v0.3.0 — Token savings (supervisor −88% in the benchmark) + 0.2.0 handoff

Single release covering both the token-savings work and the 0.2.0 handoff work (v0.2.0 was never tagged, so it ships here — see the changelog).

Install the tested version:

git clone --branch v0.3.0 https://github.com/mrchatam/Grok-workhorse.git grok-workhorse

Token savings (0.3.0)

• wait_task: long-poll one task or a batch until finished/parked (~45 s cap to stay under MCP timeouts) — no more task_status polling
• view: "brief" for task_result / wait_task: ~0.5–1 KB decision-ready result (verdict, summary, files, tests, next action + resume call)
• Compact JSON from the MCP shim (9–18% fewer characters)
• delegate_tasks: up to 10 tasks per call
• Presets and size routing, automatic fix rounds / escalation / cheap auto-reviews with hard token & cost caps
• usage_report / workhorse stats: worker tokens, cost, and a conservative estimate of supervisor tokens avoided
• Opt-in worker savers: terse / minimal_code prompt fragments and rtk command rewriting (--token-savers)
• Operator gate for approvals (approvals.require_operator), confirmed by a human through the workhorse approve CLI with a request id and operator token

Measured: supervisor tokens read per successful task dropped from ~3,060 to ~360 (about −88%) in the stub-backend benchmark. Details and method: docs/token-savings.md, Measured savings.

Task handoff (0.2.0, included)

• Structured handoff record on every finished task: state, failed checks, owner, the one exact next_action, and a suggested resume call
• Parked needs_approval / needs_input states with update_handoff / approve_task
• New filters and CLI: workhorse attention, handoff, approve, reject, tasks --status/--owner
• Fixes: graceful-shutdown finalization, restart socket race, cleanup races, audit log field collisions

Compatibility

No breaking changes to statuses, verdicts, or tool arguments; everything is additive.

Verification

Unit tests, the stub end-to-end suite, and the pinned-CLI lockfile checks all run in CI on this commit. Local run of npm run test:unit: 50/53 pass; the 3 failures are environment artifacts from this checkout's path containing a space (new URL(...).pathname becomes %20), not present in CI.

Grok Workhorse v0.1.0

First public release.

Added

• Stdio MCP server (workhorse-mcp) with list_repos, list_models, delegate_task, task_status,
  task_result, task_details, continue_task, cancel_task, list_tasks and cleanup_task,
  backed by one persistent daemon (unix socket plus token).
• Pluggable worker backends. kilo and opencode adapters are tested against the full mock
  integration suite and a live hello task. claude-code (PreToolUse guard hook) and codex adapters are implemented but
  untested. gemini and aider are TODO skeletons.
• A git worktree per task, an outer bwrap sandbox per task (hides the data dir, other tasks, the
  secret store, the main clone, /home, /tmp, …), per-task agent session data, and a test sandbox
  with no network.
• Guard plugin for Kilo/OpenCode (git mutation, network clients, secret access, paths outside the
  worktree) and the same rules as a Claude Code PreToolUse hook.
• Profiles for any OpenAI-compatible provider, with fallback chains and retries with back-off. Example
  profiles for NVIDIA NIM, OpenRouter and custom endpoints.
• Credentials from the daemon env, the MCP connector env, or an optional secret-store fallback.
• Structured results with daemon-run tests, verdicts, integrity checks, usage and timings, plus an
  append-only audit log.
• Stall detection (15 min), timeouts, cancel, recovery after restart, retention sweeps and
  workhorse health.
• Idempotent installer (scripts/install.sh) with pinned Kilo CLI, optional pinned OpenCode, config
  locking, a hello-world smoke task and optional systemd; uninstall.sh.
• Draft Grok Bot skills in grok-template/ (getting started, delegation).