consolidate: add project-agnostic multi-agent kit (.claude/EN/)

[AdaWorldAPI]

Summary

Adds the project-agnostic multi-agent kit distilled from AdaWorldAPI/WoA

• AdaWorldAPI/woa-rs on 2026-05-17 (reference handover:
  META/HANDOVER-WOA-RS-AGENT-HARDENING-2026-05-17.md).

This PR is additive only. It does not replace, rename, or remove
any pre-existing .claude/ content in this repo.

What's added

.claude/EN/
├── README.md                                          how to adopt the kit
├── CLAUDE-AGENT-PATTERN.md                            agnostic cheat-sheet
├── knowledge/
│   ├── autoattended-multi-agent-pattern.md           6-step loop, 4-savant
│   │                                                  taxonomy, worker iron
│   │                                                  rules, board pattern
│   └── a2a-workarounds.md                            file-blackboard, branch
│                                                      pub/sub, role-teleport,
│                                                      handover files
└── agents/
    ├── README.md
    ├── worker-template.md                             slot-based worker brief
    ├── meta-agent.md                                  13th agent (review + inbox)
    ├── brutally-honest-tester.md     (PP-13)          POST-IMPL gate
    ├── baton-handoff-auditor.md       (PP-15)         DURING-IMPL gate
    └── preflight-drift-auditor.md     (PP-16)         PRE-SPAWN gate

Also adds .claude/settings.json with a permissive ** allowlist for Read/Write/Edit + cat/tee/python/bash + per-language toolchains (Rust / Python / TypeScript / Go), with a deny list that preserves .git, .archive, credentials, and force-push.

Provenance

Distilled from:

• AdaWorldAPI/WoA .claude/EN/knowledge/*.md (canonical agnostic docs)
• AdaWorldAPI/WoA .claude/EN/knowledge/RUST/agents/*.md (genericized to remove Rust-specific bits)
• AdaWorldAPI/WoA .claude/CLAUDE.md §3.5 + §3.6 (extracted as standalone EN doc)
• AdaWorldAPI/lance-graph .claude/agents/ (most mature 4-savant ensemble — used as comparison reference)
• AdaWorldAPI/ndarray .claude/agents/ (compact 7+3 ensemble — used as comparison reference)

Full distillation rationale in META/HANDOVER-AGENTKIT-CONSOLIDATION-2026-05-17.md (in WoA).

How to adopt in this repo

After this PR merges:

1. Read .claude/EN/CLAUDE-AGENT-PATTERN.md (the cheat-sheet)
2. For multi-agent sprints, read .claude/EN/knowledge/autoattended-multi-agent-pattern.md
3. For cross-session coordination, read .claude/EN/knowledge/a2a-workarounds.md
4. When spawning workers, fill the slots in .claude/EN/agents/worker-template.md

Repo-specific Iron Rules stay in .claude/CLAUDE.md (not touched by this PR).

Test plan

• Files render correctly in GitHub UI (Markdown)
• .claude/settings.json parses as valid JSON
• No existing .claude/ content was removed
• Reviewer verifies the kit makes sense for this repo's domain

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 982c9f235e

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".

[chatgpt-codex-connector]

Remove shell entries that bypass the deny list

In any Claude session using this settings file, allowing Bash(bash *) and Bash(sh *) lets a denied operation be wrapped as an allowed shell invocation, e.g. bash -c 'rm -rf .git' or sh -c 'git reset --hard'. That bypasses the deny rules below that are meant to preserve .git, source trees, and other protected paths, so the new settings are less restrictive than advertised; remove these blanket shell allowances or restrict them to reviewed scripts.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Block bare hard resets too

Because the allow list includes Bash(git *), this deny pattern only covers git reset --hard when there is another argument after --hard. A bare git reset --hard is valid and destructive (git reset -h describes --hard as resetting HEAD, index, and working tree), so sessions using this config can still discard all local changes despite the protection. Add an exact Bash(git reset --hard) deny entry, and consider the same exact-form coverage for the force-push denies.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Stop advertising missing knowledge files

The README tree tells users to open knowledge/reading-depth-ladder.md and knowledge/lie-detector.md, but this commit does not add those files; I checked .claude/EN/knowledge and only the autoattended and A2A docs exist. Users following the adoption guide will hit broken paths for two of the listed primitives, so either add the files or remove these entries and point to the sections that actually exist.

Useful? React with 👍 / 👎.