chore(claude): replace agent-framework with hivemind plugin config

[brenpike]

This repository migrated from the agent-framework Claude Code plugin to hivemind. The migration was applied to the working tree but never committed, and the old plugin's references were never removed — so CLAUDE.md has been describing a plugin this repo no longer uses, while the actual configuration sat uncommitted.

Removes every agent-framework reference

• CLAUDE.md — the Multi-Agent Governance section is rewritten for hivemind. Agent and skill names were verified against the installed plugin's agents/ and skills/ directories rather than recalled; this file is the repository's agent-instruction file, so a name that doesn't exist would silently misroute future sessions. The JSON snippet now also shows the "agent" key, which settings.json has always set but the doc never mentioned.
• .claude/settings.json — drops the dead "agent-framework@brenpike": true enable.
• docs/performance-todo.md — "a planner or coder agent" named two old plugin agents. Reworded to "an implementing agent" rather than substituting the current plugin's names, so the doc doesn't need editing again after the next migration.
• .claude/settings.local.json (gitignored, not in this diff) — six stale Skill(agent-framework:*) grants removed locally.

A working-tree grep now returns zero occurrences outside .git/.

Commits the pending plugin configuration

• .gitignore gains .hivemind/ and .claude/worktrees/, plus an EOF-newline fix. This is load-bearing: without it, run ledgers and brood worktrees show up as uncommitted repo content.
• .claude/settings.json — the hivemind enable, "agent": "hivemind:overlord", the caveman and codex plugin config, the SubagentStart hook, and the permission allowlist.
• .claude/hooks/caveman-ultra-subagent.sh — previously untracked, and referenced by that hook. Committing the hook config without the script would have broken the repo for any other clone.

Two defects fixed along the way

A dead permission rule. settings.json carried Bash(node /path/to/.claude/plugins/cache/openai-codex/codex/*) — /path/to/ is a literal unsubstituted placeholder inherited from a plugin seed template, so the rule matched nothing. It's removed rather than corrected: the machine-correct path would embed an absolute home directory into a public repository. The working rule now lives in the gitignored settings.local.json, which is where this repo already keeps per-machine grants.

A latent CRLF breakage. core.autocrlf is true and the repo had no .gitattributes. Committing the first tracked .sh file without one means a future clone receives #!/usr/bin/env bash\r and the hook fails to execute — a breakage that would never reproduce in the current working tree. A minimal *.sh text eol=lf rule pins it. Deliberately scoped to that one pattern: a broad * text=auto in a repo with existing checked-in CRLF content could rewrite line endings across many files on the next checkout.

Verification

• grep -rI "agent-framework" over the working tree → zero hits
• jq . on both settings files → valid
• bash .claude/hooks/caveman-ultra-subagent.sh | jq . → emits well-formed SubagentStart JSON, so the hook is functional as committed, not merely present
• git cat-file -p :.claude/hooks/caveman-ultra-subagent.sh → shebang stored with \n, confirming the .gitattributes rule applies
• git ls-files -s → hook recorded 100755, executable bit preserved
• dotnet build -c Release → 0 errors, 3 pre-existing warnings

Versioning

None. No files under src/ are touched, so the version-check gate short-circuits. Chatter.Rest.Hal stays 2.1.0, Chatter.Rest.Hal.CodeGenerators stays 0.4.0. No CHANGELOG entries — those track package behavior, not repository tooling.

Note

The hivemind seed template still ships the /path/to/ placeholder, so re-running seed-hive against this repo would re-add the dead rule. Worth fixing upstream in brenpike/hivemind.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: aba64f904c

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".