ci: release packages

[github-actions]

This PR was opened by the Changesets release GitHub action. When you're ready to do a release, you can merge this and the packages will be published to npm automatically. If you're not ready to do a release yet, that's fine, whenever you add more changesets to main, this PR will be updated.

Releases

echadron@0.32.0

Minor Changes

• #50 de0c08c Thanks @YaseenHQ! - Point out AGENTS.md files the agent has not read. Instructions load once, only along the project-root to cwd chain, so a tool reaching into a sibling directory works there without ever seeing its rules. When that happens, the tool result now carries a short reminder naming the file — once per file. Off by default; enable under Feature controls or with ECHADRON_EXPERIMENTAL_AGENTS_MD_REMINDER=1.

• #45 29e0da0 Thanks @YaseenHQ! - Add a --thinking flag for session-scoped effort overrides, including resumed sessions. Pass --thinking max to use maximum effort without changing the config.toml default.

• #48 32d3a6b Thanks @YaseenHQ! - Tell the agent when the date changes. The date is rendered into the system prompt once when a session starts, so a session left open past midnight kept reasoning against a stale date. A reminder now announces the new date at the next step instead of re-rendering the prompt, which would throw away the whole prompt cache.

• #45 29e0da0 Thanks @YaseenHQ! - Refresh xAI models from models.dev when opening /model so Grok 4.6 appears, add Inherit primary to /secondary_model, and let you click in the prompt to place the cursor. Pass --thinking and ←/→ still set effort.

• #56 1b9722b Thanks @YaseenHQ! - models.dev now updates every provider it knows about, not just those added by API key. Only the API-key catalog import recorded where a provider came from, so a provider created by an OAuth login — xAI, for instance — was skipped by every refresh and its model list stayed frozen at the moment it was added. Provenance is now inferred from the provider id when the catalog knows it, so new models reach you without waiting for an Echadron release. Providers models.dev has never heard of are left alone, and no extra catalog fetch is made.

• #47 63ba96d Thanks @YaseenHQ! - Re-baseline the vendored pi-tui on 0.84.1 and add an experimental fullscreen mode. Pick it under /settings -> Display mode (or set [tui] tui_mode = "fullscreen", or ECHADRON_TUI_FULL_SCREEN=1 for a one-off run) to run on the alternate screen, where the transcript scrolls inside its own view and the activity, todo, queue, btw and editor chrome stays docked at the bottom instead of flowing away through the terminal's scrollback. The screen is chosen when the TUI starts, so switching asks you to restart. Inline mode is unchanged and remains the default. The upgrade also brings fullscreen transcript search, a large reduction in per-frame allocation on the alternate screen, SSH-aware escape timing, and Windows input fixes.

• #45 29e0da0 Thanks @YaseenHQ! - Chad's face is now the live activity indicator. The welcome header lives in the transcript and scrolls away as soon as work starts, so its look/blink/wink was only ever visible before the first prompt. The same character now drives the activity line, and what it does follows what the agent is doing: eyes scan while thinking, blink steadily while composing, sit half-lidded while a tool runs, look straight at you while an approval is pending, and dart during a retry. The AgentSwarm progress row keeps its own indicator.

• #52 7cc95f7 Thanks @YaseenHQ! - Added a ReadDocument tool. Reads Word, PowerPoint, Excel, OpenDocument, RTF, EPUB, CSV and PDF files as Markdown, so the agent can work with documents instead of getting raw bytes back from Read. Conversion runs locally through @firecrawl/anydoc; nothing is uploaded. An optional dependency, matching how the native terminal helpers already ship: platforms with no prebuilt binary install fine and report that the file cannot be read, rather than failing the install or crashing.

• #45 29e0da0 Thanks @YaseenHQ! - Reject /goal complete until configured workspace commands exit 0. Set [goal_completion_gates] commands or KIMI_GOAL_GATE_COMMANDS.

• #45 29e0da0 Thanks @YaseenHQ! - Route subagents at any configured model, not just two roles. Agent and AgentSwarm now accept a [models] entry id for model in addition to "primary" and "secondary", so a task that needs vision, a long context, or a cheaper model can be sent to the model that fits it. The tool description lists what the session has configured, each with its resolved capability flags. Role names still win over an identically-named entry, and agent files can set a concrete model that takes precedence over model_preference.

• #53 5c54482 Thanks @YaseenHQ! - Agent and AgentSwarm accept a result_schema: a flat JSON Schema the subagent must satisfy by ending its final message with a matching JSON object. The parent then receives parsed objects it can merge — a swarm report of JSON instead of a pile of prose — while a child that fails to comply keeps its prose plus a marker, so nothing is lost.

• #45 29e0da0 Thanks @YaseenHQ! - Sync four upstream correctness improvements.

  Ctrl+C is responsive while a request is retrying: provider SDK clients are built with maxRetries: 0, because their internal backoff sleep never observed the turn's abort signal, so cancelling during a 429 or 5xx only took effect after the sleep elapsed — and those hidden attempts double-counted the engine's own retry budget.

  A session interrupted while the model was still streaming its reasoning no longer bricks: the assistant message left holding only a reasoning fragment is now sent with an explicit empty content, so strict Chat Completions gateways stop rejecting every later request in that session.

  Steering an idle session launches a turn instead of failing, and updates the session title and last prompt the way an ordinary prompt does.

  The built-in coder profile no longer carries Agent and AgentSwarm. coder is the profile subagents are spawned as, so a subagent could spawn its own subagents and make fan-out recursive. The main agent profile keeps both, and a custom profile that lists them still opts in.

  Retry status lines are also capped, so a provider that answers with a large error body (occasionally a whole HTML page) can no longer flood the activity line.

Patch Changes

• #45 29e0da0 Thanks @YaseenHQ! - Gate autonomous goal completion through a bounded, fresh-context verifier that inherits the active model and thinking effort, falsifies broad claims with structural and transitive-contamination probes, rechecks the last concrete rejection first, lands a tool-free verdict when the step ceiling is reached, and permits only one format recovery before failing closed.

• #45 29e0da0 Thanks @YaseenHQ! - Keep the end of long Bash output in the model context so test and compiler failures stay visible when the log is truncated.

• #52 7cc95f7 Thanks @YaseenHQ! - Stop repeating the file and change count above every Edit diff. The tool header already reads Used Edit (…/readTool.ts) · +1 -1, and the diff below it printed +1 -1 /full/path/to/readTool.ts again — two extra lines on the most common tool in a session, with the full path wrapping mid-word on a narrow terminal.

• #56 1b9722b Thanks @YaseenHQ! - Refresh existing models.dev-backed providers on startup and with echadron update --models. Newly published models and catalog metadata now appear without re-login or replacing the provider's saved credentials, default model, or explicit model overrides.

• #53 5c54482 Thanks @YaseenHQ! - Fix the fullscreen layout stranding a short conversation at the top of the screen. The transcript now sits against the input box, with empty space above it rather than a gap in the middle, and a squeezed terminal gives up transcript rows instead of dropping the activity and todo lines.

• #45 29e0da0 Thanks @YaseenHQ! - Report every gap an independent completion review finds. A verdict listing more than eight gaps was previously treated as malformed, which replaced the concrete findings with a generic "could not produce a valid verdict" message; the list is now truncated instead. Long Bash output in tail mode is also cheaper to collect — the buffer is compacted on a slack threshold rather than on every chunk, which was quadratic on exactly the chatty commands tail mode exists for.

• #52 7cc95f7 Thanks @YaseenHQ! - Wrap the goal panel's "no stop condition" line instead of truncating it. Every other row in that panel wraps to the panel width; this one did not, so a narrow terminal cut the sentence off mid-word.

• #45 29e0da0 Thanks @YaseenHQ! - Give Dark and Light gray chrome with a teal model name, magenta thinking/running, and gold plan/shell. Run /theme to switch.

• #45 29e0da0 Thanks @YaseenHQ! - Close the horizontal seam through Chad, the welcome character. He was drawn with foreground block glyphs, and a font only paints its own glyph box — so in terminals that add line spacing (Terminal.app among them) two stacked rows of █ showed a gap between them. Solid cells now also set the cell background, which the terminal fills edge to edge regardless of font metrics. The half-block edges and the quadrant that carves the eyes stay foreground-only so their empty parts still show through.

• #45 29e0da0 Thanks @YaseenHQ! - Chad now winks. Roughly once a minute a blink lands as a wink instead — one eye closes while the other keeps looking wherever he was already glancing, held a little longer than a blink so it reads as deliberate rather than a dropped frame.

• #45 29e0da0 Thanks @YaseenHQ! - Pin prompt-cache prefix stability in the test suite. Provider caching only pays off when each request extends the previous one without rewriting earlier history, and nothing checked that before — a regression would only have shown up as a cost and latency drift. Three cases now cover it: the steps of a tool-calling turn, consecutive turns, and a swarm-mode toggle, whose exit reducer pops a reminder back off the history. An opt-in live test additionally proves a real provider cache hit end to end.

• #45 29e0da0 Thanks @YaseenHQ! - Pulse thinking traces on a magenta rail, paint finished tool headers green or red, and show Thought for Xs when thinking ends. Restart the TUI to pick it up.

• #45 29e0da0 Thanks @YaseenHQ! - Animate Chad, the welcome character, with a periodic specular glint and idle eye look/blink. The highlight is narrow enough to read as travelling across the seven-cell logo, with the lower row catching the light a beat after the upper one. Restart the TUI to see it.

• #45 29e0da0 Thanks @YaseenHQ! - Spin the terminal tab title while the agent is working, and flash an attention mark when a prompt is waiting in an unfocused tab.

• #45 29e0da0 Thanks @YaseenHQ! - Stop listing configure-once options twice. /theme, /editor, /features, and /secondary_model no longer appear in / completion or /help — they are grouped under /settings, which now also offers Secondary model. /reload-tui is hidden as a subset of /reload. All five still work exactly as before when typed, so existing habits and scripts are unaffected. Session-level toggles you change per task — /model, /effort, /permission, /plan, /swarm, /usage — stay in the list.

• #45 29e0da0 Thanks @YaseenHQ! - Announce swarm mode through the context-injection layer instead of pushing and popping reminders directly. Entering or leaving swarm mode no longer mutates conversation history as a side effect, and a swarm-mode reminder lost to compaction is re-announced while the mode is still active.

• #52 7cc95f7 Thanks @YaseenHQ! - Report context windows at the size the provider states. Token counts were formatted in 1024-based units, so a 1M-token window displayed as "977k", 500k as "488k", and 200k as "195k" — every configured window under-reported by 2.4%. Tokens are not bytes, and context sizes are configured and advertised in decimal, so they are now formatted that way.

• #45 29e0da0 Thanks @YaseenHQ! - Restore the original rounded prompt box with a > cue, keep the slash-command menu outside that box, and stop painting sent user turns as a tight filled strip. Welcome no longer repeats the pitch, and the footer is one muted status line. Restart the TUI to see it.

@moonshot-ai/agent-core@0.16.1

Patch Changes

• #54 85786c1 Thanks @YaseenHQ! - Download ripgrep from upstream's GitHub releases instead of a mirror, removing a third-party dependency from the search tools' first-run path.

@moonshot-ai/agent-core-v2@0.3.1

Patch Changes

• #54 85786c1 Thanks @YaseenHQ! - Download ripgrep from upstream's GitHub releases instead of a mirror, removing a third-party dependency from the search tools' first-run path.

@moonshot-ai/kimi-code-sdk@0.16.1

Patch Changes

• #45 29e0da0 Thanks @YaseenHQ! - Stop listing internal agent profiles as delegatable subagents. Session summaries returned by resumeSession, forkSession, and reloadSession now filter profiles marked internal, so the built-in completion verifier no longer appears alongside coder, explore, and plan.