CHANGELOG.md

Changelog

[2.30.4] - 2026-04-26

Changes

• optimus-fleet disabled — Temporarily disabled the shipped optimus-fleet skill so it no longer forces strict optimus_orchestrate passthrough while wait_for_completion=true can exceed MCP request timeouts and fail with MCP error -32001.
• Fleet recovery guidance — The disabled skill now directs agents to use manual todo decomposition and direct sub-agent dispatch until the underlying MCP-safe async orchestration behavior is fixed and validated.

[2.30.3] - 2026-04-26

Bug Fixes

• Async delegation defaults — delegate_task_async and dispatch_plan_async now return promptly by default instead of holding the MCP request open until terminal state; callers must opt in with wait_for_completion: true.
• Runtime proxy E2E stability — Runtime-backed worker execution now starts runs via /api/v2/agent/start and polls status, avoiding long synchronous HTTP requests that could black-hole fleet workers.
• Runtime restart races — Hardened runtime readiness checks and retry handling so connection-loss races after daemon restart do not immediately fail worker execution.
• Runtime async admission — /api/v2/agent/start now schedules prompt execution on the next tick so large prompts cannot block the start response.
• ACP busy-state tracking — ACP adapters now report busy immediately during invocation setup, closing a concurrency-test race.
• ADO work item listing — Implemented Azure DevOps work item listing for unified VCS operations and fixed result-limit handling.

Tests

• Added regression coverage for non-blocking async tool defaults, ACP busy-state setup, ADO work item listing, runtime proxy polling, and async runtime start behavior.
• Verified direct runtime HTTP E2E with github-copilot / gpt-5.5 and a real delegate_task_async worker smoke task to a verified terminal state.

[2.30.2] - 2026-04-25

Bug Fixes

• Fleet runtime bootstrap stability — Hardened automatic runtime startup so fleet workers prefer workspace-local bundles, pass the active workspace to the HTTP runtime, and emit actionable startup diagnostics when the runtime cannot become ready.
• Infrastructure failure handling — Classified runtime/proxy failures as infrastructure failures and skipped normal self-heal loops that depend on the same broken runtime path.
• Runtime capacity controls — Added v2 runtime admission control for synchronous and asynchronous generic runs, including overflow routing, 429 capacity errors, and async slot release on terminal status.
• Path containment hardening — Tightened async plan output-path validation to avoid Windows prefix containment bypasses such as sibling .optimus2 directories.

Improvements

• Fleet diagnostics consistency — Aligned optimus_status and /api/v2/health engine reporting so user-level configured engines are visible alongside built-in engines.
• ACP/runtime contracts — Consolidated duplicated ACP session typing and clarified adapter reuse behavior for multi-session ACP process pooling.

Tests

• Added regression coverage for runtime bootstrap candidate selection, startup diagnostics, infrastructure failure self-heal guards, v2 capacity helper behavior, engine diagnostics, and async plan path containment.
• Verified a real Optimus fleet smoke sample through optimus_orchestrate to a verified terminal state.

[2.30.1] - 2026-04-18

Improvements

• Agent-friendly runtime guidance — Added explicit shipped guidance across system instructions, master onboarding, and default scaffold prompts so upgraded workspaces immediately understand the current runtime model: a per-user multi-workspace HTTP daemon routed by workspace_path and X-Optimus-Workspace.

[2.30.0] - 2026-04-18

Features

• HTTP per-user multi-workspace daemon completion — Hardened the user-level runtime into the current release target: a shared HTTP daemon with explicit workspace-scoped routing, stricter runtime readiness checks, and overflow workers that preserve the same per-request workspace model instead of binding themselves to a single repo.

Improvements

• Runtime admission and queryability — /api/v1/agent/start now behaves more like a strict admission boundary: runs are persisted before async execution proceeds, immediate status recovery is more reliable, and runtime metadata remains visible even when the task manifest has not fully caught up yet.
• Workspace root normalization — Nested project paths and linked worktrees now normalize to the correct workspace root before runtime bootstrap, status lookups, and generic runtime operations, reducing repo-bound leakage in the daemon layer.

Tests

• Added targeted runtime regression coverage for per-request workspace enforcement, immediate run lookup after admission, bootstrap candidate resolution, and worktree-root normalization.

[2.29.0] - 2026-04-17

Features

• End-to-End Async Task Orchestration: Standardized wait_for_completion with timeouts across all async MCP tools (optimus_orchestrate, delegate_task_async, dispatch_plan_async). Added mandatory post-execution artifact verification rules to prevent agents from hallucinating task results and abandoning background tasks.

[2.28.0] - 2026-04-16

Added

• Agent-native intent classification: Added intent_signals to mcp_spartan-swarm_optimus_orchestrate schema to allow LLMs to actively classify intent instead of relying on regex detection, improving internationalization (e.g. Chinese requests).

[2.27.1] - 2026-04-16

Improvements

• Delegate-first fleet guidance — Updated the shipped README, architecture docs, scaffold instructions, onboarding guidance, and MCP tool descriptions so broad or multi-step requests now clearly start with optimus_orchestrate, while delegate_task_async is reserved for already-scoped single-worker execution and dispatch_plan_async for explicit dependency graphs.

Bug Fixes

• Blocked task dependency healing — TaskManifestManager.reapStaleTasks() now synchronously reconciles blocked tasks so downstream work no longer hangs forever when an upstream dependency failed, disappeared, or was already verified. Added machine-readable failure classifications for dependency and watchdog timeout outcomes, plus regression tests covering dependency failure propagation and automatic unblock recovery.

[2.25.0] - 2026-04-16

Features

• optimus-fleet wrapper skill — Added optimus-fleet skill providing a deterministic MCP dispatch layer. Enables structured orchestration patterns with explicit fleet routing for multi-agent coordination. (feat(skill): add optimus-fleet wrapper skill for deterministic MCP dispatch)

Improvements

• optimus_orchestrate rename — Renamed fleet orchestrator tool to optimus_orchestrate for consistency with the MCP tool naming convention. (refactor(mcp): rename fleet orchestrator to optimus_orchestrate)

[2.24.1] - 2026-04-16

Bug Fixes

• Windows async spawn proxy timeout (P1) — Fixed Runtime server proxy timed out after 330000ms on Windows. The proxy timeout was based on activityTimeoutMs (300s inactivity window) instead of heartbeatTimeoutMs (600s total task budget). Multi-step agent tasks (reading files, writing code, building TypeScript) easily exceed 5.5 minutes total elapsed time even if individual ACP messages arrive within 300s. Fix increases the proxy budget to 630s (heartbeatTimeoutMs + 30s grace), resolving Issue #568. (#568)
• Runtime server process detection — Extracted isRuntimeServerProcess(), shouldRouteViaRuntimeServer(), and resolveRuntimeProxyTimeoutMs() as standalone testable functions; added OPTIMUS_RUNTIME_SERVER=1 env var to the runtime server child process for reliable detection without process name inspection.
• Proxy double-settle guard — Added settled/finishReject/finishResolve guards in executeViaRuntimeServer to prevent both req.timeout and proxyTimer from firing on the same request.

[2.24.0] - 2026-04-14

Features

• Models now optional — available_models in available-agents.json is no longer required. Users only need to configure engines; the engine CLI selects its default model automatically. Explicit available_models still works as a whitelist for backward compatibility.
• Engine-default health tracking — Health tracking and cross-engine fallback now work in engine-default mode (no model specified), using a "default" sentinel for consistent health keys.
• T2 auto-repair on stale models — When a T2 role template has an obsolete model: that the CLI no longer supports, the system automatically clears it and retries with the engine default instead of quarantining the role.

Bug Fixes

• delegate_task_async startup timeout — Fixed TASK_STARTUP_TIMEOUT where async workers failed silently because the HTTP runtime server wasn't ready. The MCP server now awaits ensureRuntimeServer() before spawning async workers.
• Silent worker death — Async worker stderr is now logged to .optimus/logs/worker-{taskId}.log instead of being swallowed by stdio: "ignore". Spawn errors are immediately recorded as task failures.
• Spawn path validation — spawnAsyncWorker now validates mcp-server.js exists before spawning, providing an immediate actionable error instead of a 120-second timeout.
• Pre-flight validation for empty models — Fixed available_models: [] being treated as a valid whitelist (empty array is truthy). Now correctly treated as "no whitelist".

[2.23.4] - 2026-04-08

Bug Fixes

• Azure CLI path fallback — ADO auth now resolves az via explicit executable lookup and Windows Azure CLI install paths, so VCS tools still work when the MCP process starts with an incomplete PATH.
• Actionable Azure CLI diagnostics — When Azure CLI still cannot be found, the error now reports the executable resolution diagnostics and checked fallback locations instead of only saying 'az' is not recognized'.

[2.23.3] - 2026-04-08

Bug Fixes

• Azure DevOps timeout resilience — Added explicit timeouts for az account get-access-token and Azure DevOps HTTP requests so a degraded ADO or Azure CLI session no longer hangs the whole VCS tool path.
• VCS config diagnostics — Added vcs_config_diagnostics plus richer VCS failure output so users can see the resolved vcs.json, configured provider, git remote, and provider-cache status when troubleshooting.
• Provider cache correctness — VCS provider caching now keys on workspace, resolved config path, and git remote to avoid stale provider reuse after repo/config changes.

[2.23.2] - 2026-04-08

Bug Fixes

• Windows ACP process cleanup — AcpAdapter.cleanup() now uses taskkill /T /F for shell-spawned ACP processes on Windows so Copilot/Claude child processes do not survive after the wrapper shell exits.
• Runtime shutdown hygiene — HTTP runtime shutdown now drains AcpProcessPool on SIGTERM/SIGINT, and startup cleans up the previous runtime PID tree before binding the port.
• Pool exit cleanup — AcpProcessPool now shuts down all adapters on host process exit, preventing stale warm-pool workers from accumulating across sessions.

[2.23.1] - 2026-04-08

Features

• Cursor IDE support — Added .cursor/rules/optimus.mdc for Cursor IDE multi-agent orchestration, bringing Cursor in line with VS Code/Copilot IDE support.
• Copilot launcher scripts — Added copilot-optimus (bash), copilot-optimus.cmd (Windows CMD), and copilot-optimus.ps1 (PowerShell) convenience launcher scripts that start Copilot with the Optimus MCP config.

Tests

• delegate_task compat test — Added src/test/delegate-task-compat.test.ts to validate that delegate_task is properly advertised as a blocking compatibility wrapper and enforces workspace_path.

Bug Fixes

• package-lock.json version sync — Fixed optimus-plugin/package-lock.json version stale at 2.17.13; synced to match current 2.22.1/2.23.1 release.

Chores

• Added PATH note comments to .copilot/mcp-config.json and .mcp.json to guide ACP executable path debugging.
• Pruned stale MCP delegate proxy and user-level runtime sections from runtime-integration skill docs.
• Added nul to .gitignore (Windows filesystem artifact).

[2.23.0] - 2026-04-07

Features

• Runtime server preheat — MCP startup now preheats the runtime server and detects spawn crashes early, preventing cold-start latency and surfacing failures before the first delegate call.

Bug Fixes

• Cold start concurrency — Adapter pool reuses initializing adapters instead of spawning duplicates, eliminating race conditions on parallel cold-start requests.

[2.22.1] - 2026-04-07

Bug Fixes

• Cold start concurrency — Pool no longer replaces initializing adapters with new ones. Multiple concurrent tasks on a cold pool now share the same spawn+init via _readyPromise mutex instead of racing with duplicate processes.
• Init timeout — Increased ACP handshake timeout from 15s to 30s for cold start resilience.

Features

• Pool preheat API — AcpProcessPool.preheatEngines() enables eager warm-up of engine processes at startup.

[2.22.0] - 2026-04-06

Features

• User-level resource sharing — resolveOptimusPath now falls back to ~/.optimus/ for shared resources (config, roles, skills, dist, memory). Resolution: project → main worktree → ~/.optimus/. No more duplicating roles/skills/config across projects.
• Upgrade all projects — optimus upgrade defaults to upgrading all registered projects. Use --single for current-dir only.

Bug Fixes

• role-registry.json — Fixed hardcoded path to use resolveOptimusPath (was missing worktree + user-level fallback).
• mcp-servers.json — Now resolves via resolveOptimusPath with user-level fallback.

[2.21.1] - 2026-04-06

Features

• Upgrade all projects — optimus upgrade now defaults to upgrading all registered projects in ~/.optimus/projects.json. Use --single for current-directory only.

[2.21.0] - 2026-04-06

Features

• User-level HTTP Runtime — Runtime server is now cross-project. Start once with node ~/.optimus/dist/http-runtime.js --port 3100, serve all workspaces via per-request workspace_path. No --workspace required.
• MCP auto-start uses user-level path — ~/.optimus/dist/http-runtime.js is the primary auto-start candidate, enabling any project's MCP server to find the runtime.

Stability

• Crash guard — HTTP runtime catches uncaughtException/unhandledRejection without crashing.
• Session concurrency limit — AcpAdapter enforces maxConcurrentSessions (default: 10) to prevent OOM.
• MCP proxy auto-reconnect — Detects ECONNREFUSED/ECONNRESET, auto-restarts runtime server and retries.

Bug Fixes

• Auto-start TDZ — Fixed cwd used before declaration in ensureRuntimeServer.
• No silent fallback — Throws actionable error when runtime server unavailable instead of silently falling back to direct ACP spawn.
• v2 API workspace warning — Logs warning when workspace_path is missing (custom engines won't resolve).

Documentation

• Updated runtime-integration skill with User-Level Runtime section and cross-project examples.
• Updated system-instructions.md with runtime server dependency note.

[2.20.2] - 2026-04-06

Stability

• Crash guard — HTTP runtime server now catches uncaughtException and unhandledRejection without crashing, keeping the server alive for subsequent requests.
• Session concurrency limit — AcpAdapter enforces maxConcurrentSessions (default: 10) to prevent unbounded session growth on a single ACP process.
• MCP proxy auto-reconnect — When runtime server connection drops (ECONNREFUSED/ECONNRESET), the MCP proxy automatically restarts it and retries once.

Bug Fixes

• Auto-start TDZ — Fixed cwd variable used before declaration in ensureRuntimeServer, preventing runtime server auto-start from working.
• No silent fallback — When runtime server fails to start, throws actionable error instead of silently falling back to direct ACP spawn (which fails for copilot).

[2.20.1] - 2026-04-05

Bug Fixes

• Manifest cross-process safety — saveManifest now uses PID+timestamp-unique temp files, retry with exponential backoff on Windows EPERM/EACCES/EBUSY, and graceful degradation (never crashes the process). Fixes random EPERM: operation not permitted, rename task-manifest.json.tmp crashes when MCP server, runtime server, and cron patrol write concurrently.

[2.20.0] - 2026-04-04

Features

• Multi-session ACP concurrency — AcpAdapter now supports multiple concurrent sessions on a single ACP process. Replaces the _busy single-task lock with per-session context routing via _activeSessions Map. Notifications are routed by sessionId with backward-compatible fallback for single-session mode.
• Runtime server proxy for delegates — When running as MCP server (inside a host agent), all engine executions are transparently routed through the runtime HTTP server (:3100). This eliminates the nested Copilot auth conflict and decouples ACP process lifecycle from the MCP server.
• Initialization mutex — _readyPromise prevents double-spawn when multiple concurrent invocations hit a cold adapter simultaneously.
• Per-session activity timers — Each concurrent session has its own timeout watchdog instead of a shared global timer.

Bug Fixes

• Copilot delegate auth — Fixed Authentication required error when delegating to github-copilot engine from within a Copilot CLI host. Requests are now proxied through the runtime server which runs as an independent process.
• Pool ephemeral fallback removed — AcpProcessPool no longer spawns ephemeral processes for concurrent requests; instead reuses the persistent adapter's multi-session capability.
• v2 API workspace_path — Generic runtime API now passes workspace_path to engine resolver, enabling custom engines (e.g. qwen-code) configured in available-agents.json.

[2.19.1] - 2026-04-04

Bug Fixes

• Copilot ACP auth — Only strip classic PATs (ghp_ prefix) from child env; preserve OAuth (gho_) and fine-grained (github_pat_) tokens to prevent "Authentication required" errors (b6909ed)
• ACP shell mode — Use shell mode only for non-.exe executables on Windows to prevent double-quoting and spawn errors (d0bceea)
• ACP PATH injection — Inject Node.js bin directory into ACP child process PATH on Windows so node-based executables resolve correctly (0f67a45)

[2.19.0] - 2026-04-03

Features (Claude Code Architecture Research — Issue #555)

• Structured task notifications — delegateTaskSingle() returns TaskDelegationResult with status, tokens, timing, and metadata instead of plain text. check_task_status now shows execution metrics table.
• Coordinator synthesis gate — Research tasks with synthesis_required: true auto-extract key findings before unblocking dependent implementation tasks, preventing understanding loss in agent handoffs.
• Lifecycle hooks — 16 hook points across task lifecycle, execution pipeline, and MCP dispatch with veto support, best-effort error isolation, and 3 built-in hooks (metrics, logging, slow-execution).
• Prompt cache optimization — Parallel workers share byte-identical prompt prefixes via cache_control: ephemeral, saving ~70-80% of repeated token processing in council dispatch.
• Dynamic skill prompts — SKILL.md files support template syntax ({{role}}, {% if engine == "claude-code" %}, {% include %}) for context-aware prompt generation. Static skills unchanged (zero overhead).
• Memory staleness warnings — Entries older than 1 day get a warning: "claims about code behavior may be outdated."
• Memory snapshot initialization — .optimus/memory/snapshots/ enables team onboarding with pre-populated project conventions.

Security

• Path security hardening — New pathSecurity.ts module with null byte rejection, URL-encoded traversal detection, Unicode fullwidth character detection, and case-insensitive comparison for Windows/macOS.

Bug Fixes

• ACP executable resolution — Multi-strategy path resolver scans common npm install locations as fallback when system PATH doesn't include ACP executables. Auto-updates to absolute path on resolution.
• ACP config completeness — Fixed missing protocol and acp blocks in user-level available-agents.json that caused wrong executable/protocol fallback.

[2.18.0] - 2026-04-03

Features

• Lifecycle hooks system — Task execution pipeline now supports lifecycle hooks for pre/post task processing, enabling custom automation at each task phase.

Improvements

• Shared isPidAlive utility — Consolidated 3 duplicate PID-alive checks (TaskManifestManager, meta-cron-engine, worker-spawner) into a single shared utility.
• Optimized engine health fallback — pickHealthyFallbackCandidate now uses early return for confirmed-healthy candidates instead of building a full sorted array.

[2.17.13] - 2026-04-03

Improvements

• Claude Code auto-accept permissions — optimus go --cli claude now launches with --dangerously-skip-permissions by default for a frictionless autonomous workflow.

[2.17.12] - 2026-04-03

Bug Fixes

• Claude --resume Esc fallback — When using optimus go --cli claude, pressing Esc in the session picker now starts a new session instead of exiting, matching Copilot CLI behavior.

[2.17.11] - 2026-03-31

Improvements

• Clarified upgrade scope — optimus upgrade now states more explicitly that it upgrades the current workspace and its local launchers/config. The completion message and docs also explain that npx github:cloga/optimus-code upgrade picks up the latest published CLI package itself, including optimus go.

[2.17.9] - 2026-03-31

Features

• CLI update notices — optimus now checks the latest GitHub release on a cached interval and prints an upgrade reminder when a newer version is available. Cached notices continue to work offline, and optimus serve skips the check to keep MCP stdio output clean.

[2.17.8] - 2026-03-31

Bug Fixes

• Prune stale optimus go project entries — The global project registry now self-heals on load by removing paths that no longer exist or no longer contain .optimus, so deleted smoke and temp workspaces stop polluting the project picker.

Improvements

• Clearer upgrade reminder — optimus upgrade now ends with a direct reminder to run npx github:cloga/optimus-code upgrade in other installed workspaces so they pick up the latest fixes.

[2.17.7] - 2026-03-31

Bug Fixes

• Published optimus upgrade --disable-project-available-agents migration flag — Upgrades can now disable an active project-level .optimus/config/available-agents.json override by renaming it to a non-active backup (for example .optimus/config/available-agents.project.disabled.json) and reactivating the user-level ~/.optimus/config/available-agents.json as the authoritative engine registry.

[2.17.6] - 2026-03-31

Bug Fixes

• Init handshake timeout and config-driven adapters — Added init handshake timeout to prevent startup hangs, made adapters config-driven, and reduced timeouts for faster failure detection. (a1cf103, #538)
• Resolve acp_process_crashed across all engines — Fixed acp_process_crashed error propagation so all engine adapters report crash events consistently. (ab3a5f1, #538)

[2.17.5] - 2026-03-30

Features

• User-level available-agents is now the default config source — optimus init now seeds ~/.optimus/config/available-agents.json (or OPTIMUS_USER_AVAILABLE_AGENTS_PATH) as the primary engine registry. Projects receive .optimus/config/available-agents.project.sample.json as an opt-in override sample instead of an always-active override.

Bug Fixes

• Engine config hierarchy and fallback hardening — Runtime engine resolution now layers built-in defaults under user-level and project-level config, preserves nested object overrides without concatenating arrays, skips malformed user config with a warning, and retains raw fallback for malformed project config during static validation.
• optimus upgrade now preserves the new default model cleanly — upgrades sync the user-level engine registry, keep existing project-level overrides when present, and refresh the project-level sample when no active override exists.

[2.17.4] - 2026-03-30

Bug Fixes

• All MCP configs now use absolute paths — Copilot CLI 1.0.13 auto-discovers .vscode/mcp.json but cannot resolve ${workspaceFolder} variables, causing Connection closed. All generated MCP configs (.vscode/mcp.json, .copilot/mcp-config.json, .mcp.json) now use absolute paths for command, args, and env values. These files are gitignored so absolute paths don't affect other users.

[2.17.3] - 2026-03-30

Bug Fixes

• Reverted --disable-mcp-server in optimus go — The v2.17.2 fix caused spartan-swarm to show as disabled (⊘) because --disable-mcp-server takes precedence over --additional-mcp-config. Reverted to simple --additional-mcp-config injection, which per Copilot CLI docs "overrides any installed MCP server configuration with the same name."

[2.17.2] - 2026-03-30

Bug Fixes

• MCP duplicate registration via IDE auto-connect — optimus go now passes --disable-mcp-server for each server defined in .copilot/mcp-config.json before re-injecting via --additional-mcp-config. This prevents duplicate spartan-swarm registration when Copilot CLI auto-connects to a running VS Code instance that already loaded .vscode/mcp.json.

[2.17.1] - 2026-03-30

Bug Fixes

• MCP config node ENOENT on Windows — optimus init/upgrade now writes the absolute Node.js path (via process.execPath) into generated MCP configs instead of bare node. Fixes spawn node ENOENT when MCP hosts like Copilot CLI spawn the server without shell: true. Fixes #533.

[2.17.0] - 2026-03-30

Features

• optimus_status MCP tool — New tool to verify Optimus Swarm activation. Returns version, workspace, skills/roles/engines count, system instructions and memory status. Provides actionable next steps.
• Enhanced system instructions (v3) — Injected copilot-instructions.md, CLAUDE.md, and cursor.mdc now include a MANDATORY FIRST ACTION directing agents to call optimus_status immediately, ensuring reliable Optimus activation across all clients.

Bug Fixes

• Security: 8 npm vulnerabilities resolved — Fixed path-to-regexp ReDoS (high), hono prototype pollution, brace-expansion hang, and esbuild/vite/vitest chain (5 moderate). npm audit now reports 0 vulnerabilities.
• Council capacity test fixed — The pre-existing council-capacity.test.ts failure ("rejects malformed config entries") is now resolved. Added readRawEngineEntries() fallback for lenient config reading when strict parsing rejects malformed entries. 245/245 tests now pass.
• Meta-cron patrol concurrency (#511) — Fixed safety timer leak in meta-cron-engine.ts that allowed overlapping patrol runs. All exit paths now properly clear both checkInterval and safetyTimer.
• Agent output_path enforcement (#382) — Added post-execution artifact rescue in worker-spawner.ts that detects when agents write to self-chosen filenames instead of the specified output_path, and auto-corrects by moving the content.

Refactoring

• worker-spawner.ts modularization — Extracted 989 lines (37% reduction) into two new modules:
  • src/mcp/engine-resolver.ts (1023 lines) — engine config, model validation, protocol resolution, health tracking, ACP auto-discovery
  • src/mcp/t3-tracker.ts (80 lines) — T3 usage tracking, role name sanitization, usage log persistence
  • All functions re-exported from worker-spawner.ts for backward compatibility

Dependencies

• vitest upgraded from 2.1.9 to 3.2.4

[2.16.28] - 2026-03-28

Features

• Agent Runtime streaming (SSE) — New GET /api/v1/agent/runs/:id/stream endpoint provides real-time Server-Sent Events during agent execution. Emits text, thinking, status, error, and done event types with sequence numbers for reconnection support.
• Streaming pipeline — executePrompt() now accepts an onChunk callback that receives streaming chunks from ACP adapters. Events flow from engine → AcpAdapter → genericExecutor → worker-spawner → in-memory event buffer → HTTP SSE.
• SDK streaming — OptimusRuntime client gains streamRun() async generator and runAndStream() convenience method for consuming SSE events from TypeScript/JavaScript.
• Event buffer — Per-run in-memory ring buffer (2000 events) with subscriber notification, sequence-based replay for reconnection, and automatic TTL cleanup (5 min after completion).

[2.16.27] - 2026-03-28

Bug Fixes

• Agent Runtime state file lifecycle sync — worker-spawner now appends running, completed, and failed history entries to .optimus/state/agent-runtime/run_*.json during task execution. Previously, state files remained stuck at pending even after successful completion. Fixes #532.

[2.16.26] - 2026-03-28

Bug Fixes

• Runtime frontmatter ReferenceError — Fixed bare frontmatter identifier in worker-spawner.ts that caused ReferenceError: frontmatter is not defined in orchestrator code paths. The variable is now properly scoped as roleFrontmatter from the parsed role template. Fixes #531.
• TypeScript baseline errors resolved — Fixed outputSchema type mismatch in genericExecutor.ts. npm run check-types now passes with zero errors.

[2.16.25] - 2026-03-27

Bug Fixes

• optimus go default --resume — Restored --resume as a default argument for both Copilot and Claude adapters, which was accidentally dropped during the multi-CLI refactor in v2.16.24.

[2.16.24] - 2026-03-27

Features

• Multi-CLI optimus go — optimus go now supports both GitHub Copilot CLI and Claude Code CLI via a client adapter layer. Use --cli copilot or --cli claude to override per-launch.
• Per-project CLI preference — Each project can store a preferredCli in ~/.optimus/projects.json. Set via optimus go set-cli <project> <client>.
• Global default CLI — Set via optimus go set-default-cli <client>. Resolution order: --cli flag → project preferredCli → global default → copilot.
• Client adapter architecture — New go-clients.js module provides extensible adapter definitions for each supported CLI (executable, config path, injection flags).

[2.16.23] - 2026-03-26

Features

• Lightweight CLI installer — New scripts/install-cli.ps1 and scripts/install-cli-remote.ps1 install only the optimus go launcher (~25KB) to ~/.optimus/cli/, auto-adding to user PATH. No roles, agents, skills, or dist bundles needed.
  • One-line remote install: irm https://raw.githubusercontent.com/cloga/optimus-code/master/scripts/install-cli-remote.ps1 | iex
  • Local install: .\scripts\install-cli.ps1
  • Uninstall: .\scripts\install-cli.ps1 -Uninstall
• optimus go --resume by default — optimus go now automatically passes --resume to Copilot CLI, enabling session resumption without manual flags.

[2.16.22] - 2026-03-26

Maintenance

• Removed redundant dependencies — Cleaned up 4 unused packages from package.json:
  • lucide-react (frontend icon library, never imported in CLI codebase)
  • node-fetch (no code references; Node 18+ has native fetch)
  • @types/marked (for marked v5; project uses v17 which ships its own types)
  • @types/strip-ansi (for strip-ansi v3; project uses v7 which ships its own types)

[2.16.21] - 2026-03-26

Features

• optimus go cross-project Copilot launcher — Added a new optimus go command that launches GitHub Copilot CLI for any registered Optimus workspace without requiring a manual cd.
• Global project registry — Optimus now stores registered workspaces in ~/.optimus/projects.json, supports alias-aware project lookup, and can scan the home directory plus its direct child folders for .optimus/ workspaces.
• Automatic workspace registration — optimus init and optimus upgrade now auto-register the current project in the global registry so it is immediately available to optimus go.
• Regression coverage for project registry — Added tests covering project registration, scan behavior, and Copilot launcher argument construction.

[2.16.20] - 2026-03-25

Bug Fixes

• Azure DevOps az-cli auth support — AdoProvider now respects ado.auth from .optimus/config/vcs.json. When configured as \"az-cli\", Optimus falls back to az account get-access-token if no PAT environment variable is present.
• ADO authorization header handling — PAT-based auth continues to use Basic, while Azure CLI access tokens now use the correct Bearer authorization header.
• ADO auth regression coverage — Added tests to verify az-cli fallback works and that PAT environment variables still take precedence when present.

[2.16.19] - 2026-03-25

Improvements

• Copilot CLI launcher generation — optimus init and optimus upgrade now generate repo-local Copilot CLI launchers (copilot-optimus.ps1, copilot-optimus.cmd, and copilot-optimus) that automatically preload the project's .copilot/mcp-config.json.
• Copilot CLI documentation — Updated the setup guide to explain that Copilot CLI should be started via the generated launcher so project-level MCP configuration is applied consistently.

[2.16.18] - 2026-03-22

Refactoring

• Architecture: Runtime Decoupling — delegateTaskSingle() now calls genericExecutor.executePrompt() instead of directly invoking AcpAdapter. This establishes the correct layering: AcpAdapter → genericExecutor (infra) → Harness → Optimus Orchestration (business) → Transport (MCP/HTTP).
  • Output validation runs once inside executePrompt() (previously duplicated in both worker-spawner and genericExecutor)
  • Session ID, usage metrics, and stop reason are returned via ExecuteResult instead of reading adapter properties directly
  • Doom loop detection and metadata backfill now use execResult properties instead of adapter.lastSessionId/adapter.lastUsageLog/adapter.lastStopReason
  • No behavioral changes — all 219/220 tests pass (1 pre-existing failure)

[2.16.17] - 2026-03-22

Features

• Harness Engineering: Mechanical Linter — Deterministic (non-LLM) validation of Optimus artifacts. Lints role templates (frontmatter fields, naming, engine/model validity, thin template detection), skill files (structure, naming), artifact outputs (type/status/author/date format), and engine-model cross-consistency against available-agents.json. Includes lintWorkspace() for batch validation.
• Harness Engineering: Entropy Patrol — Periodic workspace health checks designed for meta-cron execution. 5 built-in checks: structural lint (roles + skills), stale T1 agent detection (>7 days), orphaned skill references, quarantined role alerts, and memory file health. Generates formatted markdown reports.

[2.16.16] - 2026-03-22

Features

• Harness Engineering: Output Validation Gate — Agent output is now validated before being written to artifact files. Built-in rules catch: empty output, schema non-compliance, premature completion declarations, unfinished code markers (TODO/FIXME), and error trace leaks. Validation runs on both v1 (delegate_task) and v2 (generic API) paths.
• Harness Engineering: Doom Loop Detection — Tracks per-session file edit counts. When an agent edits the same file 3+ times, a warning is logged suggesting the agent reconsider its approach. Inspired by LangChain's LoopDetectionMiddleware.
• Harness Engineering: Self-Verification Prompt — All delegated agents now receive a mandatory verification checklist appended to their prompt, requiring them to re-read the task spec, verify output completeness, and check for errors before submitting. Controllable via verification_level in role frontmatter (strict | normal | skip).

[2.16.15] - 2026-03-22

Features

• Generic Agent Runtime v2 API — New decoupled API layer (/api/v2/) that provides prompt-based agent execution without Optimus-specific concepts (no role, skill, workspace, or tier system required).
  • POST /api/v2/agent/run — Synchronous prompt → result execution
  • POST /api/v2/agent/start — Async run initiation
  • GET /api/v2/agent/runs/:id — Status polling with result
  • POST /api/v2/agent/runs/:id/cancel — Run cancellation
  • GET /api/v2/health — Health check with available engines list
• Built-in engine defaults — github-copilot and claude-code engines work out-of-the-box without available-agents.json configuration.
• Structured output via v2 — Pass output_schema (JSON Schema) to get validated structured responses.
• Generic executor (genericExecutor.ts) — Direct AcpProcessPool-based execution, no TaskManifestManager dependency.
• Generic runtime service (genericRuntime.ts) — In-memory run tracking with lightweight envelope format.

[2.16.14] - 2026-03-22

Bug Fixes

• Artifact file normalization — When structured output is extracted from prose/markdown wrapping, the clean JSON is now written back to the output_path artifact file. This ensures downstream consumers reading the file directly get machine-readable JSON, consistent with the API result field.

[2.16.13] - 2026-03-22

Features

• Runtime envelope metadata — runtime_metadata now includes usage (token counts: input_tokens, output_tokens) and stop_reason (why the engine stopped, e.g. end_turn, max_tokens). Captured from ACP promptResult for both Copilot and Claude Code engines.
• ACP structured output preference — ACP adapter now prefers promptResult.content (structured response) over streaming agent_message_chunk chunks. Falls back to streaming when engines don't populate content (current behavior for both Copilot and Claude Code).

[2.16.12] - 2026-03-22

Bug Fixes

• Structured output JSON extraction — When engines return JSON wrapped in markdown prose/code fences, Optimus now extracts the JSON payload before validation. Supports ```json ``` code fences and brace-matching extraction. Previously this caused invalid_structured_output errors even when the engine produced correct JSON.
• 100% agent-friendly error coverage — All remaining error paths now include actionable Fix: guidance:
  • Worker-spawner: engine resolution failure and model pre-flight errors
  • ACP adapter: generic fallback error now suggests recovery steps
• System defaults skip verbose configs — applyEngineDefaults() now detects old-format configs (with explicit protocol, cli, or capabilities fields) and skips default injection to avoid conflicts.

[2.16.11] - 2026-03-22

Bug Fixes

• CAPIError detection — Copilot backend API errors (CAPIError: 400/500) now caught by fail-fast error detection in worker-spawner and classified with actionable fix guidance in HTTP runtime.
• Automation policy fix hint — buildAutomationCompatibilityFixHint no longer incorrectly suggests downgrading to continuation: "single" for claude-code. Now recommends running upgrade to refresh system defaults.
• New HTTP error category — automation_policy_invalid (422) with guidance to run upgrade to get system defaults.

[2.16.10] - 2026-03-22

Improvements

• Simplified available-agents.json — Config now only contains available_models, acp.path, and timeout. ACP protocol, warm pool, and autopilot mode are injected as system defaults by applyEngineDefaults() in code — no need for users to configure them.
• Engine system defaults are defined per-engine in ENGINE_SYSTEM_DEFAULTS (worker-spawner.ts) and deep-merged at config load time.

[2.16.9] - 2026-03-22

Features

• HTTP runtime auto-scaling — When the primary instance reaches max concurrent runs, overflow instances are automatically spawned on adjacent ports. Overflow instances auto-shutdown after idle timeout (default: 60s). Total capacity = MAX_CONCURRENT * (1 + MAX_OVERFLOW) = 20 by default.
  • OPTIMUS_MAX_OVERFLOW env var controls max overflow instances (default: 3)
  • OPTIMUS_OVERFLOW_IDLE_TIMEOUT env var controls idle shutdown in seconds (default: 60)
  • Primary instance acts as reverse proxy, routing overflow requests transparently
  • /api/v1/health reports all instances with per-instance active_runs and idle_ms
  • X-Optimus-Instance response header identifies which instance handled the request

[2.16.8] - 2026-03-22

Breaking Changes

• ACP-only mode — All engines (github-copilot, claude-code, qwen-code) now use protocol: "acp" exclusively. CLI adapters (GitHubCopilotAdapter, ClaudeCodeAdapter) are deprecated and emit a warning if triggered. All agent interactions go through ACP warm pool with autopilot mode.

Features

• HTTP timeout protection — server.requestTimeout = 0 to prevent Node.js from killing long-running agent connections (default was 5 min, agent tasks can run 10-30 min).
• Concurrency control — MAX_CONCURRENT_RUNS = 5 (configurable via OPTIMUS_MAX_CONCURRENT env var). Returns 429 concurrency_limit with actionable fix when at capacity.
• Health endpoint enhanced — /api/v1/health now reports active_runs and max_concurrent for observability.

[2.16.7] - 2026-03-22

Improvements

• Actionable HTTP runtime errors — Every error response now includes a fix field with concrete recovery steps. Error JSON format: { error: { code, message, fix } }. Covers all 15+ error codes across input validation, run lifecycle, engine errors, and route matching.
• invalid_state now shows actual status — Resume errors include current status: <status> so callers know why the operation was rejected.
• 3 new error classifications — role_quarantined, skill_preflight_failed, engine_resolution_failed are now properly classified at the HTTP layer instead of falling through to generic internal_error.

[2.16.6] - 2026-03-22

Features

• ACP autopilot mode & model selection — ACP adapter now sends configOptions (mode + model) via session/new, attempts session/configure and session/setConfiguration with graceful fallback. Ensures delegated tasks run in autopilot mode when available.
• AgentAdapter options interface — invoke() accepts optional { model, autopilot, maxContinues } to control session behavior per-task.

[2.16.5] - 2026-03-22

Improvements

• Agent-friendly error messages — All ACP adapter errors now include recovery guidance (auth setup, retry hints, timeout config). HTTP server errors return structured error_code values (auth_failed, rate_limit, task_timeout, acp_process_crashed, invalid_model, invalid_engine, workspace_not_initialized) instead of generic internal_error.
• Runtime integration SKILL.md — Added comprehensive error code reference table, authentication setup guide, warm pool behavior docs, and troubleshooting section for common failures.

Error Classification (New)

• auth_failed (401) — engine authentication missing or expired
• rate_limit (429) — API rate limit exceeded
• task_timeout (504) — no activity from engine within timeout window
• acp_process_crashed (500) — engine process exited unexpectedly (auto-recovers)
• invalid_model (400) — model not available for the specified engine
• invalid_engine (400) — engine not found in config
• workspace_not_initialized (400) — .optimus/ directory not found
• body_too_large (413) — now includes 10 MB limit in message
• invalid_json (400) — now includes parse error details

[2.16.4] - 2026-03-22

Fixes

• ACP session/prompt uses array format first — Copilot ACP requires prompt: [{type:'text', text:...}] (not text string). Swapped compatibility order: try content-array first, fallback to text for legacy agents.
• isInvalidParamsError matches -32603 — Copilot returns -32603 (Internal error) with schema validation data instead of -32602 (Invalid params). Error matcher now covers both codes.

Verified

• HTTP runtime E2E: 4 consecutive tasks via POST /api/v1/agent/run
• Warm pool reuse confirmed across all tasks (invocations 1→2→3→4)
• Models tested: gpt-5.4, claude-opus-4.6-1m, gemini-3-pro-preview
• All via Copilot ACP with warm pool — single process, multiple sessions

[2.16.3] - 2026-03-22

Fixes

• Resolver fails for explicit protocol with sub-object capabilities (#499) — getTransportConfig() returned the parent engine config instead of the protocol sub-object when protocol !== 'auto', making capabilities.automation_continuations invisible to all check functions. Now prefers the cli or acp sub-object when it exists.
• Upgrade merge preserves stale protocol field — deepMergePreserveUser treats scalars as atomic, so user's old protocol: "cli" overrode template's "auto". Post-merge migration now normalizes protocol to "auto" when both acp and cli sub-objects are present.
• Health endpoint hardcoded version 2.14.0 — Now injected at build time via esbuild define from package.json. Health correctly reports the installed version.
• Stale dist bundles in git — Rebuilt and committed all 3 bundles so npx github:cloga/optimus-code#v2.16.3 gets current code.

Tests

• 4 new regression tests for #499: explicit protocol with sub-objects, explain output verification, flat config backward compatibility.

[2.16.2] - 2026-03-22

Fixes

• Copilot autopilot resolver failure after upgrade — deepMergePreserveUser treated arrays atomically, so users upgrading from older configs had stale automation_continuations: ["single"] that suppressed the template's ["single", "autopilot"]. Both ACP and CLI transports failed the continuation check, causing selectedProtocol = null and timeout-like behavior. (Fixes #499)
• Scaffold ACP config stale — Updated scaffold available-agents.json so new installs get correct Copilot ACP capabilities (autopilot in continuations, no --stdio flag).
• Upgrade now unions capability arrays — Post-merge migration ensures engine capability arrays are unioned with template values instead of being overridden by old user arrays. Also strips stale --stdio from ACP args.

[2.16.1] - 2026-03-22

Fixes

• init/upgrade now copies all 3 dist bundles — Previously only mcp-server.js was copied to .optimus/dist/, making http-runtime.js and runtime-cli.js unavailable in user workspaces. Now all 3 bundles are deployed on optimus init and optimus upgrade.

Additions

• runtime-integration skill — Agent-discoverable documentation for HTTP REST, TypeScript SDK, and CLI contract integration. Agents find it via list_knowledge() or .optimus/skills/ scan.

[2.16.0] - 2026-03-22

Features

• Copilot CLI ACP Warm Pool — GitHub Copilot now runs via ACP protocol (copilot --acp) with persistent process pooling. Every delegation reuses a warm Copilot process instead of spawning fresh copilot -p "prompt" subprocesses (~4s cold start eliminated).
• Agent Runtime In-Process Execution — HTTP server, CLI, and SDK now execute tasks in-process, sharing the warm AcpProcessPool across runs. Previously each API call spawned a detached subprocess with its own pool (no warm reuse).
• Unified Warm Pool — All transports (MCP delegate_task, HTTP /agent/run, CLI optimus-runtime run, SDK runtime.runAgent()) now share the same AcpProcessPool for both Claude and Copilot engines.

Changes

• Copilot ACP capabilities updated to declare autopilot support — ACP protocol inherently supports continuation (agent runs to completion within session/prompt).
• Protocol auto-resolution now correctly selects ACP (preferred) for Copilot instead of falling back to CLI.
• Removed 5 Agent Runtime MCP tools (run_agent, start_agent_run, get_agent_run_status, resume_agent_run, cancel_agent_run) — replaced by native HTTP/CLI/SDK transports from v2.15.0.
• Added runWorkerInProcess() in council-runner — process-safe variant that never calls process.exit().
• Removed redundant --stdio flag from Copilot ACP args.

[2.15.1] - 2026-03-22

Features

• Copilot CLI ACP Warm Pool — GitHub Copilot now runs via ACP protocol (copilot --acp) with persistent process pooling, matching the warm start optimization previously available only for Claude ACP.
  • Before: Every delegation spawned a fresh copilot -p "prompt" process (~4s cold start: Node.js bootstrap + MCP server init + tool registration)
  • After: A persistent copilot --acp process stays alive in the pool; subsequent tasks skip all initialization overhead
  • Verified: copilot --acp implements full ACP JSON-RPC protocol (initialize, session/new, session/prompt, session/update streaming, session/request_permission auto-approval)
  • Session context preserved across multi-turn interactions within the same warm process

Changes

• available-agents.json — Copilot ACP capabilities updated to declare autopilot support in automation_continuations. ACP protocol inherently supports continuation (agent runs to completion within session/prompt), so autopilot is a natural fit.
• Protocol resolution now correctly selects ACP (preferred) for Copilot instead of falling back to CLI. Previous behavior was caused by ACP not declaring autopilot capability while the engine automation requested it.
• Removed redundant --stdio flag from Copilot ACP args (Copilot --acp defaults to stdio transport).

[2.15.0] - 2026-03-22

Features

• Agent Runtime HTTP Server — Native REST API for host application integration, no MCP transport required:
  • POST /api/v1/agent/run — Synchronous run (blocks until complete)
  • POST /api/v1/agent/start — Async start (returns immediately)
  • GET /api/v1/agent/runs/:id — Get run status/result
  • POST /api/v1/agent/runs/:id/resume — Resume blocked run
  • POST /api/v1/agent/runs/:id/cancel — Cancel active run
  • GET /api/v1/health — Health check
  • Start: node .optimus/dist/http-runtime.js --port 3100 --workspace /path
• Agent Runtime CLI Contract — JSON-in/JSON-out CLI for app embedding:
  • optimus-runtime run < request.json — sync run
  • optimus-runtime start < request.json — async start
  • optimus-runtime status --run-id <id> — get status
  • optimus-runtime resume < resume.json — resume blocked
  • optimus-runtime cancel --run-id <id> — cancel
  • All output is structured JSON on stdout, logs on stderr
• TypeScript SDK (src/sdk/runtime-client.ts) — Thin HTTP client wrapping the REST API:
  • const runtime = new OptimusRuntime({ baseUrl: 'http://localhost:3100' })
  • await runtime.runAgent({ role: 'writer', input: {...} })
  • Typed methods: runAgent, startRun, getStatus, resumeRun, cancelRun, waitForCompletion, health

Architecture

• AgentRuntimeService (src/runtime/agentRuntimeService.ts) — Transport-agnostic service layer extracted from MCP server. Shared by MCP tool handlers, HTTP server, and CLI. Clean API: runSync(), startRun(), getRunStatus(), resumeRun(), cancelRun(), waitForCompletion().
• Three build outputs via esbuild: dist/mcp-server.js (MCP stdio), dist/http-runtime.js (HTTP REST), dist/runtime-cli.js (JSON CLI).
• RuntimeError class with code and httpStatus for clean error propagation across all transports.

Compatibility

• MCP tool handlers (run_agent, start_agent_run, etc.) now delegate to the shared service — behavior is identical.
• Addresses #517: Host applications no longer need MCP transport.

[2.14.0] - 2026-03-22

Features

• ACP persistent process pool — Eliminates cold-start overhead (~1-2s) for ACP engine tasks by keeping agent processes alive between invocations:
  • First task spawns the ACP process and performs the initialize handshake
  • Subsequent tasks reuse the warm process — only session/new + session/prompt are needed
  • Idle processes are automatically evicted after 5 minutes
  • If a persistent process crashes, the next invocation auto-recovers (respawn + reinitialize)
  • When the pool adapter is busy with a concurrent task, an ephemeral adapter is created as fallback

Architecture

• AcpProcessPool (src/utils/acpProcessPool.ts) — Singleton pool managing warm ACP adapter instances, keyed by engine. Provides idle sweep, graceful shutdown, and usage stats (reuses/creations).
• AcpAdapter dual lifecycle — Now supports both ephemeral (per-task spawn/kill, original behavior) and persistent (process stays alive) modes. New pool management API: isAlive(), isBusy(), shutdown(), idleSince, invocationCount.
• Per-task MCP env injection — In persistent mode, per-task environment variables (delegation depth, role, etc.) are injected into MCP server configs via session/new params, ensuring child processes get correct context even though the ACP process is reused.
• Process crash safety — Exit handler uses closure-captured process identity to prevent stale handlers from clobbering newly spawned processes during auto-recovery.

Compatibility

• Fully backward compatible — Ephemeral adapters (non-pool) behave identically to before. The pool is transparent to the delegation system.

[2.13.0] - 2026-03-22

Features

• Worktree orchestration tools — 3 new MCP tools for automated multi-branch parallel development:
  • create_worktree — Create a git worktree for a branch, with automatic .optimus/ state initialization
  • list_worktrees — List all active worktrees with branch, HEAD, and Optimus state status
  • remove_worktree — Clean up a worktree after work is complete
• branch parameter on delegate_task / delegate_task_async — When specified, automatically creates a worktree for the target branch and runs the agent there. Enables Master Agent to dispatch parallel feature work across isolated branches with zero git conflicts.

Architecture

• WorktreeManager (src/utils/worktreeManager.ts) — Lifecycle management for git worktrees: create, list, find-by-branch, remove, ensure-for-branch. Auto-generates worktree paths using the convention ../<repo>-wt-<branch>/.
• Seamless delegation — spawnAsyncWorker already passes cwd: workspacePath to child processes, so worktree-targeted tasks run in the correct directory with full state isolation (v2.12.0) and shared config/roles/skills from the main worktree.

Compatibility

• Fully backward compatible — Omitting the branch parameter preserves existing single-workspace behavior. The 3 new tools are additive.

[2.12.0] - 2026-03-22

Features

• Git worktree support — Optimus now detects git worktrees and resolves .optimus/ paths accordingly. Shared resources (config, dist, roles, skills, memory) resolve from the main worktree, while runtime state (task manifests, results, reviews, agent instances) stays isolated per worktree. This enables running multiple features simultaneously on one machine.
• ensureWorktreeStateDirs() at server startup — The MCP server automatically creates required state directories when running inside a worktree, so optimus init in the main worktree is all you need.
• optimus init now detects worktree context — Shows informational output about worktree status and whether the main worktree has .optimus/ set up.

Architecture

• Centralized path resolution via resolveOptimusPath() — All .optimus/ path construction across 12+ source files now routes through src/utils/worktree.ts, replacing ~40 scattered path.join() calls. Paths are auto-categorized as shared or state based on the first directory segment.
• Path categories: config, dist, roles, skills, memory, specs, tasks → shared (main worktree). state, results, reviews, system, agents → per-worktree.

Compatibility

• Fully backward compatible — Non-worktree workspaces behave identically (both roots resolve to the same directory). No changes to CLI commands, MCP tools, or configuration formats.

[2.11.0] - 2026-03-22

Features

• First-class Agent Runtime abstraction — New application-facing runtime layer above raw delegation and transport. Host applications can now call agents as domain services without coupling to delegate_task, CLI transport details, or task-manifest internals. Addresses #516.
• 5 new Agent Runtime MCP tools — run_agent (sync execution), start_agent_run (async), get_agent_run_status (normalized polling), resume_agent_run (unblock manual intervention), cancel_agent_run (graceful cancellation).
• Normalized result envelopes — All runtime tools return a consistent AgentRuntimeEnvelope with status, result, error_code, error_message, requires_manual_intervention, action_required, and runtime_metadata (engine, model, session_id, duration_ms, retries_attempted).
• Structured output validation — When output_schema is provided, the runtime detects malformed JSON output and returns error_code: "invalid_structured_output" instead of silently passing broken data.
• Engine fallback and retry policy — run_agent supports runtime_policy.retries and runtime_policy.fallback_engines for automatic retry with engine rotation.
• Task cancellation support — New cancelled status in the task manifest. Running tasks respect cancellation; the async runner checks cancellation before writing final status.

Fixes

• Async tasks now record resolved engine, model, and session_id — Worker spawner backfills resolved_engine, resolved_model, and session_id into the task manifest after execution, enabling runtime metadata to surface actual execution details.
• completed_at timestamp added to task records — Terminal task states now record completion time for accurate duration tracking.

Compatibility

• Backward compatible with existing orchestration tools — delegate_task, delegate_task_async, dispatch_council, and check_task_status remain unchanged.
• Recommended for application-side integrations — Teams building domain services (script generation, classification, extraction, structured content generation) on top of Optimus should adopt the runtime tools for a stable contract.

[2.10.0] - 2026-03-21

Features

• Unified project MCP config is now first-class — Optimus now treats .optimus/config/mcp-servers.json as the single source of truth for workspace MCP server definitions instead of hardcoding .vscode/mcp.json as the canonical config.
• optimus init / optimus upgrade now generate multi-client MCP configs — Workspaces now get synchronized project-local MCP files for VS Code / GitHub Copilot (.vscode/mcp.json), GitHub Copilot CLI (.copilot/mcp-config.json), and Claude Code (.mcp.json) from one shared definition.

Fixes

• Claude Code and ACP flows no longer depend on VS Code config layout — Runtime adapters now prefer the canonical Optimus MCP config and only fall back to legacy client files, removing the old coupling to .vscode/mcp.json.
• Cross-platform workspace config is now portable across Windows and macOS — Generated Claude and Copilot CLI config files use project-relative paths, while VS Code keeps its native workspace macro format.

Compatibility

• Existing VS Code-based workspaces remain supported — Legacy .vscode/mcp.json workspaces still work as a fallback while newer workspaces can adopt the unified source config.
• Recommended for teams validating multiple MCP clients — This release is intended for shared testing across Claude Code, VS Code Copilot, and Copilot CLI with one project-owned configuration model.

[2.9.0] - 2026-03-19

Features

• Azure DevOps updateWorkItem is now implemented — The unified VCS layer can now update ADO work items through JSON Patch, including title, description, state, assignee, and priority changes.
• vcs_update_work_item now exposes richer ADO fields — MCP callers can pass description, assigned_to, and priority, while provider-specific workflow states are forwarded correctly for Azure DevOps.

Fixes

• GitHub update validation is stricter and clearer — GitHub work-item updates now reject unsupported state values and unsupported field-only payloads with actionable errors instead of failing ambiguously.
• Label-only GitHub updates no longer require a redundant issue patch — The provider now refetches issue data when only labels change, preserving the unified update flow without unnecessary PATCH requests.

Compatibility

• Backward compatible for existing GitHub workflows — GitHub retains its existing open/closed state model and label handling.
• Recommended for mixed GitHub + ADO environments — Teams using Azure DevOps now get parity for unified work-item updates without changing MCP tool names.

[2.8.0] - 2026-03-19

Features

• Available-agents resolution is now explainable at runtime — Engine selection no longer lives only in worker-spawner.ts. Optimus now exports a resolved explanation view with requested automation policy, candidate transports, final protocol choice, and selection reason.
• New explain_available_agents MCP tool — Agents and operators can query the fully resolved available-agents.json behavior directly, either for one engine or the full config, without re-implementing the resolver logic.
• roster_check now shows resolved runtime behavior — The roster output surfaces configured protocol, resolved protocol, requested automation, and why each engine resolves the way it does.

Fixes

• Explicit transport configs no longer masquerade as both protocols — Runtime transport selection now correctly distinguishes pinned cli versus pinned acp engine declarations instead of treating one top-level config as valid for both protocol branches.
• Engine resolution is now reusable instead of duplicated — Transport preview, selection reason, and runtime adapter resolution now share a single explanation path, reducing drift between human-visible summaries and actual execution behavior.

Compatibility

• Backward compatible for existing configs — Legacy available-agents.json shapes remain accepted, but the runtime surface is now more explicit and machine-readable.
• Recommended upgrade for agent-native workflows — If you rely on protocol: "auto", upgrade to get explainability and more accurate runtime introspection.

[2.7.0] - 2026-03-19

Features

• Copilot ACP transport is now first-class in engine config — GitHub Copilot can be declared with protocol: "auto" and preferred_protocol: "acp", with headless ACP launches defaulting to copilot --acp --stdio when ACP args are not explicitly provided.
• Automation policy is now normalized across engines — automation.mode and automation.continuation are treated as separate, explicit policy axes. Copilot CLI autopilot and ACP transport selection now resolve from the same config surface and are exposed consistently in the MCP entrypoints.
• Async council dispatch gets immediate queued artifacts — dispatch_council_async now creates STATUS.md and the task-manifest entry before background execution starts, improving observability for queued reviews.

Fixes

• Fail-fast automation/transport validation — Invalid combinations such as Copilot ACP + continuation: "autopilot" or Claude + unsupported autopilot continuation are now rejected during engine resolution instead of failing later at runtime.
• ACP session compatibility retries — AcpAdapter now retries session/prompt and session/load with compatibility fallbacks when ACP servers reject one parameter shape, and falls back to a fresh session when persisted session resume is not accepted.
• ADO links now resolve to browser URLs — ADO work item creation/comment APIs now return stable web URLs, including GUID-backed project resolution and correct work item comment anchors.
• Meta-cron and council capacity hardening — stale lock files are detected and cleaned more safely, malformed engine config entries are excluded earlier, and council async queue setup is covered by dedicated Vitest tests.

Compatibility

• Low-risk for default users; review custom engine configs — Legacy vendor-specific automation aliases are still parsed, but new configs should use normalized values such as auto-approve, deny-unapproved, single, and autopilot.
• Copilot autopilot remains CLI-only — ACP and autopilot are orthogonal concepts, but Copilot autopilot continuation is still only supported on the CLI transport. protocol: "acp" plus continuation: "autopilot" is intentionally rejected.

[2.6.2] - 2026-03-16

Fixes

• optimus init auto-fills GitHub owner/repo from git remote — No longer leaves owner/repo empty in vcs.json. Parses git remote get-url origin and fills github.owner and github.repo automatically. Fixes wrong-owner issue when init is run in org-owned repos.
• SAML SSO 403 errors get clear remediation message — GitHub 403 responses containing "SAML enforcement" now produce: "GitHub token is valid but NOT authorized for this organization via SAML SSO. Action required: authorize your token for the organization." Instead of generic "VCS unavailable". Applied to all 8 GitHub API call paths in GitHubProvider.

[2.6.1] - 2026-03-16

Fixes

• vcs_add_comment Markdown→HTML for ADO — Comments on ADO work items and PRs now auto-convert Markdown to HTML, matching vcs_create_work_item behavior. Fixes broken rendering of tables, headings, and formatting in ADO comments.

[2.6.0] - 2026-03-16

Features

• Agent/Role Lifecycle GC — Patrol new Phase 6.5: Agent Hygiene. Auto-cleans stale T1 agents (>7 days), orphaned lock files, and unused T2 roles (>30 days). Task manifest auto-trimmed to 30 days. (#475)

Fixes

• MCP server path resolution — optimus init/upgrade now copies mcp-server.js to .optimus/dist/ and uses ${workspaceFolder} relative path in mcp.json. Fixes MCP server not starting after running from npx/temp directory. (#477)
• npx ECOMPROMISED — Removed prepare lifecycle scripts that triggered cascading builds during npx install, exceeding npm v11 lock timeout. (#471)

[2.5.2] - 2026-03-16

Fixes

• AcpAdapter path resolution — loadMcpServers now uses OPTIMUS_WORKSPACE_ROOT env var for reliable path resolution, fixing issues where the workspace root was incorrectly inferred.

[2.5.1] - 2026-03-16

Fixes

• Scaffold release-policy.json missing from v2.5.0 — The auto-release opt-in config template (enabled: false) was referenced in the v2.5.0 release notes but was not included in the release commit. Added to optimus-plugin/scaffold/config/.

[2.5.0] - 2026-03-16

Features

• get_user_memory MCP tool — Master Agent can now read the same user preferences that sub-agents see. Single source of truth at ~/.optimus/memory/user-memory.md.
• ACP activity-based timeout — Detects live-but-stuck ACP sessions (process alive but zero output). Default 20 min, configurable per-engine via timeout.activity_ms. Fixes 5+ hour zombie tasks (#433).
• Delegation Scope Decision Matrix — Added to system-instructions and delegate-task skill. Master now knows when to delegate to PM vs dev vs specialist (#393).
• Competitive Discovery skill rewrite — Now auto-adds high-confidence competitors (score ≥5) to watchlist via Read-Modify-Write protocol. Medium-confidence candidates go through request_human_input.
• Auto-release skill + cron — New auto-release skill with release-gate cron (every 4h). Conventional commits determine semver. max_auto_bump: minor. Scaffold default: enabled: false (opt-in).

Improvements

• Master onboarding Step 0 — User Memory loading is now the first numbered step in master-onboarding skill for more reliable execution.

[2.4.0] - 2026-03-16

Features

• competitive-discovery skill — New weekly discovery skill for finding unknown competitors via GitHub topic search. Separate from daily monitoring. Weekly cron registered (0 9 * * 1).
• 9 new competitors added to watchlist — Zeroshot, ComposioHQ/agent-orchestrator, Babysitter, OpenCastle, Ruflo, TAKT, DeerFlow, AG2, Google ADK.

Improvements

• Unified prompt for all engines — Removed ACP lean prompt. All engines (CLI and ACP) now receive the same full prompt with inline memory, user memory, persona, skills, and context. Fixes: ACP sub-agents were missing user preferences and project memory because they didn't read the file paths provided in the lean prompt.

[2.3.2] - 2026-03-15

Fixes

• Patrol no longer auto-closes swarm-council Issues — Council verified only means the review finished, not that recommendations were implemented. Now only swarm-task verified Issues are closed. Council Issues get a status comment but stay open until manually closed.

[2.3.1] - 2026-03-15

Features

• optimus init auto-creates Health Log Issue — Detects GitHub repo from git remote, creates a "[Optimus] System Health Log" issue, and links it in meta-crontab.json. Patrol reports are appended as comments. Requires GITHUB_TOKEN in .env; skips gracefully if unavailable.

Improvements

• Scaffold default cron → hourly-patrol — New workspaces get hourly-patrol (patrol-manager + project-patrol skill) instead of night-steward. No dry-run, max_actions=999.
• VCS tools graceful degradation — All 7 VCS tools (vcs_create_work_item, vcs_create_pr, vcs_merge_pr, vcs_add_comment, vcs_update_work_item, vcs_list_work_items, vcs_list_pull_requests) now return ⚠️ VCS unavailable warning instead of crashing when no token or config is present. Agents continue working without Issue/PR tracking.

[2.3.0] - 2026-03-15

Features

• vcs_list_pull_requests tool — New VCS tool to list pull requests with state filter. Returns PR number, title, mergeable status (clean/CONFLICTING/unknown), head/base branches, and labels. Enables patrol PM to automatically discover and handle open PRs.

Improvements

• Patrol PR phase upgraded — Patrol skill now uses vcs_list_pull_requests instead of unreliable git log --remotes. Adds decision matrix for conflicting bot PRs (auto-close) and conflicting human PRs (escalate via request_human_input).

[2.2.1] - 2026-03-15

Fixes

• ACP sub-agents can now call MCP tools — AcpAdapter was passing mcpServers: [] to session/new, making sub-agents unable to call any MCP tools (e.g. request_human_input, vcs_update_work_item). Now reads .vscode/mcp.json, resolves ${workspaceFolder}/${env:VAR} macros, and converts VS Code object format to ACP array format. Verified with request_human_input end-to-end test.

[2.2.0] - 2026-03-15

Features

• vcs_update_work_item tool — New VCS tool to update work items (close/reopen issues, change titles, add/remove labels). Platform-agnostic — works on both GitHub and Azure DevOps.
• vcs_list_work_items tool — New VCS tool to list work items with state and label filters. Replaces legacy github_sync_board.

Breaking Changes

• Removed github_update_issue and github_sync_board — These legacy tool schemas had no handler implementation. Replaced by vcs_update_work_item and vcs_list_work_items which use the unified VCS provider layer.

Fixes

• Patrol PM was unable to close issues — Root cause: patrol skill referenced github_update_issue which had a schema but no handler. All skill references updated to use vcs_update_work_item / vcs_list_work_items.

[2.1.1] - 2026-03-15

Fixes

• Scaffold model name alignment — Scaffold available-agents.json claude-code models changed from ACP-native names (default, opus, etc.) to standard identifiers (claude-opus-4.6-1m, gpt-5.4, gemini-3-pro-preview) matching project config.

[2.1.0] - 2026-03-15

Features

• Autonomous Memory Writing — Agents now autonomously write to project/role memory when encountering non-trivial lessons (bug postmortems, tool gotchas, conventions). Removed the previous "MUST NOT write" restriction.
• User Memory auto-init — optimus init now creates ~/.optimus/memory/user-memory.md with a starter template. Also creates .optimus/memory/roles/ directory for role-level memory.

[2.0.0] - 2026-03-15

Breaking Changes

• Claude Code defaults to ACP protocol — claude-code engine now uses claude-agent-acp (ACP protocol) instead of legacy CLI text parsing. optimus init and optimus upgrade auto-install @zed-industries/claude-agent-acp. No Anthropic API key needed — uses existing Claude Code OAuth session.

Features

• ACP Lean Prompt — ACP engines receive a minimal prompt (role + task only). System instructions, memory, and skills are referenced by file path instead of inline injection. Reduces prompt from 40K+ chars to ~2K chars.
• AcpAdapter Permission Auto-Approve — Automatically approves session/request_permission requests from agents in headless mode. Fixes tool calls (file reads, etc.) hanging indefinitely.
• AcpAdapter Windows PATH Resolution — shell: true on Windows for ACP spawn, enabling simple command names like claude-agent-acp instead of absolute paths.
• 10-min Heartbeat Timeout Default — All engines in scaffold now default to 600,000ms heartbeat timeout.
• Competitive Intelligence v2 — Autonomous discovery with three-frequency model (daily monitor, weekly discovery, first-run bootstrap).

Fixes

• ACP complex task stall — Root cause: session/request_permission from agent was unhandled, causing indefinite hang on any task involving tool calls. (#440)

[1.2.0] - 2026-03-15

Features

• Configurable Task Timeouts — Engine-level and per-task timeout overrides via available-agents.json timeout block and timeout_ms parameter on delegate_task. Resolution order: per-task > per-engine > hardcoded fallback. (#407)
• AcpAdapter Claude Code Compatibility — AcpAdapter now supports claude-agent-acp (Zed's Claude ACP bridge). Handles cwd/mcpServers in session/new, array prompt format, and agent_message_chunk streaming. (#413)
• Patrol Auto-Close Task Issues — Patrol PM now correlates open swarm-task/swarm-council Issues with task-manifest.json and auto-closes verified ones. Issue lineage cascade notifies parent Issues. (#415)
• Competitive Intelligence Agent — New competitive-intel meta-skill + competitive-watchlist.json config. Daily cron monitors competitor repos for releases, star velocity, and architectural changes. Escalates notable findings via human-input-needed. (#419)
• Human Escalation Protocol — Agents now have explicit rules for when to call request_human_input: strategic decisions, merge conflicts, repeated failures, missing credentials, destructive operations, and unmerged verified work. (#425)

Fixes

• Council role_descriptions not passed to manifest — dispatch_council_async was not storing role_descriptions in the task manifest, causing new council roles (without pre-existing T2 templates) to consistently fail T3→T2 precipitation. (#428)

[1.1.0] - 2026-03-15

Features

• Task Dependencies — delegate_task_async now accepts depends_on: [task_ids]. Tasks register as blocked and auto-spawn when all dependencies are verified. (#395)
• Meta-Cron Session Persistence — Cron-triggered agents (e.g., patrol PM) now resume their previous session via last_agent_id, maintaining context across patrol cycles. (#398)
• Patrol PM Diagnose & Re-delegate — Patrol skill gains Phase 2.5: diagnoses open Issues by correlating with task manifest, auto-closes verified Issues, re-delegates failed tasks to last executor with session continuity. (#400)
• Generic Release Process Skill — release-process skill refactored into a config-driven meta-skill with release-config.json schema. Ships in scaffold for all users. (#403)
• Configurable Timeouts — Engine-level and per-task timeout overrides via available-agents.json and timeout_ms parameter on delegate_task. (#407)
• Landing Page Claude Teams Comparison — New "Optimus vs. Agent Teams" comparison section on both landing page and pitch deck.

[1.0.6] - 2026-03-15

Fixes

• T2 Guard regression fix — Thin T2 templates (existing but < 25 body lines, e.g., pm.md) were incorrectly triggering the new "missing role_description" error. Now: if a T2 file exists (even thin), delegation proceeds without requiring role_description. The error only fires when truly no T2 exists.

[1.0.5] - 2026-03-15

Features

• Self-Execution Pre-Flight Rule — Master Agent must now check .optimus/skills/ for a matching Skill before executing any multi-step workflow. Added to both project and scaffold system-instructions.

Fixes

• Council Issue title readability — dispatch_council_async now extracts the topic from the proposal's # PROBLEM: / # PROPOSAL: heading instead of using the filename. Issue titles change from [Council] 00 PROBLEM (Review) to [Council] Agent-driven automation architecture (Review). (fixes #392)

[1.0.4] - 2026-03-15

Features

• Multi-Level Memory System — New MemoryManager.ts with project + role memory scopes, YAML frontmatter parser, relevance scoring (+3 role match, +2 recency), tiered token budgets, and greedy-fill loader. append_memory now accepts level: "project" | "role". Workers receive OPTIMUS_CURRENT_ROLE env var. (#375)
• User-Level Cross-Project Memory — Opt-in ~/.optimus/memory/user-memory.md for persistent user preferences across projects. append_memory accepts level: "user" with write-time safety filtering. optimus memory CLI subcommands: init, list, edit, add, remove. CI-guarded (disabled when CI=true). (#380)
• Context Continuity System — New list_knowledge MCP tool returns structured manifest of all .optimus/ artifacts (specs, memory, reports, reviews). Master delegation pre-flight now includes Context Check step. Sub-agents can self-discover context. (#381)

Fixes

• Enforce T3→T2→T1 lifecycle — ensureT2Role now throws an error (instead of silent null) when role_description is missing for a new role. T1 instance creation skipped when no T2 exists, preventing orphaned empty-shell agent files.

[1.0.3] - 2026-03-14

Features

• Master Onboarding Skill — New master-onboarding skill teaches Master Agent the first-run protocol (read system-instructions, roster_check, Problem-First SDLC)
• Inject Template v2 — IDE configs now prompt "First action: read master-onboarding SKILL.md" on every session
• Role Creation Decision Rules — Master must roster_check before creating roles, no near-duplicates, require role_description

Fixes

• Auto-label all VCS items — vcs_create_work_item and vcs_create_pr now auto-append optimus-bot label

[1.0.2] - 2026-03-14

Features

• Skills Quick Reference expanded — Critical constraints (async-only, roster_check first, no busy-poll) inlined into system-instructions for delegate-task, council-review, git-workflow
• Sync dispatch_council warning — Returns a warning nudging users toward dispatch_council_async
• Docs updated for v1.0.0 — ARCHITECTURE.md, HYBRID_SDLC.md, building-autonomous-swarm-guide.md, IDEA_AND_ARCHITECTURE.md

Fixes

• Remove orphan code: MemoryManager.ts, SharedTaskStateManager.ts, Calculator.ts (1035 lines deleted)

[1.0.1] - 2026-03-14

Features

• Qwen Code Auto-Discovery: ACP engine config supports "path": "auto" — automatically scans ~/.vscode/extensions/ for Qwen Code CLI, no hardcoded paths needed
• Scaffold now ships qwen-code engine entry with auto-discovery (users just need Qwen Code VS Code extension installed)

Fixes

• Fix AcpAdapter TypeScript errors (invalid chat mode, undefined encodeMessage)
• Fix scaffold available-agents.json leaking local paths

[1.0.0] - 2026-03-14

Features

• ACP Protocol Support: Implement AcpAdapter with NDJSON-based JSON-RPC transport, verified with Qwen Code CLI
• Multi-ACP Vendor Architecture: Engine config uses protocol field to route adapters — add new ACP vendors (Kimi, Cursor, etc.) with zero code changes
• Problem-First SDLC Workflow: New lifecycle: PROBLEM → PROPOSAL → SOLUTION → EXECUTE, replacing the old proposal-first approach
• Artifact Directory Routing Table: Centralized routing rules in system-instructions.md — every agent knows what goes where
• YAML Frontmatter Format Templates: 7 artifact types with standardized templates (PROBLEM, PROPOSAL, SOLUTION, VERDICT, task result, report, review)
• Cross-Model Council Diversity: Automatic greedy round-robin assignment of engine:model combos to council participants
• VERDICT Template Externalization: Council verdict format loaded from .optimus/config/verdict-template.md (customizable)
• Mandatory Council Review Rule: High-impact changes require expert council review before implementation
• Engine Health Tracking & Fallback: Per engine:model health tracking with automatic fallback on failures
• Agent Pause/Resume: Human-in-the-loop input mechanism via request_human_input
• Patrol Manager System: project-patrol skill with automated cron-based project monitoring
• Docs Site Enhancements: Interactive tutorial guide, investor pitch page, bilingual EN/ZH support, terminal hero animation
• Init/Upgrade: Scaffold now includes specs/ and results/ directories, verdict-template.md, protocol field in config

Fixes

• Strip Tool-Call Traces from Output: Artifact files no longer contain adapter process traces (GPT-5.4/Claude Opus outputs now clean)
• ESM/CJS Bundling: Bundle all deps via esbuild to eliminate runtime crashes
• Windows Path Mangling: Prevent backslash issues in agent delegations
• Meta-Cron Race Condition: Fix concurrent cron execution and add log capture
• strip-ansi CJS Compatibility: Downgrade to v6 for CommonJS support

Improvements

• Consolidated System Prompt: Single Source of Truth at .optimus/config/system-instructions.md
• Frozen proposals/ Directory: New work goes to specs/, legacy proposals preserved
• Replaced protocol.md: Obsolete blackboard protocol replaced with redirect stub
• Release Process Skill v1.1: Fixed scaffold sync rules, added validation warnings

[0.4.0] - 2026-03-12

Features

• Inject project memory into agent prompts at spawn time (#181)
• Add release-process skill SOP (#189)
• Default to squash merge and sync local master after merge (#185)
• Add agent attribution signatures to all VCS tools (#152)
• Inject auto-created issue ID into agent prompt to prevent duplicates (#166)
• Implement agent retirement, quarantine and T1 GC (#161)
• Role-skill decoupling, enhance roster_check (#163)
• Ensure optimus-bot label on all auto-created issues (#153)
• Auto-delete source branch after PR merge (#150)

Fixes

• Preserve user config during optimus upgrade (#175)
• Prevent async council tasks from getting stuck in running state (#64)

Improvements

• Add systemic safeguards from vcs.json wipe postmortem (#178)
• Fix misleading MCP config section in Quick Start (#176)
• Update README and system-instructions for v0.3.0 features (#144)
• Enforce issue lineage in skills and instructions (#155)
• Remove auto-skill genesis (#160)

[0.3.0] — 2026-03-12

Features

• Delegation Depth Control — MAX_DELEGATION_DEPTH = 3 prevents infinite agent recursion. Tracked via OPTIMUS_DELEGATION_DEPTH env var.
• Plan Mode for Orchestrators — mode: plan strips write permissions from PM/orchestrator roles, forcing delegation instead of direct coding.
• write_blackboard_artifact MCP Tool — Allows plan-mode agents to write exclusively to .optimus/ directory with symlink-safe path validation.
• Issue Lineage Tracking — OPTIMUS_PARENT_ISSUE env var injected into child agents, enabling GitHub Issue parent-child tree visualization.
• optimus upgrade CLI Command — Safe incremental upgrade that force-updates skills/roles/config while preserving user agents and runtime data.
• Enhanced ADO vcs_create_work_item — New params (area_path, iteration_path, assigned_to, parent_id, priority), vcs.json defaults section, auto-tag created-by:optimus-code, Markdown→HTML body conversion.
• Auto-Skill Genesis — Auto-generate SKILL.md after successful T3 precipitation so new roles are born with operational playbooks.
• Rich T3→T2 Precipitation via agent-creator — Replace thin fs.writeFileSync templates with full agent-creator invocation for professional-grade role definitions.
• Engine/Model Validation — Validate engine and model names against available-agents.json before writing to T2 frontmatter, preventing invalid engine corruption.

Improvements

• Feature-dev skill rewritten with 6-phase PM-driven autonomous workflow
• Senior Full-Stack Developer role replaces generic dev
• Session Continuity Rule for agent memory reuse
• Environment variable interpolation fix for nested MCP worker processes
• Explicit tool call examples added to each feature-dev phase

[Unreleased]

Self-Evolving Agent System (T3→T2→T1 Complete Lifecycle)

• T3→T2 Immediate Precipitation: First-time T3 role usage auto-creates a T2 role template in .optimus/roles/. No threshold — instant on first delegation.
• T2→T1 Session Instantiation: When a task completes and returns a session_id, the system auto-creates a T1 agent instance in .optimus/agents/ from the T2 template. T1 is frozen after creation.
• Master-Driven T2 Evolution: Master Agent can update T2 templates with new role_description, role_engine, role_model via delegate_task params. T1 instances are never retroactively modified.
• Structured delegate_task Params: Added role_description, role_engine, role_model, required_skills fields. Master Agent provides all T2 info — no more guessing.
• Engine/Model Fallback Chain: Master override → frontmatter → available-agents.json → claude-code hardcoded fallback.

Skill System

• Skill Pre-Flight Check: required_skills field in delegate_task/delegate_task_async. Missing skills → rejection with actionable error listing what to create.
• Skill Auto-Injection: Found skills automatically injected into agent prompt as EQUIPPED SKILLS section.
• skill-creator Bootstrap Meta-Skill: Teaches agents how to create new SKILL.md files.
• agent-creator Bootstrap Meta-Skill: Teaches Master Agent the T3→T2→T1 lifecycle, role selection, engine binding.

Infrastructure

• DOTENV_PATH via mcp.json env mount: Replaces hardcoded .env path. Users can point to any env file.
• Auto-generate .vscode/mcp.json: optimus init creates or merges MCP config for VS Code/Copilot users.
• [Optimus] Auto-Tagging: All Issues/PRs created via MCP tools get [Optimus] prefix and optimus-bot label.
• Zero-Config Scaffold: optimus init ships no pre-built roles/agents. System bootstraps at runtime.
• Inject-Only Instruction Bridging: optimus init appends reference to existing CLAUDE.md/copilot-instructions.md but never creates new ones.
• Windows CRLF Fix: parseFrontmatter normalizes \r\n for cross-platform compatibility.
• Path Traversal Prevention: sanitizeRoleName() strips dangerous characters from role names.
• T3 Log File Mutex: Prevents concurrent write corruption on t3-usage-log.json.
• windowsHide: true: Background child processes no longer pop up terminal windows.

Removed

• Lazy-sync of built-in roles to user projects (was polluting .optimus/roles/ with phantom T2 files)
• Instruction bridging that copied full system-instructions.md content into CLAUDE.md
• Threshold-based precipitation (was 3 invocations + 80% success rate, now immediate)

[0.0.8] - 2026-03-08

• Enhancement: Planner Consensus Voting Threshold: _computeIntentFromPlanners() now requires min(2, numPlanners) agreeing votes before routing to action or skip. Single-planner setups behave as before; with 2+ planners, at least 2 must agree, preventing a single aggressive planner from overriding the majority.
• Fix: Plan Synthesis Without Dedicated Executor: In plan mode with multiple planners but no executor configured, the synthesis step now falls back to the first plan adapter (invoked in read-only plan mode). Previously, synthesis was silently skipped, forcing users to manually reconcile planner outputs.
• UI: Plan Synthesis Label: The synthesis output is now labelled 📋 <AgentName> (Plan Summary) in the UI to distinguish it from individual planner outputs.

[0.0.6] - 2026-03-08

• Feature: Smart auto-routing (inferMode): When mode is set to Auto (default), prompt heuristics now automatically route pure questions to Plan mode and short explicit edits (with prior context) to Direct mode. Reduces unnecessary planner+executor round-trips for straightforward interactions.
• Feature: Prompt prefix shortcuts: Type /plan <prompt> or /exec <prompt> (or /direct) to override the mode for a single submission without touching the mode buttons. The prefix is stripped before the prompt reaches the host.
• Feature: Mode inference indicator: When smart auto-routing changes the mode, a small inline notification ("⚡ Auto-routed to Plan mode") appears in the chat so the user knows which execution path was taken.
• Feature: Three-mode execution routing: Added a Plan / Auto / Exec toggle in the input area. Plan runs planners only (analysis without code changes). Auto (default) runs the full planner → executor pipeline. Exec skips planners and sends user prompt directly to the executor for immediate action. Mode selection is preserved in queue items.
• Feature: Queue persistence: Pending queue items are now saved to VS Code globalState and restored when the webview reloads or VS Code restarts, preventing queued prompts from being lost.
• Feature: Queue content visibility: A collapsible panel below the Queue button shows all queued prompts (truncated to 60 chars) with per-item delete (✕) support. Click the badge count to toggle.
• Fix: Execution Trace leak in summary: ExecutorOutcomeRecord.summary now uses the clean output (post _extractThinking) instead of execCleaned, preventing tool trace markers (•, ✓, ✗, ↳) from leaking into the task state summary strip.
• Fix: Multi-line process line handling: appendProcessLines now splits multi-line entries (e.g. "• tool\n↳ summary") into individual lines before dedup, preventing malformed process text accumulation.
• Fix: Consistent result preview label: Unified first= to preview= in summarizeStructuredToolResult for multi-line results, matching the format already used by bash/shell tool-specific summarization.
• Fix: Copilot adapter missing captureProcessLinesAfterOutputStarts: GitHubCopilotAdapter.extractThinking() now sets captureProcessLinesAfterOutputStarts: true, so tool trace lines appearing after LLM output starts are correctly moved to thinking instead of leaking into output.
• Fix: Incomplete process line regexes: Both COPILOT_PROCESS_LINE_RE and CLAUDE_PROCESS_LINE_RE now include ↳, ✓, and ✗ characters, ensuring continuation and completion markers are recognized as process lines.
• Fix: Defense-in-depth in _summarizeText: _summarizeText() now strips tool trace lines (starting with •, ✓, ✗, ↳, etc.) before collapsing whitespace, preventing any residual trace markers from reaching the summary.
• Fix: History resume scrolls to last 3 turns: restoreTaskSessions now positions the scroll to the most recent 3 turns instead of jumping to the top.
• Fix: Direct mode hides dropped planner indicator: The "is not currently available" notification for dropped planners is now suppressed in Direct mode to avoid confusing messages.

[0.0.5] - 2026-03-08

• Repo Cleanup: Removed debug log files, temp test scripts, and temp_debug/ directory from root.
• Gitignore Fix: Replaced garbled .gitignore entries with clean rules; added patterns for *.log, [LOG], temp/, temp_debug/.

[0.0.4] - 2026-03-08

• Modes-aware Agent Selection: Agents now declare modes: string[]. Planning agent checkboxes show only agents with "plan" mode; executor dropdown shows only agents with "agent" mode.
• Council → Executor Two-Phase Pipeline: Phase 1 runs all selected planners concurrently; Phase 2 feeds synthesized plans to a single designated executor in agent mode.
• Multi-turn Shared Task State: SharedTaskStateManager tracks task history, open questions, blocked reasons, and auto-compacts context when token usage exceeds threshold.
• Active Editor Context Injection: The visible code range (or selection) from the active editor is automatically prepended to every planner prompt.
• Streaming Callbacks Debounced: Per-agent streaming updates are debounced at 100 ms to reduce webview re-render churn.
• Image Paste Support: Users can paste images into the prompt bar; images are saved to global storage and their paths are injected into the prompt.
• Apply Code Block: Executor output code blocks can be applied directly to workspace files via a one-click button.
• Task Resume / Rename / Delete / Pin: Full task lifecycle management from the history panel.
• Debug Mode: Optional optimusCode.debugMode setting surfaces per-adapter command, cwd, pid, and token usage in the UI.
• Code Quality Fixes: Replaced inline ANSI regex with shared ANSI_RE utility; extracted _buildDebugObject helper to eliminate duplicate debug-info construction; fixed ReturnType<...>[0] type annotations to use AgentAdapter; removed TOCTOU race in image storage; capped history view at 20 turns; added random suffix to image filenames to prevent collisions.

[0.0.3] - 2026-03-07

• Architecture Direction Clarified: Documented the long-term move toward app-level multi-turn shared task state, including structured planner contributions, resumable task history, and executor context synthesis owned by the orchestrator layer.
• UI Evolution: Abandoned the vertical chat layout in favor of a Side-by-Side "Kanban Card" approach, allowing multiple models to dynamically run horizontally without drowning out the conversation. Added rich-text Markdown formatting using marked.
• Architectural Shift - Dynamic Configuration: Removed hardcoded Agent registries. The extension now deeply integrates with VS Code workspace.getConfiguration, enabling users to swap Models (gpt-5.4, claude-opus-4.6, etc.) on the fly without a reload.
• Session History Persistence: Added a "History" toggle to the chat sidebar. Conversations and agent responses are now persisted in context.globalState across VS Code sessions. Users can pull up previous architectural discussions instantly.
• Architectural Shift - Streaming Execution: Completely refactored the underlying OS execution layer from synchronous cp.exec to asynchronous continuous streams via cp.spawn. This immediately pumps stdout/stderr token-by-token directly to Webview.
• Removed Hard Timeouts: Set timeout: 0 for all CLI tool executions, giving advanced Coding Agents the infinite runtime needed to utilize their own terminal-based "tools/skills" before formulating an output.
• Structural Standardization: Refactored project directory structure, moving extension components into src/providers/ and standardizing the architecture according to project documentation.
• Dependency Cleanup: Removed DoubaoAdapter place-holder to keep the underlying structure clean but natively backwards compatible to ANY future adapter type string inputs.

[0.0.2] - 2026-03-07

• Implemented asynchronous streamed UX updates. Instead of blocking the UI until all models finish, the chat interface now renders a dedicated task list indicating which agents are "Thinking" (🔄), completed (✅), or failed (❌).

[0.0.1] - 2026-03-07

• Initial MVP creation.
• Shifted architecture from command-palette invocation to a persistent Sidebar Webview Chat.
• Implemented Adapter Design Pattern in src/adapters/.