v0.5.1

What's Changed

• worktreeOpencodeConfig — per-loop inline opencode config written to worktree opencode.jsonc (git-excluded, removed before teardown commit to keep loop history clean). MCP servers declared here run as host processes from the worktree directory.

Full changelog

• feat: add worktreeOpencodeConfig for per-loop opencode config (#63)
• chore: bump version to 0.5.1

v0.5.0

Highlights

Parallel feature groups — A new launch-group workflow takes a broad source of work (a PRD, a set of GitHub/Linear issues, a milestone) and splits it into implementation-coherent features. Each feature is planned by the architect-auto agent and runs as its own loop, scheduled with a per-group concurrency cap. Track and control runs with group-status and group-cancel.

Redesigned dashboard — Rebuilt on SolidJS with a live activity feed and session observability. Loop and section views were restructured, and live transcripts now stream over SSE with per-session filtering.

Also new

• new-session loop mode — execute-plan can launch a plan in a fresh standalone session instead of a git worktree loop.
• Post-action phase — new loop state for review/fixup after the main work completes.
• Custom sandbox mounts — configurable bind-mounts for sandbox containers; tool-output directory now exposed to loop/audit sessions.
• Model variant defaults — model/variant selection driven by config priority.
• forge install CLI — installs bundled prompts and feature modules, kept in sync via content-hash manifests.

Fixes & internals

• Deeply-nested sub-agent sessions resolve correctly; TUI loop follower no longer auto-follows child sessions.
• Sandbox reliability: sandbox.enabled honored on attach, Docker socket reachable and pnpm store writable as non-root.
• Test suite migrated from bun:test to Vitest.
• README split into focused reference docs (configuration.md, sandbox.md, tools.md).

v0.4.21

Full Changelog: v0.4.20...v0.4.21

v0.4.20

feat: add restart support for execution service with forge naming utilities

v0.4.19

What's Changed

• Dashboard: Add copy buttons to markdown sections
• Review tool fix: Scope review-read to non-loop findings when outside a loop
• Review tool enhancement: Allow review-read to target findings from a specific loop by name

Full changelog: v0.4.18...v0.4.19

v0.4.18

Full Changelog: v0.4.17...v0.4.18

v0.4.17

Full Changelog: v0.4.16...v0.4.17

v0.4.16

Fix: exclude minified marked.js from lint, fix Solid reactivity warning

v0.4.15

feat: improve audit loop with coder decisions and finding recurrence tracking

• Coder decisions: coding agents can now document their decisions and verification methods via HTML comments, which the auditor prompt includes to avoid re-reporting explained findings.
• Finding recurrence tracking: bug findings that recur across audit rounds are automatically escalated after 3 occurrences, appearing as "Recurring blocking findings" in coder prompts and "Recurring findings — re-evaluate" in auditor prompts. Recurrence resets when a finding is resolved or a section passes audit clean.
• Sandbox host fallback: glob/grep tools with absolute paths outside the workspace mount (e.g., tool-output directories) now defer to host execution instead of erroring inside the container.

v0.4.14

Fixes

• Add PRAGMA busy_timeout=5000 to the readonly forge.db connections in the TUI loop store and dashboard launcher, preventing intermittent "database is locked" errors when reads collide with WAL checkpoints during an active loop. The dashboard's long-lived reader made this previously latent collision likely to surface.
• Suppress git stderr in worktree cleanup and sandbox manager git probes so handled non-zero exits no longer leak fatal: not a git repository to the terminal.