Skip to content

Releases

v0.3.1

Choose a tag to compare

@lenitain lenitain released this 20 Nov 10:06

[0.3.1] - 2026-11-20

Added

  • Optional auto-submit: autoSubmit (default false) fills a ready suggestion into the prompt and submits it like Enter after autoSubmitDelayMs (default 30000, 1000–600000 ms) without keyboard activity. Any keypress or paste cancels the pending submit while keeping the suggestion available; the timer also skips a prompt that is unfocused, non-empty, in shell mode, or blocked by a dialog or queued message

v0.3.0

Choose a tag to compare

@lenitain lenitain released this 01 Nov 04:02

[0.3.0] - 2026-11-01

Changed

  • Suggestion lifecycle: a prediction is made when the session is idle, no permission or form dialog is waiting, and no message is queued. Typing, focus changes, and shell mode only hide the suggestion; they do not cancel an in-flight call or discard a ready one. Accepting fills the input without consuming it, so deleting the accepted text reveals the same suggestion again. The suggestion is keyed by its turn, not by what the user did; a new turn, revert, compaction, or session switch discards it
  • Situation context instead of a flat transcript: the prompt now carries the recent exchange with the latest assistant reply kept at full fidelity, older assistant turns reduced to their final prose paragraph, fenced code collapsed to a [code: N lines, language] fact, and the last completed tool's name and outcome (input/output only with includeToolContext). The latest assistant reply's ending is the one invariant the budget never cuts, because it holds the question, proposal, or result the user is about to answer. There is no session origin anchor: the first message is often a dead topic, and stale context would pull the prediction backwards
  • One call, no retry, no fallback: predictions run once in a disposable, tool-denied child session and are removed afterwards. Timeouts abort the call instead of retrying; a failed helper creation no longer falls back to the active session or to the global generate.text (which cannot resolve project-scoped models)
  • Output: a prediction that exceeds 170 characters is cut at a clause boundary; if no boundary exists, no suggestion is shown. The prompt asks for NO_SUGGESTION when the next message is not strongly implied, because silence costs less than noise
  • Cleanup: helper removal is idempotent, bounded, and also runs when a helper is created after the deadline or aborted mid-flight

Removed

  • The 800 ms retry, the timeout retry budget, the runtime version-warning toast, and the active-session/global fallback chain
  • Older-host compatibility: the OpenTUI 0.5 getLayoutNode().markDirty() fallback, the OpenCode 2.0.x session.idle event without a session id, the guards for a missing session.create/generate/remove, and the shell placeholder heuristic for hosts without a keymap mode. The plugin now requires OpenCode ≥ 2.1.4 (OpenTUI ≥ 0.6)

Docs

  • Documented that OpenCode Zen free models (opencode/*-free) cannot power predictions: the upstream free-tier gateway rejects the tool-stripped prediction request with a misleading Free usage limit reached error (anomalyco/opencode#49723), so a non-free prediction model is required

Testing

  • Rewrote the conversation, predictor, and plugin suites around the new contract: tail preservation, tiered retention, code collapse, tool facts, dialog/queued blocking, suggestion preservation across typing/focus/shell, accept/deletion round-trip, no retry, and helper cleanup