platform: promote 10 generic rules from custom/ for family-agent inheritance by fitz123 · Pull Request #125 · fitz123/claude-code-bot

fitz123 · 2026-05-17T21:24:59Z

Summary

Promotes 5 rules (down from 10 after owner review) from a private workspace .claude/rules/custom/ to upstream .claude/rules/platform/. Family agents inherit them via the .claude/rules/platform/ directory symlink.

What lands

File	State	Note
`anti-amnesia.md`	as-is	universal verify-before-claim
`bot-code-readonly.md`	expanded	paths frontmatter grew 4 -> 12 entries; added paragraph clarifying `protect-files.sh` hook only enforces `.claude/skills/**` for crons
`contradictions.md`	depersonalized	Ninja -> the user
`runtime-context.md`	parameterized	dropped workspace-specific paths and agent roster
`sops-no-decrypt-stdout.md`	generalized	dropped specific incident filenames

Plus CLAUDE.md: added @reference/governance/decisions.md to the existing @USER.md / @IDENTITY.md / @MEMORY.md import block. Replaces the previously promoted every-session.md rule — auto-import achieves the same effect without procedural instruction.

What does NOT promote (per owner review)

every-session.md — replaced by @reference/governance/decisions.md import above
implementation-protocol.md — main + coder only, non-code agents do not use ralphex
public-repo-pii.md — only main pushes to public repos
ralphex-notify-thread.md — ralphex-specific
show-evidence.md — main-only

These stay in main workspace custom/. Plus the original 8 stay custom rules: local-tools, reference-structure, release-flow, reminders, safety-workspace, task-tracking, fix-verification, heartbeats.

Dual-review iterations

Iter 1: family-angle reviewer found 2 Critical ("Ninja" hardcoded) + 3 High; main-angle clean.
Iter 2: family-angle found 2 High; addressed.
Iter 3: clean.
Iter 4 (owner review): 7 inline comments → 5 files dropped from promotion, bot-code-readonly expanded, CLAUDE.md import added.
Iter 5 (final): reviewer reports 1 advisory High (non-blocking @-import warning on fresh installs without decisions.md — owner explicitly requested the import).

Companion workspace PR

After this merges:

Delete now-duplicate custom files in workspace
Symlink reference/governance/decisions.md into anna/yulia/cyber-architect workspaces (so the new @-import from upstream CLAUDE.md resolves there)
Add reference/runbooks/create-family-workspace.md documenting required symlinks

Out-of-scope follow-up

Pre-existing platform rules with hardcoded "Ninja" (delegation, memory-protocol, communication, bot-operations). Separate small PR.

Test plan

Pre-commit hook (gitleaks + author-identity) passes locally — needed an allowlist update in fitz123/gitleaks-config for @reference / @memory literals first (merged separately as PR Telegram polling may silently stop after network failure (unconfirmed) #4).
CI passes on this PR.

…ritance Promotes these rules from private-workspace custom/ to upstream platform/ so all consumers (main agent + family agents Anna/Yulia/Coder/Cyber-Architect) inherit them via .claude/rules/platform/ directory symlink: As-is: - anti-amnesia, bot-code-readonly, contradictions, every-session, ralphex-notify-thread, show-evidence, sops-no-decrypt-stdout With edits (depersonalization + carve-outs for non-code agents): - implementation-protocol: "Main agent" -> "You"; carve-outs in both "Separation of Concerns" and "Ralphex - When to Use" sections so non-code agents skip cleanly - public-repo-pii: drop hardcoded fitz123/claude-code-bot reference - runtime-context: parameterize workspace paths, drop main-only agent roster (agents discover their roster via config.yaml) Stays in private workspace custom/: - local-tools, reference-structure, release-flow, reminders, safety-workspace, task-tracking, fix-verification, heartbeats (workspace-specific tooling, repo paths, or Ninja-only access) Dual-review iterations: - Iter 1: family-angle reviewer found 2 Critical (Ninja-hardcoded in show-evidence + contradictions) + 3 High; main-angle clean. Addressed. - Iter 2: family-angle reviewer found 2 High (implementation-protocol needed second carve-out; every-session needs decisions.md in family workspaces). One fixed in rule; the other deferred to companion workspace PR that symlinks decisions.md into family workspaces and documents the symlink set as a runbook. - Iter 3: family-angle reviewer confirms clean. Follow-up (separate PR): pre-existing platform rules with hardcoded "Ninja" (delegation, memory-protocol, communication, bot-operations) — out of scope here. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Copilot

Pull request overview

Promotes 10 generic rule files from a private workspace's custom/ directory into upstream .claude/rules/platform/ so all consumers (main agent + family agents) inherit them via the platform symlink. Seven files land as-is; three (implementation-protocol.md, public-repo-pii.md, runtime-context.md) are depersonalized/parameterized for family-agent reuse.

Changes:

Add 7 rules verbatim from the private workspace (anti-amnesia, bot-code-readonly, contradictions, every-session, ralphex-notify-thread, show-evidence, sops-no-decrypt-stdout)
Add 3 depersonalized rules with carve-outs for non-code agents and parameterized workspace paths
No code changes; all additions are markdown rule docs

Reviewed changes

Copilot reviewed 10 out of 10 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
`.claude/rules/platform/anti-amnesia.md`	New rule: verify before claiming, write to file.
`.claude/rules/platform/bot-code-readonly.md`	New rule: bot/platform files are upstream-managed, read-only in workspace.
`.claude/rules/platform/contradictions.md`	New rule: flag contradictions instead of silently resolving.
`.claude/rules/platform/every-session.md`	New rule: read USER.md + decision log at session start.
`.claude/rules/platform/implementation-protocol.md`	New rule with non-code-agent carve-outs in both sections.
`.claude/rules/platform/public-repo-pii.md`	New rule, generalized (no hardcoded repo).
`.claude/rules/platform/ralphex-notify-thread.md`	New rule: set RALPHEX_NOTIFY_THREAD when launching ralphex from non-default topic.
`.claude/rules/platform/runtime-context.md`	New rule: explains Claude Code subprocess runtime; parameterized paths, no agent roster.
`.claude/rules/platform/show-evidence.md`	New rule: assertions need inline evidence; includes PR self-check.
`.claude/rules/platform/sops-no-decrypt-stdout.md`	New rule: never decrypt sops to stdout in agent transcripts.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Owner left 7 inline comments. Resolutions: - every-session.md: deleted (retired entirely). Added @reference/governance/decisions.md to CLAUDE.md so the directive becomes an auto-import instead of a procedural rule. USER.md, IDENTITY.md, MEMORY.md were already imported. - implementation-protocol.md: deleted (main-only - non-code agents do not use ralphex). - public-repo-pii.md: deleted (main-only - only main pushes to public repos). - ralphex-notify-thread.md: deleted (main-only - ralphex-specific). - show-evidence.md: deleted (main-only). - bot-code-readonly.md: expanded paths frontmatter from 4 to 12 entries (.github/workflows/**, .githooks/**, .gitleaks.toml, .gitleaksignore, CLAUDE.md, README.md, config.yaml, config.local.yaml.example). Added paragraph noting protect-files.sh hook only enforces .claude/skills/** for crons - interactive sessions are self-discipline only. Final scope: 5 promoted rules + CLAUDE.md import. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Per owner request — the bot-code-readonly rule was soft self-discipline. Hook now hard-blocks writes to upstream-owned platform files from all sessions (interactive + cron), mirroring the rule paths list. Bypass triggers (all log to stderr for transcript visibility): 1. PROTECT_FILES_BYPASS=1 — explicit opt-out 2. CLAUDE_PROJECT_DIR contains /.ralphex/worktrees/ — ralphex pipeline 3. git remote.origin.url at CLAUDE_PROJECT_DIR matches *fitz123/claude-code-bot(.git)?(/)?$ — upstream dev checkout CLAUDE.md and config.yaml removed from the rule paths list (workspace edits these locally via merge=ours / per-workspace bindings). Dual-review iterations: - Iter 1: 2 Critical found (ralphex worktrees blocked, upstream dev blocked). Addressed via bypass logic. - Iter 2: 2 Critical in bypass (basename spoofable, PWD fallthrough) + 1 High (silent env bypass). All addressed: git-remote check instead of basename, fail-closed on missing CLAUDE_PROJECT_DIR, stderr logging. - Iter 3: clean. Test matrix (7 cases, all pass): T1 USER.md → allowed T2 workspace bot/src → BLOCKED T3 upstream dev (real remote) → BYPASS + log T4 spoofed /tmp/claude-code-bot (no remote) → BLOCKED T5 CLAUDE_PROJECT_DIR unset → BLOCKED (fail-closed) T6 ralphex worktree → BYPASS + log T7 PROTECT_FILES_BYPASS=1 → BYPASS + log Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Copilot

Pull request overview

Copilot reviewed 7 out of 7 changed files in this pull request and generated 1 comment.

Comments suppressed due to low confidence (1)

.claude/hooks/protect-files.sh:100

The README.md glob */README.md|README.md will match every README.md anywhere in the tree (e.g., reference/README.md, docs/README.md, .claude/skills/<x>/README.md), not just the repository-root README that the bot-code-readonly.md frontmatter entry README.md is meant to designate. Since reference/ and docs/ are explicitly listed as workspace-local/editable in the rule (lines 23–27), this hook will block legitimate edits to nested READMEs. Consider anchoring the match to the project root (e.g., compare against $PROJECT_ROOT/README.md after realpath normalization) rather than a trailing-component glob. The same concern applies to */config.local.yaml.example|config.local.yaml.example.

  */README.md|README.md) match=1 ;;
  */config.local.yaml.example|config.local.yaml.example) match=1 ;;

Copilot flagged that case patterns like */bot/*|bot/* match the `bot/` segment at ANY depth — false-blocking nested paths like `reference/bot/notes.md` or `docs/bot/x.md`. The frontmatter list in bot-code-readonly.md is rooted (bot/**, .github/workflows/**, etc) so the hook must match the same way. Compute REL_PATH = FILE_PATH minus $CLAUDE_PROJECT_DIR prefix once, then match against rooted patterns (`bot/*`, `.github/workflows/*`, ...). Same anchoring applied to the existing skills block. Tests (8 cases, all pass): T1 USER.md → allowed T2 /workspace/bot/src/foo.ts → BLOCKED T3 /workspace/reference/bot/notes.md → ALLOWED (was wrongly blocked) T4 /workspace/docs/bot/x.md → ALLOWED (was wrongly blocked) T5 /workspace/reference/.github/workflows/x.yml → ALLOWED T6 .github/workflows/x.yml at root → BLOCKED T7 CLAUDE.md → allowed T8 upstream dev repo bypass → BYPASS + log Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

fitz123 · 2026-05-17T22:32:09Z

@copilot please re-review — fixes pushed since the last review

Agent-Logs-Url: https://github.com/fitz123/claude-code-bot/sessions/c0947a7f-0ad9-4031-ba90-7eb97358feb9 Co-authored-by: fitz123 <10243861+fitz123@users.noreply.github.com>

Copilot · 2026-05-17T22:35:18Z

@copilot please re-review — fixes pushed since the last review

Re-reviewed the latest updates and added one test-only follow-up to keep hook tests deterministic with rooted path matching. Commit: ad63baf.

Copilot AI review requested due to automatic review settings May 17, 2026 21:25

Copilot started reviewing on behalf of fitz123 May 17, 2026 21:25 View session

Copilot AI reviewed May 17, 2026

View reviewed changes

fitz123 commented May 17, 2026

View reviewed changes

fitz123 and others added 2 commits May 18, 2026 00:42

Copilot AI review requested due to automatic review settings May 17, 2026 22:12

Copilot started reviewing on behalf of fitz123 May 17, 2026 22:12 View session

Copilot AI reviewed May 17, 2026

View reviewed changes

Comment thread .claude/hooks/protect-files.sh Outdated

Copilot started work on behalf of fitz123 May 17, 2026 22:32 View session

test: stabilize protect-files hook tests with explicit project root env

ad63baf

Agent-Logs-Url: https://github.com/fitz123/claude-code-bot/sessions/c0947a7f-0ad9-4031-ba90-7eb97358feb9 Co-authored-by: fitz123 <10243861+fitz123@users.noreply.github.com>

Copilot AI review requested due to automatic review settings May 17, 2026 22:34

fitz123 review requested due to automatic review settings May 17, 2026 22:34

Copilot finished work on behalf of fitz123 May 17, 2026 22:36

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

platform: promote 10 generic rules from custom/ for family-agent inheritance#125

platform: promote 10 generic rules from custom/ for family-agent inheritance#125
fitz123 wants to merge 5 commits into
mainfrom
platform/promote-generic-custom-rules

fitz123 commented May 17, 2026 •

edited

Loading

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

fitz123 commented May 17, 2026

Uh oh!

Copilot AI commented May 17, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

fitz123 commented May 17, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

What lands

What does NOT promote (per owner review)

Dual-review iterations

Companion workspace PR

Out-of-scope follow-up

Test plan

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Uh oh!

fitz123 commented May 17, 2026

Uh oh!

Copilot AI commented May 17, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

fitz123 commented May 17, 2026 •

edited

Loading