CHANGELOG.md

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[0.2.4] - 2026-04-17

Added

• Markdown rendering for Mailbox messages via Earmark: headings, bold/italic, code blocks, and lists render formatted instead of as raw text. Includes XSS sanitization that strips script/iframe/object/embed tags, event handlers, and javascript: URLs.

Fixed

• Credential-alias token limits: profiles containing multiple credentials that share one underlying model (e.g., the same Gemini model registered under two credential names) now use the real model's context/output limits instead of silently defaulting to 128K. Previously this caused Vertex/Gemini to reject requests whose maxOutputTokens exceeded the true model cap.
• Child-dismiss cost decrease bug: cost absorption on child dismiss now runs inside a single Repo.transaction (delete + insert), eliminating the read-committed gap where dashboard totals briefly decreased. Defense-in-depth monotonic guard added to the dashboard as a fallback. Fifth and final fix for this recurring regression.
• Async shell Phase 1 results no longer leave pending_actions without acked: true, which previously caused incoming messages to silently queue during shell execution.

Removed

• Dead interrupt_wait code paths from Router, ClientHelpers, WaitHandlers, and TaskManager (zero callers; would have crashed if invoked due to references to a non-existent state field).

[0.2.3] - 2026-03-27

Changed

• Dashboard per-node scalar extraction: AgentNode only reads the fields it needs from agent state, reducing LiveView diff overhead.
• Log detail truncation with lazy-load: long log entries are truncated in the list view and expand on click.
• Removed duplicate agent cost PubSub subscriptions that caused redundant re-renders.
• Upgraded LLMDB to v2026.3 with new snapshot format.

Fixed

• Livebench coding category scoring issues.
• Benchmark grove adherence improvements for more reliable agent behavior.

[0.2.2] - 2026-03-18

Added

• Quoracle logo (SVG) displayed in navigation header, replacing the text link.
• Strict confinement mode for groves with grove_vars template variable resolution at spawn time.

Changed

• AgentNode evolved to recursive LiveComponent for per-node DOM diffing instead of full-tree re-renders.
• Log panel uses debounce timer and pre-computed display_logs to reduce render overhead.
• Dashboard cost queries batched with precomputed data flow for lower DB pressure.
• Mailbox panel fully isolated with dedicated PubSub subscriptions.
• CostDisplay breakdown cache invalidated automatically when total_cost changes.
• Initial cost data hydrated at mount time instead of lazy-loading.
• Header streamlined: gear icon replaces Settings text link, reduced whitespace.

Removed

• Dead signals, stale cost_id_prefix references, and unused assigns cleaned up from dashboard.

[0.2.1] - 2026-03-12

Fixed

• Child tracking race condition: use Registry fallback when injected children lag state updates, and deduplicate spawn_child results to keep tracking consistent under async races.
• Dashboard cost refresh storm: debounce cost_recorded PubSub events (2s window) to prevent O(N²) re-query feedback loop that caused UI lockup with 100+ agents.

Changed

• Removed 20-child injection limit that caused agents to lose visibility of children beyond 20, triggering unbounded respawning.
• Deferred per-agent GenServer calls (todos, budget) in Dashboard to after mount to prevent longpoll timeout with many live agents.

Removed

• Dead registry_spawned_at/1 function and unreachable code paths in ChildrenInjector.

[0.2.0] - 2026-03-10

Added

• Groves: declarative multi-agent orchestration via GROVE.md manifests. Define topology, bootstrap, governance, confinement, schemas, and spawn contracts in a single file — then run it.
  • Bootstrap: pre-fill task creation forms from file references or inline values.
  • Governance: dual-layer rule enforcement — prompt guidance (LLM reads the rule) AND runtime blocking (code enforces the rule). Supports shell_pattern_block and action_block hard rules.
  • Confinement: per-skill filesystem restrictions enforced at the file_read, file_write, and execute_shell action layer.
  • Schema validation: JSON Schema (Draft 2020-12) validation on file_write actions matching glob path patterns.
  • Spawn contracts: declared parent→child topology edges with auto-injection of skills, profile (with fallback), and constraints.
  • PathSecurity: three-layer defense against path traversal and symlink attacks across all grove resolvers.
• Two example groves shipped: livebench (6-category benchmark) and mmlu-pro (14-subject multiple-choice benchmark).
• Children context enrichment: agents see their children's latest message preview and status directly in the prompt.
• Correction feedback injection: parent feedback is injected into child prompts with lifecycle tracking and root stall notification when children stop progressing.

Fixed

• Per-model consensus queries now run in parallel instead of sequentially.
• Context window overflow from insufficient token safety margins across multiple providers.
• Correction feedback cleared by queued messages arriving during retry.
• Missing :id field in system stall messages crashing the Mailbox UI.
• Empty children signal not injected when no live children exist.
• GPT wait-stall pattern causing agents to idle indefinitely.
• LLM receive timeout too low for slow providers (increased to 300s).

Changed

• Comprehensive Groves documentation added to README.
• DRY refactors: children tracking, skill metadata construction, single-model persist path, PerModelQuery extraction.

[0.1.17] - 2026-02-28

Added

• Profile hot-reload: agents reload profile config at runtime (model pool, max refinement rounds, system prompt) without task restart.
• Forced reflection: single-model agents with force_reflection enabled undergo mandatory refinement in round 1.

Fixed

• Condensation overflow crashing when to_discard exceeded model context window.
• GenServer crash when LLM returns unknown sub-action type in batch_sync or batch_async.
• Reflector response extraction failing for reasoning models with :object content parts.

Changed

• Warning log broadcast on model pool switch failure during hot-reload.
• Non-negative validation for max_refinement_rounds in hot-reload handler.

[0.1.16] - 2026-02-23

Added

• System prompt caching in the consensus pipeline for improved performance.
• MCP Client lifecycle monitoring with liveness guards and crash-to-error logging.

Fixed

• Cost display budget timeouts and model attribution loss after child dismissal.
• Infinite continuation loop in fast-path consensus.
• Decimal.decimal(nil) crash in Tracker.over_budget? for N/A budgets.
• Response.text() blindness to pure JSON responses.
• Google Vertex string error body handling in RetryHelper v3.3.
• Silent result loss from HTTP actions missing timeout overrides.

Changed

• DRY Aggregator SQL with extracted ResponseLogger module.
• Test suite optimization: removed ~230 redundant tests, split heavy modules, migrated 23 files from DataCase to ExUnit.Case.

[0.1.15] - 2026-02-21

Fixed

• Consensus interleaving — Defer consensus when self-contained actions are still pending, preventing premature consensus queries.
• quoracle.reload change detection — Fix false positives from non-deterministic compilation.
• RetryHelper v3.2 — Retry HTTP 429/5xx errors from crashed response parsers with improved malformed response logging and stacktrace context.
• Vertex native model compatibility — Skip response_format for native Vertex models (Gemini, Claude) that don't support it.

Changed

• Apply ResponseTruncator to async shell completion path.

Removed

• Remove obsolete auto_complete_todo parameter.

[0.1.14] - 2026-02-21

Fixed

• Async shell Phase 2 results discarded — Shell completion results from long-running async commands were silently lost. ActionResultHandler now preserves pending_actions entries for Phase 1 acks, and ShellCompletion passes action_atom via 4-arity cast for standard continuation.
• Flaky shell Port crash test — Condensed core.ex stop_requested handler and stabilized the shell Port crash test.

Added

• Vertex MaaS JSON mode — Added response_format support for Vertex Model- as-a-Service, fixed ClientAPI timeouts.

[0.1.13] - 2026-02-20

Added

• Local model support — Connect to locally-hosted LLMs (Ollama, LM Studio, etc.) with conditional API key handling, LLMDB bypass for embeddings, custom model_spec input, and local model indicators in the UI.

Fixed

• MCP reliability — Guaranteed error delivery on dispatch task crash, added MCP_Client reconnection with dead status tracking, retry logic with reconnect and timeout increase, sync timeout for system-level actions, and fixed Streamable HTTP base_url path doubling.
• Budget timeout on child dismiss — Replaced blocking GenServer.call budget updates with cast-based approach, eliminating child timeout when parent is busy. Fixed Decimal-to-String drift in cost tracking.
• Cost details total — Fixed cost details total dropping on child dismiss.

Changed

• MCP protocol version — Set protocol_version per transport in connection manager.

[0.1.12] - 2026-02-16

Fixed

• Hot-reload false positives — Fixed quoracle.reload reporting spurious changes due to non-deterministic compilation order and MD5 algorithm mismatch, then always reporting no changes after the algorithm fix.
• Task cost tracking — Fixed task costs dropping to zero when dismissing children with N/A budgets.
• Agent stall during sync actions — Fixed agents stalling when user messages arrive while an always-sync action is executing.
• Consensus clustering — Fixed consensus clustering incorrectly handling mergeable params.

Changed

• Shell error messages — Enriched command_not_found errors with a list of available command IDs for easier debugging.
• Reflector token budget — Added 5% safety margin to Reflector max_tokens calculation to prevent truncation.

[0.1.11] - 2026-02-16

Fixed

• Dead Router PID crash — Guard check_id routing against dead Router PIDs, preventing crashes when routing to a terminated Router process.
• Reflector echoing action JSON — When conversation history was dominated by action JSON, the Reflector LLM would pattern-match to action output instead of producing lessons and state extraction. Split into system/user message structure with XML boundary tags to keep the LLM on task.

Changed

• ConfigBuilder simplification — Extracted inherit_key/4 helper to reduce ConfigBuilder complexity.

[0.1.10] - 2026-02-14

Fixed

• Parent ID lost after pause/resume — Child agents now correctly persist their parent ID when the parent process has exited, preventing nil parent_id after restoration.
• Root agent hijack on restore — Dashboard now uses first-writer-wins guard on root_agent_id, preventing orphaned agents from overwriting the root during task restoration.
• Restore halts on single agent failure — TaskRestorer now continues past individual agent failures instead of aborting the entire restoration.
• In-flight spawns missed during pause — Post-pause sweep catches agents that register after the initial Registry query, ensuring all agents are stopped before persistence.
• Orphaned agents from previous session — New cleanup pass terminates agents surviving from a prior session that were not part of the restoration set.

[0.1.9] - 2026-02-14

Fixed

• Action executor regressions — Fixed three regressions: error-state stall preventing recovery, shell action check_id misrouting, and Router process leak on action completion.
• Malformed LLM response handling — Reflector now retries when LLM responses fail to parse, instead of stalling.
• Flaky deadlock test — Fixed non-deterministic batch_sync test.

Changed

• LLM prompt robustness — Strengthened JSON output instructions for weaker models, clarified batch_sync/batch_async exclusions, added enum constraints to action JSON schemas, and improved parent-child timing guidance.
• Dynamic max_tokens safety margin — Token estimation now includes a 5% safety margin to prevent truncation.
• Action lifecycle log broadcasts — Restored action lifecycle events to the UI log panel.
• Reflector refactor — Extracted retry_ctx map to reduce function arity; deduplicated extract_check_or_terminate_id and removed dead code.
• Test performance — Deleted redundant tests, reduced timeouts, and simplified setup for faster test suite execution.

[0.1.8] - 2026-02-13

Added

• Hot-reload mix task — mix quoracle.reload enables hot-reloading modules on a running node without restart.

Fixed

• Hostname resolution in reload task — Fixed node name lookup for the hot-reload mix task.
• Embedding crash on non-binary input — Embeddings.get_embedding/1 no longer crashes when receiving non-binary input.
• Embedding cost context not propagated — Embedding cost metadata is now properly threaded through HistoryTransfer and batch_sequence_merge.

[0.1.7] - 2026-02-13

Added

• Multi-provider embedding support — Agents can now use different embedding providers and models via configurable opts, with token-based chunking for inputs exceeding provider token limits.

Fixed

• Action executor deadlock — Actions are now dispatched via Task.Supervisor instead of inline callbacks, eliminating a deadlock where synchronous action execution blocked the agent's message loop.
• Race condition in message handling — Extracted ActionResultHandler from MessageHandler to fix an R64 race condition where action results could be processed before the agent was ready.
• FunctionClauseError in Aggregator — extract_action_type/1 now handles unexpected input shapes gracefully instead of crashing.
• Batch validator crash on non-map entries — Guard added to prevent crashes when batch actions contain malformed entries.

Changed

• Replaced manual on_exit cleanup across test files with register_agent_cleanup helper for consistency.
• Added dynamic max_tokens to Reflector query path to prevent context window overflow.

[0.1.6] - 2026-02-12

Fixed

• batch_sync stall with wait:true — batch_sync actions with wait:true stalled indefinitely when all sub-actions were self-contained (e.g., file_read, orient). Now dynamically inspects sub-actions and auto-corrects wait:true to false when no sub-action can trigger an external response. Mixed batches preserve wait:true as expected.

[0.1.5] - 2026-02-12

Fixed

• Orphaned agents on task deletion — Deleting a task now force-kills its agents instead of sending a graceful stop request. Agents stuck in retry loops could not process :stop_requested messages, leaving them running indefinitely.
• Send message target cleanup — Stripped bracket artifacts from LLM-generated targets (e.g., [parent] → parent). Removed dead all_children and user target types; added error logging for invalid targets.
• Announcement guardrails — Tightened send_message schema and prompt guidance to prevent LLMs from misusing announcement for status updates. Announcement is now explicitly documented as broadcast-only for directives and corrections.

[0.1.4] - 2026-02-11

Changed

• Release workflow now triggers on tag push and creates the GitHub release automatically with changelog notes and tarball.

Note: v0.1.3 was skipped due to a GitHub ghost-ref bug that blocked tag re-creation.

[0.1.3] - 2026-02-11

Fixed

• Budget enforcement for spawn — Budgeted parent agents now correctly require a :budget parameter when spawning children. Budget data and spent amounts propagate through ActionExecutor to child agents, and budgets are reconciled when children are dismissed.

Changed

• Replaced dev-centric role names with research/business examples in SVG diagrams.
• DRY refactor of wait coercion, pipeline opts, and background dismissal extraction in Core.

[0.1.2] - 2026-02-11

Added

• DeepSeek reasoning support — OptionsBuilder now configures thinking/reasoning parameters for DeepSeek models alongside existing Claude support.
• llm_db update script for refreshing the local LLM database from upstream sources.

Fixed

• Dynamic max_tokens calculation — Prevents context window overflow by computing available output tokens from the model's context limit minus input token count, capped at the model's output limit. Includes proactive history condensation when output space falls below a safety floor.

Removed

• Outdated PostgreSQL setup script (superseded by standard Ecto workflows).

[0.1.1] - 2026-02-10

Added

• Per-profile max refinement rounds — Profiles can now set the maximum number of consensus refinement rounds (0–9). Temperature descent adapts automatically to the configured round count. Defaults to 4 (matching previous hardcoded behavior).
• Configurable skills path — The skills directory is now configurable via Settings > System. Defaults to ~/.quoracle/skills/.

Fixed

• Docker startup crash when CLOAK_ENCRYPTION_KEY is missing — now shows an actionable error message (#2).
• Improved Reflector prompt quality for more useful lesson extraction.

[0.1.0] - 2026-02-07

Initial open-source release. Recursive Mixture-of-Agents orchestration with multi-LLM consensus, 21 action types, three-panel LiveView dashboard, task persistence with pause/resume, hierarchical budget system, cost tracking, profile-based agent permissions, and file-based skills system.