v1.1.0 — Eval harness repair, PR-task benchmark, work-item E2E driver

Highlights

Eval harness works again on the runtime-backend architecture. npm run eval had rotted after the execution-backend split — it now registers local/docker backends, builds config per scenario so config_overrides.AGENT_COMMAND_TEMPLATE actually reaches the coding agent (per-scenario model switching verified live), and prints a per-category summary.

Robust judges. Diff judge resolves origin/<base> with fetch fallback (single-branch clones), gate verdicts reconstruct from events.jsonl (checkpoint.json was never written on local runs), llm_judge survives fenced/truncated verdicts, and a new expected_outcome judge lets 'correctly did nothing' PASS.

PR-task benchmark suite (evals/benchmark/, docs/benchmark.md): 10 scenarios against a real Rails app across delivery / exploration / clarification / scope. Sonnet 4.6 baseline: 9/10 — the failure is bench-impossible, where the agent invents a nonexistent feature (~80x cost of honest work) instead of flagging it. A/B model comparison documented.

Work-item label-walk E2E driver (npm run label-walk): drives the feature-delivery flow (ai:assist adoption → QA/UAT comment → squash + automerge) against a real GitHub PR; verified on EpicCoders/pxls#840. Documents that labels only set flags — state advances on live CI-green events.

Also merges the NetsoftHoldings master lineage (61 commits) to align public and private repos.

v1.0.0 — First Public Release

Gooseherd v1.0.0

First public release — self-hosted AI coding agent orchestrator.

Highlights

• Pipeline engine with configurable YAML pipelines (code, verify, fix loops)
• Slack integration — trigger runs from Slack, get PR links back
• Browser verification — Stagehand + video recording of deploy previews
• Observer daemon — auto-trigger from Sentry errors, GitHub events, Slack messages
• Real-time dashboard — run tracking, agent activity replay, cost analytics
• Sandbox isolation — Docker-out-of-Docker with non-root user
• Scope judge — AI-powered PR scope review

Docker

docker pull ghcr.io/chocksy/gooseherd:latest
docker pull ghcr.io/chocksy/gooseherd-sandbox:latest
cp .env.example .env  # add your tokens
docker compose up -d
open http://localhost:8787

What's new since pre-release

• Docker entrypoint with gosu for proper socket permissions (OrbStack/Docker Desktop)
• Graceful degradation when Docker socket unavailable (no more crash loops)
• Dashboard error display fix for New Run modal
• OCI labels for GHCR package ↔ repo linking
• Full deployment guide (docs/deployment.md)