audit: bead/mcb-o96i-16-hook-commit-recovery-v3 -> main

.code-review-graphignore

@@ -0,0 +1,31 @@
+# BEGIN AI-HUB CRG GLOBAL POLICY
+build/**
+.worktrees/**
+worktrees/**
+.agents/worktrees/**
+.code-review-graph/wiki/**
+*.db
+*.db-shm
+*.db-wal
+*.sqlite
+*.sqlite3
+*.min.js
+*.min.css
+coverage/**
+.coverage/**
+.pytest_cache/**
+.ruff_cache/**
+.mypy_cache/**
+.tox/**
+.cache/**
+tmp/**
+temp/**
+.claude/worktrees/**
+.codex/worktrees/**
+.gemini/worktrees/**
+.kimi-code/worktrees/**
+.cursor/worktrees/**
+.config/worktrees/**
+.vscode/worktrees/**
+.aider/worktrees/**
+# END AI-HUB CRG GLOBAL POLICY

.mise.toml

@@ -0,0 +1,28 @@
+# Generated by `flext_infra codegen conform`.
+# === SECTION: header (managed) ===
+# Source: template (base/.mise.toml.j2)
+# Free: no
+# End SECTION: header
+# NOTE: mise carries the Python runtime and native tools from the config SSOT.
+# Python dependencies and uv are resolved by the caller-provided environment.
+
+# === SECTION: tools (managed) ===
+# Source: config:python_version, config:codegen.scaffold.project.dev native toolchain floors
+#         (kubectl_version, helm_version, kind_version, taplo_version, ast_grep_version,
+#          gitleaks_version, tokei_version, beads.selector, beads.version)
+[tools]
+python = "3.13"
+# Native toolchain (single source: Infra.codegen.toolchain in codegen.yaml).
+kubectl = "1.32.0"
+helm = "3.19.4"
+kind = "0.31.0"
+taplo = "0.10.0"
+ast-grep = "0.45.0"
+gitleaks = "8.30.1"
+tokei = "14.0.0"
+# mise resolves go: backend selectors through a dependency graph, so the Go
+# runtime must be a declared tool; without it mise looks for go on PATH and
+# every runner without an ambient toolchain fails to install beads.
+go = "1.26.5"
+"go:github.com/steveyegge/beads/cmd/bd" = "423afdcb2813e36b2bc4c96b07e0fc3516a34495"
+# End SECTION: tools

AGENTS.md

@@ -1,138 +1,11 @@
-> Universal rules: `~/.agents/AGENTS.md` loads first. This file is the
-> project-canonical overlay for `/home/marlonsc/mcb`.
-
-# AGENTS.md - MCB Project Rules
-
-<!-- BEGIN UNIVERSAL AGENT LAW (portable; regenerable; do not edit inside) -->
-## Universal Agent Law (portable core)
-
-**This block is the inviolable, agent-agnostic core of engineering conduct for this repository.** It is
-self-contained: it binds any AI agent — Claude, Codex, Gemini, Cursor, Cline, GitHub Copilot, or any other —
-and any user, with or without access to the author's personal configuration. The live user's explicit
-instructions override this block; nothing else does. These rules apply to every project type and every
-session, and may not be relaxed, reinterpreted, or scoped-out for convenience, speed, or perceived triviality.
-
-### 1. Zero-Tolerance / Strict-Total
-- **Always** fix the root cause — generically, cleanly, via reuse of existing canonical code — and validate it
-  in the same turn with the actual command, its exit code, and the relevant output line.
-- **Always** remove superseded code in the same cycle the replacement lands. No dead code "for later".
-- **Always** fail loud when the single source of truth (identity, config, contract, version) is absent — never
-  substitute a guess, a local copy, or an alternative path.
-- **Never** use a fallback, compatibility wrapper, legacy branch, allowlist/carve-out, skip, suppression,
-  hardcode, stub, fake, `TODO`/`FIXME`, or a side-script to make a gate pass.
-- **Never** classify a failure surfaced by the current task as "pre-existing", "cosmetic", "unrelated", or
-  "acceptable legacy". If it appears in your flow, you own it.
-
-### 2. Fix-Forward-Only
-Multiple agents may share one working tree. Reverting to a past state silently destroys another agent's
-in-flight work. **Accept the current state and fix forward.** Discarding changes via `git checkout -- <path>`,
-`git restore`, `git reset --hard`, `git reset <path>`, `git stash` (hiding others' work), `git clean`, or
-`git revert` of another's commit is **forbidden**. If you think you must revert → **STOP and ask the user**;
-never unilaterally revert shared work.
-
-### 3. Root Cause Only — No Workarounds
-No TODOs, stubs, fakes, fallbacks, compat wrappers, or "temporary" workarounds. No suppression directives
-(`# type: ignore`, blanket `# noqa`, `@ts-ignore`, `eslint-disable`, etc.) and no escape-hatch typing
-(`Any`, bare `object`, unchecked casts) unless carrying a one-line documented justification. A bypass that
-hides a symptom is a defect even when the gate turns green.
-
-### 4. Stay In Scope
-Do exactly what the user asked — nothing more. No unrequested refactors, renames, cleanups, "obvious
-improvements", or adjacent fixes. Found something unrelated? Mention it in one sentence; do not touch it.
-
-### 5. Evidence Before Done — Report Honesty Is 100% Mandatory
-"Done" means the **complete chain validated** with objective evidence (command + exit code + output), not
-conclusion-by-sample. **Never** present partial, assumed, speculative, or unverified results as verified.
-State explicitly when a step was skipped, when a check failed (paste the output), and when a result is
-unverified. If something only worked via a workaround, say so — it is not "done".
-
-### 6. Execute As Planned, Else Stop And Ask
-Execute the agreed plan exactly. On anything that cannot be done cleanly — a blocked tool, a missing source of
-truth, a real ambiguity, or a step that would require a bad practice — **STOP and ask**, presenting concrete
-options. **Every option must be a clean, root-cause solution.** Fallback, hack, hardcode, suppression, skip,
-or stub are **forbidden as suggestions** — never offer one, even labelled "quick" or "temporary". Any
-mid-execution deviation from the plan requires explicit user confirmation **before** applying.
-
-### 7. Blocked-Operation Protocol
-When a tool, command, or edit is blocked (deny rule, security hook, sandbox, missing permission, unavailable
-integration): (1) **Stop** — do not retry a variation or seek a bypass; (2) **diagnose in one sentence** what
-was blocked and why; (3) **hand the exact command or edit to the user** to run on their side; (4) **wait for
-their output** before continuing; (5) **never claim done because a substitute ran** — a successful bypass is
-still a violation. Forbidden bypass techniques include `bash -c`/`sh -c` subshell wrapping, `eval`/`exec`,
-`env <blocked>`, `xargs <blocked>`, absolute-path swaps to dodge prefix deny rules, pipes/command-chains into a
-blocked command, and invoking it via a `subprocess` call.
-
-### 8. Strict, Most-Restrictive Typing
-Use the most restrictive type that compiles. No `Any`, no bare `object`, no suppression of type errors. Fix
-types at the source; depend on declared contracts, not loosely-typed escape hatches.
-
-### 9. Universal Engineering Principles (always, no exception)
-- **SSOT** — one authoritative source per fact; reference it, never duplicate or restate it; fail loud when
-  absent.
-- **SOLID** — SRP / OCP / LSP / ISP / DIP respected. Type-switching where polymorphism applies, fat
-  interfaces, and god-objects are defects.
-- **YAGNI** — no speculative params, dead branches, future-hooks, or single-implementation abstractions.
-  Build only what the task needs now; delete the rest.
-- **DI / DIP** — depend on abstractions (protocols/interfaces); inject collaborators; no hidden globals or
-  hard-wired construction inside business logic.
-
-### 10. User Manages Git
-Do not run `git add`/`commit`/`push`/`tag` unless the user explicitly requests it, and do not suggest
-committing. Read-only inspection (`status`/`log`/`diff`) is fine. When a commit is authorized, write it as the
-user with no agent/bot attribution — no `Co-Authored-By`, no "Generated with …" trailer, and never override
-author/committer identity.
-
-### 11. Multi-Agent Coordination
-Agents may share one working tree. Coordinate through a committed task board (e.g.
-`<repo>/.agents/coordination/tasks.md`): claim a task with an ownership + lease entry before editing, heartbeat
-the lease, set `done`/`blocked` on finish, and recover stale tasks from git history. Commit small and often so
-a fresh agent rebuilds state from `git log`. **Never overwrite or discard another agent's work** (see Rule 2);
-on a divergent approach, stop and escalate to the user.
-
-### 12. When Unsure — Ask
-If a task is unclear, ambiguous, or would expand scope → ask one focused question. If an action is hard to
-reverse, affects shared state, or could surprise the user → confirm first. Authorization is scope-specific:
-approval for one action once does not authorize it in future contexts.
-
-### 13. Destructive Commands — Archive, Don't Destroy
-Prefer non-destructive moves: archive a file as `<file>.bak` instead of deleting it. Do not escalate
-privileges (`sudo`/`su`), change ownership/permissions, perform remote operations, or fetch over the network
-without explicit user confirmation. Use the agent's structured file/search/edit tools over raw destructive
-shell commands.
-<!-- END UNIVERSAL AGENT LAW -->
-
-MCB (Memory Context Browser) is a Rust 2024 MCP server for persistent agent
-memory, semantic code search, and architecture validation.
-
-## Current Status
-
-- Source version: `0.3.1` from `Cargo.toml`.
-- Active branch observed during init: `release/v0.3.1`.
-- Rust toolchain: stable, MSRV `1.92`, edition `2024`.
-- Workspace: 7 first-party crates; `third-party/` is excluded from the
-  workspace and should not be edited unless the user explicitly asks.
-- Platform state: the v0.3 SeaQL + Loco.rs rebuild is the current baseline.
-- Public MCP surface: 24 tool names registered through `linkme` descriptors,
-  grouped into 9 handler families in `docs/MCP_TOOLS.md`.
-
-When a static document disagrees with `Cargo.toml`, `Makefile`, `make/*.mk`,
-`config/*.yaml`, or the code, trust the executable source first and update the
-doc as part of the same change.
-
-## Source Of Truth
-
-- Version, MSRV, workspace members, lint policy: `Cargo.toml`.
-- Rust toolchain components and targets: `rust-toolchain.toml`.
-- Developer commands: `Makefile` plus `makefiles/ui.mk`, `makefiles/dispatch.mk`,
-  and the canonical monopoly script `scripts/lib/mcb.sh` (exit codes, the
-  `APPLY=Y` gate, SSOT readers, the banned-pattern guard, the agent bash-guard).
-- Runtime configuration: `config/development.yaml`, `config/test.yaml`,
-  `config/production.yaml`.
-- Architecture validation config: `config/mcb-validate.toml` and
-  `config/mcb-validate-internal.toml`.
-- MCP tool contract: `docs/MCP_TOOLS.md` and `crates/mcb-server/src/args/`.
-- Architecture rules and ADR context: `docs/architecture/` and `docs/adr/`.
+# AGENTS.md — mcb
 
+> **Parent workspace law** lives in [`../AGENTS.md`](../AGENTS.md) — read it first.
+> Universal engineering core: `~/.agents/UNIVERSAL_CORE.md`. Composition: global skills + parent/root `AGENTS.md` + this scope delta. Do not re-embed universal law.
+>
+> **Standalone / independent mode:** when `../AGENTS.md` does not resolve, pin the parent raw `AGENTS.md` URL to the same branch/release as this package (never `main`).
+
+<!-- AIHUB-AGENTS-SCOPE-LOCAL-BEGIN -->
 ## Commands
 
 The whole dev cycle runs through few canonical `make` verbs backed by the single
@@ -307,3 +180,4 @@ passed unless that exact gate was run in the current turn.
   on static status.
 - `docs/architecture/ARCHITECTURE.md`: architecture overview and historical
   context.
+<!-- AIHUB-AGENTS-SCOPE-LOCAL-END -->