Orca v1.2.1

1.2.1 - 2026-10-07

Changed

• Provider responses must now contain visible non-whitespace assistant text or at least one valid tool call; empty, whitespace-only, reasoning-only, usage-only, truncated, filtered, and contradictory completions fail visibly instead of silently ending a turn.
• Active provider generation now has a ten-minute total deadline in addition to the existing connection and inactivity timeouts.
• The total generation deadline is preserved across connection and stream_options compatibility retries; deterministic localhost coverage keeps the stream active while proving the original deadline still wins.
• Tool-loop finalization now has a distinct Finalizing safely state, retains the safety trigger reason, explains denied post-finalization tool calls accurately, and gives explicit non-replay continuation guidance.
• Raw file patches now verify retained proposal integrity before Apply and reread, hash-check, and revalidate exact destination bytes after replacement.
• Verified replacement now has explicit not-committed, committed, committed-with-cleanup-warning, and recovery-failure outcomes across file, scene, Input Map, main-scene, and ProjectSettings mutations.
• The About page can copy a strict-allowlist diagnostic report containing Orca/Godot versions and coarse provider/request state without credentials, endpoint addresses, project data, conversation content, model output, logs, or file changes.
• Plan turns can request one explicit, turn-bound transition to Work mode when the user asked for implementation or execution. Approval continues the same task with Work tools on the next provider request; rejection remains in Plan, and every project mutation still requires its normal separate approval.
• Runtime loop detection now ignores elapsed-time and observation-cursor churn, warns before the tool-round boundary, gives each observation a recommended next action, and performs one controller-owned cleanup of an exact same-turn Orca process before safe finalization.
• stop_game now returns bounded final run output, diagnostics, state, and verification evidence so routine cleanup does not require another diagnostics round.
• Multi-call tool batches are now limited to known independent read-only operations. Work-mode requests, task updates, reviewed mutations, process control, observations, verifications, and unknown tools require singleton batches and are rejected atomically otherwise.
• Successful task updates now refresh exactly one request-scoped checklist before continuation or finalization, while failed and cleared updates preserve or remove that context correctly.
• Completed tool turns that exceed the full-schema context budget now attempt one no-tools safe finalization; if that cannot fit, Orca emits a deterministic sanitized local checkpoint containing only bounded receipts, tasks, and coarse cleanup outcome.
• Automatic finalization cleanup is now bound to the exact active run ID launched in that turn, exposes only a fixed coarse outcome, and cannot stop a newer replacement run. Bounded stop_game evidence explicitly labels omitted diagnostics and truncated output tails.

Fixed

• Fixed DeepSeek-compatible reasoning-only or empty completions removing the working state without displaying an answer or error.
• Fixed empty tool-loop finalization responses ending without a visible outcome.
• Fixed file-patch recovery states hiding the guarded Revert action.
• Fixed committed writes being reported as failed when private-backup cleanup failed, and fixed uncertain replacement failures being reported or persisted as successful mutations.
• Cleanup-only warnings no longer weaken typed live-state conflict checks; recovery-copy guidance is retained as bounded project-relative session metadata.
• Fixed delayed stream and terminal callbacks from an older provider request being able to mutate a newer controller turn or chat UI state. Tool follow-ups now receive distinct provider request IDs while retaining one turn ID.
• Fixed synchronous workflow and terminal signal handlers being able to launch an orphan request or have an ending turn clear ownership belonging to a reentrantly started turn.
• Fixed plain Enter being silently swallowed while a request is active; follow-up drafts now remain editable and survive matching completion, failure, and cancellation while Stop remains explicit.
• Fixed Expanded Diff retaining exclusive window ownership while Apply or Reject synchronously triggered editor work, which could collide with Godot reload or confirmation dialogs.

Tests

• Added empty, whitespace-only, reasoning-only, tool-only, truncated, filtered, and terminal-ownership transport regressions.
• Added real GDScript patch lifecycle coverage and exact duplicate class/local variable validation regressions.
• Added tool-loop trigger, finalization-state, denied-provider-text, explicit-continuation, and empty-finalization coverage.
• Added deterministic replacement, restoration, temporary-cleanup, post-write verification, cleanup-warning, recovery-classification, and retry-deadline fault coverage.
• Added transport, compatibility-probe, controller, and chat-window ownership regressions for stale and duplicate callbacks, follow-up identity, cancellation, and synchronous configuration failures.
• Added thinking/finalizing cancellation, terminal reentry, duplicate history/usage/message, and stale second-step probe failure/cancellation regressions.
• Added active-composer, real Stop-button, draft-preservation, stale-control-state, request diagnostic lifecycle, and adversarial diagnostic privacy regressions.
• Added Plan-to-Work schema, approval, rejection, cancellation, stale-decision, one-request, mixed-batch permission, protocol-ordering, session-redaction, and narrow-card regressions.
• Added real-shaped volatile-observation, semantic-progress, early round-budget, exact-process cleanup, and bounded stop-evidence regressions.
• Added successful read-only batching, singleton-only category, unknown-tool, mixed escalation, atomic rejection, and correction-request regressions.
• Added checklist refresh/finalization, context-pressure fallback, complete inspect/fix/run/fail/stop/reread/fix/rerun/pass orchestration, exact cleanup ownership/privacy, stop-evidence truncation, and expanded-dialog ownership regressions.