifying its .sha256 before writing anything. yolop-extension-logfire 0.2.0 is the first extension published with prebuilt servers.
• The Logfire exporter emits a real trace tree: spans nest by the host’s parent_span_id, carry the Gen-AI semantic conventions (gen_ai.request.model, token counts, tool names), and give thinking its own paired span.
• Model switching serves a curated, cross-provider list the user owns, administered by yolop models list|add|rm|move|use|reset, with “Browse all models…” still reaching the per-provider catalog.
• An opt-in --compact-work mode keeps each active turn on one mutable summary row, expandable with Ctrl+O.
• The global config and data directories are relocatable through --config-dir / --data-dir or YOLOP_CONFIG_DIR / YOLOP_DATA_DIR, so isolating a run no longer means moving HOME.
• Local inference gains metal and cuda backends and a default model that runs at small-model speed (Qwen3-30B-A3B-Instruct, Q4 GGUF), both gated in CI.

• Config and data directories can be relocated
• Model switching uses a curated list you manage
• Compiled extensions install from prebuilt binaries

Before upgrading, read the breaking changes: Model switching uses a curated list you manage.

See the GitHub release and the full release notes for every pull request in this version.

The archive’s .sha256 file is required: an archive without one, or with a mismatched checksum, is refused and nothing is written. yolop-extension-logfire 0.2.0 is the first extension published this way.

See Extensions and #620.

Upgrade notes: the local weights commands moved to yolop weights, so yolop models pull is now yolop weights pull. The per-provider models config key is now default_models; an existing models table is still read.

See Providers and models and #618.

Highlights

• Sessions can hand work to each other: the built-in session_coordination capability discovers opt-in local sessions in the same project, reserves one, dispatches a durable task, and wakes the coordinator on completion, all through yolop coordination.
• Extension and coordination administration moved from model tools to an attached CLI. yolop extensions ... and yolop coordination ... run against the live session over one-shot anonymous pipes, so enable, disable, and reload affect the session you are in without spending model context on tool schemas.
• Profiles now define a whole agent, not just execution settings: capabilities, mcp, skills_dir, and standing instructions are all selectable per run.
• run_command covers the entire command registry on every host. The nine-command allowlist is gone, and ACP and --print sessions get command dispatch for the first time.
• An experimental in-process local provider runs a model inside the binary, with yolop models pull owning the weights. Off in default features so source builds stay fast; the release binaries carry it.
• Memory is scoped: remember, recall, and forget take global (the default) or repository, with repository memory shared across linked worktrees.
• Editor integration widens: yolop into buzz writes a Buzz Desktop harness, and ACP clients can be handed an opt-in loopback setup page (--acp-setup-page) so API-key providers have an in-conversation path.
• Copying wrapped text out of the transcript no longer inserts newlines that were never there, on tuika 0.10.

• Set up as a Buzz Desktop harness in one command
• Extension management moves to the extensions CLI
• Profiles can define a whole agent
• Local sessions can hand work to each other
• Memory can be scoped to a repository

Before upgrading, read the breaking changes: Extension management moves to the extensions CLI.

See the GitHub release and the full release notes for every pull request in this version.

The agent picks the scope when it calls the tool, and the system prompt labels global and repository memory separately. Any other scope value is rejected.

See Skills, memory, and hooks and #584.

yolop --config-dir ./tmp/yolop-config --data-dir ./tmp/yolop-data

Isolated runs, or several Yolop identities on one machine, no longer need a different HOME.

See CLI essentials and #600.

Agents and people drive it with yolop coordination (list, status, dispatch, complete, accept, drain), which acts on the live session, and a matching /coordination command.

yolop coordination list --json

See CLI essentials, #588 and #594.

• capabilities and capabilities_mode: capability entries layered after the global list or replacing it, including which extensions are enabled.
• mcp and mcp_mode: MCP servers merged by name over the global set, or replacing it.
• instructions or instructions_file: text appended to the system prompt.
• skills_dir: an extra skills scope, by default profiles/<name>/skills/.

Start a session with yolop --profile <name>. Credentials stay global.

See Session coordination for a profile example and #586.

yolop extensions install crates.io:yolop-extension-lsp

/extensions, and the commands and tools that installed extensions contribute, are unchanged.

See Extensions and #587.

yolop into buzz writes a custom harness file for Buzz Desktop, next to the existing yolop into zed and yolop into paseo commands. Buzz then starts Yolop over ACP.

yolop into buzz

Running the command again keeps your env values and any extra fields in the file; --force replaces the file completely.

See In your editor, over ACP and #579.

Highlights

• OpenRouter can be authenticated from the browser: --provider openrouter runs a PKCE login instead of requiring a hand-pasted key.
• Terra and Luna model variants join the OpenAI and Codex selectors, listed ahead of Sol.
• Malformed tool calls no longer reach execution, every provider’s model-authored arguments are validated against the authoritative JSON Schema first, and progress_checkpoint stays callable when the progress gate demands it.
• ACP recovers provider authentication mid-session rather than stalling on an expired credential.
• The tmux gallery render is now gated on every PR by a real terminal implementation, so the release spec’s terminal matrix reserves the human walk for GUI emulators.
• Tracing output no longer scribbles over terminal rows the TUI owns.

• OpenRouter connects with a browser sign-in

See the GitHub release and the full release notes for every pull request in this version.

Highlights

• Model selection is honest about what is reachable: ACP exposes only connected models, duplicate model names resolve uniquely, and provider model catalogs are listable over ACP.
• The Meta Model API joins the supported providers.
• Long-running turns behave better under stress, stall recovery shares the watchdog’s budget, blocked completion status stays hidden, and workspace grep scans stream instead of buffering.
• Terminal selection survives Ctrl and Ctrl+C, on tuika 0.8.0 with matching satellite crates.
• Skill management (search_skills / install_skill / delete_skill) is now actually enabled in the coding harness, matching what the docs already promised.

See the GitHub release and the full release notes for every pull request in this version.

OpenRouter can now be connected from the browser. In /setup, pick OpenRouter and then Sign in with browser: Yolop opens openrouter.ai/auth, waits on a loopback callback, exchanges the code for an API key and stores it the same way as a pasted key. ACP clients see the method as openrouter_browser.

A pasted key and OPENROUTER_API_KEY still work, and the environment variable wins when both are set.

See Providers and models and #567.