v0.6.0 — flow chain + one skill tree for Claude Code and Codex + two-question /bootstrap

Added

• One skill tree for every agent. Skills live once in .agents/skills/<name>/; .claude/skills/<name> is a relative symlink into it, so Claude Code, Codex, Cursor and other AGENTS.md-aware tools load the same files. AGENTS.md stays a symlink to CLAUDE.md. New scripts/link-agent-skills.mjs keeps the links in sync, run by the pre-commit hook.
• Planning and build skills (the flow chain). /flow picks the path; /grilling, /grill-me and /grill-with-docs sharpen an idea by interview; /to-spec and /to-tickets turn it into spec.md and tickets/NN-slug.md in the project; /implement builds one ticket per session. On-ramps: /triage, /diagnosing-bugs, /wayfinder. Underneath: /domain-modeling, /codebase-design, /tdd, /code-review, /prototype, /research, /handoff. All resolve the project and paths from docs/agents/project-context.md. Adapted from mattpocock/skills (MIT, see THIRD-PARTY-NOTICES.md). With /entity-research below, the bundle grows from 29 skills to 47.
• /entity-research — company and person research from public web sources, with citations.
• Codex gets the same MCP servers. .codex/config.toml is generated from .mcp.json (API keys forwarded, opt-in servers left commented out; Slack isn't available in Codex yet) and the pre-commit hook regenerates it once you've run bash scripts/install-git-hooks.sh; node scripts/codex-config-from-mcp.mjs refreshes it by hand. Hooks stay Claude Code only.
• /migrate-work reads Codex sessions too, alongside Cowork, the Code tab and Claude Code. A folder you worked on from both Claude and Codex shows up as one project, and the census says which session stores it found. It also brings over what Claude and Codex already remember about you (your global instructions, Claude Code's auto-memory, Codex's memories): facts about you go to USER.md, how the assistant should work to SOUL.md, project notes to that project's memory.md. Every line is shown first, appended word for word, and duplicates, conflicts and anything that looks like a credential are held back.
• Hooks: inject-projects (SessionStart — injects the active-project and Area-app table) and workspace-guards (append-only daily memory, no outputs at the repo root). Both read workspace paths from the Configuration block.

Changed

• Configuration: workspace.coding and the code_projects index are gone. Code lives in workspace.area_apps (shipped Area apps) and workspace.project_apps (a COD project's app/), both gitignored.
• /bootstrap re-runs by deleting the setup_completed line; there is no --reconfigure flag.
• /bootstrap asks two questions (who you are; your assistant's name and feel) and works out the rest: git name and email, timezone and language (a bundled script), plus which connectors are live. The script fills the persona files and Configuration block and verifies nothing is left unresolved. A persona file that no longer matches its template keeps your text, with its names filled in if you choose. Writing-voice capture, /migrate-work and the Slack, Atlassian and Figma connectors are offered at the end instead of asked up front; the connectors are added by apply.py add-connectors, which also refreshes the Codex config. Fixed: fresh forks were treated as already configured, because the template's placeholder date counted.
• external-action-guard also asks before Gmail send/reply/forward, iMessage and Slack canvases.
• pre-compact auto-handoffs record tool calls, edited files and errors again; tldr-read works with tldr 1.5.2.
• Skill evals no longer ship (maintainer test data); write your own with /skill-creator.
• /briefing reads mail and calendar through your Gmail and Calendar connectors (connect Google Workspace with /mcp) instead of the gws CLI.
• /migrate-work on a workspace that hasn't been through /bootstrap stops and points you there, instead of writing raw <placeholders> into your files.
• /migrate-work stays readable on a machine with years of history: past 40 memory lines, the approval shows each destination's count and first lines and puts the full numbered list in a file you can open; candidates beyond four are picked by number. Memories that are about a different person or assistant are held back, quoted YAML values are unescaped, and dated journal files in a memory folder are skipped.
• scripts/install.sh now installs the git hooks and the packages for the bundled image-analysis server, and the README setup runs bash scripts/install-git-hooks.sh.

Removed

• Hooks post-edit-diagnostics and auto-handoff-stop.
• The 3-Coding/ folder from the README map and every folder README.
• The .claude/commands/use-design.md slash command, which hardcoded workspace/4-Resources/; the /use-design skill does the same job and follows your Configuration. Delete the command file from a 0.5.x fork so the skill answers /use-design.
• The oracle and worker subagents in .claude/agents/: oracle searched only through the Python sandbox removed in 0.5.0, and worker needs an orchestrator that doesn't ship. Delete .claude/agents/ and .claude/tools/ from a 0.5.x fork unless you use them.

Fixed

• THIRD-PARTY-NOTICES.md now credits the 14 design-* skills (adapted from open-design, Apache 2.0) and skill-creator (Apache 2.0), with what was changed. Earlier releases shipped the design skills without their license.
• .claude/settings.local.json (your own permission grants) is now gitignored, so git add -A no longer picks it up.
• .claude/settings.json no longer ships permission rules or enabled plugins from the maintainer's own setup.
• In Codex, gemini-vision starts from the repo root, so it works when Codex runs in a subfolder.
• /code-review now shows up in Codex: its description wasn't valid YAML, which Claude Code tolerates and Codex silently skips.
• gemini-vision failed to start on a fresh clone because its packages were never installed; the README and .claude/mcp-servers/README.md now say to run npm install --prefix .claude/mcp-servers once.
• scripts/install.sh on a Mac without Homebrew: brew wasn't found right after installing it, every failed step printed die: command not found, and a missing gh failed the whole install. All three fixed.
• Hooks treat an unfilled <workspace.root> as the default workspace, so they work before /bootstrap; creating today's memory/YYYY-MM-DD.md with Write is no longer blocked (overwriting an existing one still is); tldr-read no longer runs the file path through a shell and only pre-approves reads inside the project.
• .gitignore keeps code (app/, apps/) and handoffs ignored after you rename the workspace folder.
• Skills no longer say HQ, /retro-project, /contact-research or /bootstrap (TBD); /archive-project's last step no longer contradicts its index step; /new-project's COD- rule matches where code lives.
• The docs now match the thin-Areas model from 0.5.0. The CLAUDE.md template, the new-project CLAUDE.md template and the folder READMEs had still described Areas holding a resources/ folder and archived projects promoting files into them; now an Area holds only its rules, memory and generated index, and archived projects move whole. They also use the PROJ-/COD-/VID- project names /new-project creates, without pre-made inputs/working/outputs folders.
• INSTALL.md matches the two-question /bootstrap: no tool-probe panel and no connector picker. /bootstrap no longer reads the installer's SBOS_TIER marker. It also now says to run bash scripts/install-git-hooks.sh once.
• Docs no longer point at skills that don't ship (/meeting-prep, /contact-add, /weekly-digest, /standup, the old ContinuousClaude skill list).
• /os-guide points at sections that exist in a fork (several rows cited maintainer-only README sections) and answers "how do I bring in my Codex work" and "does Codex get my MCPs".

Upgrading from 0.5.x

Skills moved from .claude/skills/ to .agents/skills/. Skills you never edited switch over on their own; the steps below cover the rest.

1. Before you pull, copy any skill you edited to a folder outside the repo. An edited skill conflicts with the move, and the pull replaces it.
2. Pull, taking the incoming version on conflicts in .claude/skills/, .agents/skills/ and .claude/settings.json; the old settings point at two hooks that no longer exist. Re-add any settings of your own afterwards. If CLAUDE.md conflicts (it will if /bootstrap filled it in), keep your values.
3. Update the Configuration block in CLAUDE.md: delete workspace.coding and the indexes section, and add workspace.area_apps and workspace.project_apps from the new template. Move anything in 3-Coding/ into an Area's apps/<name>/ (shipped apps) or a project's app/ (code still being built); both are gitignored.
4. Put your edited skills back under .agents/skills/<name>/, run bash scripts/install-git-hooks.sh once so the pre-commit hook keeps skills linked from now on, then run node scripts/link-agent-skills.mjs.
5. Restart Claude Code so the new hooks load.

Full changelog: v0.5.0...v0.6.0