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.

Orca v1.2.0

1.2.0 - 2026-10-05

Added

• Added first-class Ollama, LM Studio, and Local OpenAI-compatible profiles with editable conventional endpoints, optional bearer authentication, explicit keyless model discovery, and manual Model ID fallback.
• Added native Ollama /api/tags and LM Studio /api/v1/models discovery with bounded model metadata and filtering for models identified as embeddings.
• Added strict endpoint normalization, loopback/LAN/remote disclosure, exact-origin confirmation, and generated-endpoint validation for editable profiles and provider origin overrides.
• Added an isolated project-free two-step function-call compatibility probe and separate Agent-tool opt-in bound to the exact provider, origin, endpoints, model, reasoning configuration, and probe version.
• Added permanent Godot 4.7.2 editor integration coverage for dirty scripts and scenes, exact unsaved source, diagnostics logging, editor scaling, and plugin enable/disable/re-enable ownership.

Changed

• Local and editable profiles now start in Chat mode and expose Agent tools only after a matching compatibility pass and explicit user enablement.
• Provider discovery now uses bounded endpoint- and credential-aware caches; the normalized cache format was advanced to v3 for native local-provider metadata and stronger sanitation.
• Unsaved focused GDScript reads now use the exact current editor text, while reviewed file paths are canonicalized before editor-state, hash, revert, and write checks.
• Diagnostics services now share one reference-counted Godot logger, preserve warning severity, deep-copy caller records, and bound retained fields.

Fixed

• Fixed normalized path aliases and dirty editor buffers bypassing reviewed edit or revert protections.
• Fixed duplicate diagnostics and resource loading from logger callbacks during plugin reload and fallback-service ownership.
• Fixed native Ollama and LM Studio discovery exposing known embedding-only models while preserving unrestricted manual Model ID override.
• Fixed stale discovery-request ownership and programmatic model selection clearing or replacing the active local model list.
• Added bounded plain provider-error extraction and detection when a local endpoint silently serves a different model; Orca now returns model_mismatch before emitting assistant text.
• Fixed editable-provider Settings controls widening narrow docks and clipping the Done, About, and model actions.

Security And Privacy

• Editable non-loopback endpoints require confirmation bound to normalized scheme, host, and effective port; endpoint drift invalidates prior confirmation.
• Chat and discovery reauthorize endpoint trust before credential headers are constructed and reject provider redirects.
• Compatibility probes contain no project instructions, skills, editor context, task list, session history, project path, or real Orca tool schema.
• Compatibility passes persist only bounded binding metadata, start disabled, and are recomputed from each request snapshot before schema exposure and runtime execution.
• Probe requests and post-tool continuations cannot silently retry with altered stream options, preventing completed tool activity from being replayed.

Known Limits

• Orca does not bundle inference, install or download models, or manage Ollama or LM Studio processes.
• A successful compatibility probe verifies one narrow synthetic exchange; it does not guarantee planning quality or reliable Agent behavior on real projects.
• The tested qwen2.5-coder:3b Ollama model supported Chat but failed the first probe step, and the tested llama-3.2-3b-instruct LM Studio model supported Chat but repeated the tool call in the second probe step.
• Generic compatible servers may expose incomplete metadata. Conservative embedding-name filtering and manual Model ID override remain available.

Orca v1.1.1

1.1.1 - 2026-10-02

Fixed

• Recovered provider failures after complete tool rounds through bounded, sanitized, non-replayable checkpoints instead of making the conversation permanently view-only.
• Activated repetitive/no-progress tool-loop detection in the controller and changed the 12-round boundary from a system error into one tool-free finalization request.
• Raised the raw SSE transport allowance to accommodate provider framing while retaining strict accumulated-content limits, and bounded DeepSeek thinking output to 8,192 tokens.
• Connected bounded root project instructions and project-skill catalog metadata to the request-scoped controller context as originally documented for 1.1.0.

Security And Privacy

• Recovery never automatically retries tools, and its model-continuation checkpoint never persists hidden reasoning, raw tool output, call IDs, arguments, source content, hashes, process identifiers, or approval payloads. The separate visible activity log retains only its existing bounded display metadata.
• Incomplete, malformed, cancelled, pending-approval, dirty, and truncated turns remain non-resumable when Orca cannot prove complete tool protocol.

Orca v1.1.0

1.1.0 - 2026-10-01

Added

• Automatic bounded loading of root res://AGENTS.md as subordinate, private request-scoped project guidance.
• Bounded metadata discovery for res://skills/<slug>/SKILL.md and explicit read_project_skill loading for one exact skill body.
• Read-only inspect_godot_api reflection for ClassDB signatures, hierarchy, properties, signals, constants, enums, and safe editor Open Docs topics without prose scraping or object construction.
• Read-only read_gdscript_function extraction from saved scripts or exact unsaved editor source, with provenance and no patch hash for editor source.
• Read-only discover_dependencies traversal for bounded forward serialized dependencies and reverse dependents.
• Tool-loop thrashing detection for repeated calls/results, alternating cycles, repeated rounds, and no-progress rounds, followed by one no-tools finalization request beneath existing hard caps.
• Permanent automated suites for project instructions, project skills, Godot API reflection, focused function reads, dependency discovery, and tool-loop guarding.

Changed

• Improved post-1.0 responsive editor scaling, compact composer sizing, narrow action containment, transparent branding, expanded diff sizing, and active-turn auto-follow for delayed content and sequential approvals.
• Updated release metadata and About-page coverage to 1.1.0.

Security And Privacy

• Project guidance remains subordinate to system, user, mode, approval, path, and runtime safety checks and is removed from stored continuation history after each turn.
• Skill bodies are not automatically loaded, symlink traversal is rejected, and instruction/skill text and metadata use strict byte, line, and count limits.
• Unsaved editor source is never represented as a disk patch base, API reflection does not construct reflected objects, and dependency discovery does not load or instantiate resources.

Orca v1.0.0

Orca 1.0.0

Initial public release of Orca, a safety-first AI development assistant for Godot 4.7.2.

Highlights

• Integrated AI assistant dock for the Godot editor
• Plan and Work modes with explicit change
• Support for OpenAI, Gemini, xAI, DeepSeek, OpenRouter, and custom OpenAI-
• Project exploration, scene inspection, diagnostics, and reviewed file
• Local project-scoped conversation history

Installation

1. Download orca-v1.0.0.zip below.
2. Extract the archive into the root of your Godot project.
3. Open the project in Godot 4.7.2
4. Enable Orca under Project > Project Settings > Plugins.