Context Manifest: Validate orchestrator-discipline plugin

Context Manifest: Validate orchestrator-discipline plugin

Groomed: 2026-02-20 Source backlog item: .claude/BACKLOG.md lines 23-28

NOTE (2026-02-27): BACKLOG.md was removed. Backlog items now live in .claude/backlog/ per-item files; GitHub Issues are the source of truth.

Plugin Structure

Location: plugins/orchestrator-discipline/

plugins/orchestrator-discipline/
├── .claude-plugin/
│   └── plugin.json                          (670 bytes)
├── hooks.json                               (597 bytes — at ROOT, not in hooks/)
├── hooks/
│   ├── pre-tool-orchestrator-read-warning.cjs    (2648 bytes, chmod +x)
│   └── pre-tool-diagnostic-command-gate.cjs      (3302 bytes, chmod +x)
├── rules/
│   └── CLAUDE.md                            (6615 bytes)
└── skills/
    └── orchestrator-discipline/
        ├── SKILL.md                         (4082 bytes)
        └── references/
            └── investigation-escalation.md  (13264 bytes)

Total files: 7 (including plugin.json). No agents directory. No scripts directory.

ANOMALY: hooks.json exists at the plugin root AND the plugin references ./hooks.json from plugin.json. A hooks/ directory contains the JS files, but there is no hooks/hooks.json. This is the correct layout — plugin.json hooks field points to ./hooks.json at root, and that file's command values reference ${CLAUDE_PLUGIN_ROOT}/hooks/*.js.

Hook Configuration

File: plugins/orchestrator-discipline/hooks.json (at plugin root)

{
  "description": "Orchestrator discipline enforcement hooks — context window protection",
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Read|Grep",
        "hooks": [{ "type": "command", "command": "node \"${CLAUDE_PLUGIN_ROOT}/hooks/pre-tool-orchestrator-read-warning.cjs\"" }]
      },
      {
        "matcher": "Bash",
        "hooks": [{ "type": "command", "command": "node \"${CLAUDE_PLUGIN_ROOT}/hooks/pre-tool-diagnostic-command-gate.cjs\"" }]
      }
    ]
  }
}

Event: PreToolUse Matchers: Read|Grep (source file read warning) and Bash (diagnostic command gate) Both hooks: type: command, invoke Node.js scripts via ${CLAUDE_PLUGIN_ROOT} variable

Hook Logic Summary

pre-tool-orchestrator-read-warning.cjs

Trigger condition: Read or Grep tool call where target path matches:

• Extensions: .py .toml .yaml .yml .js .ts .jsx .tsx .json .cfg .ini .env .sh .bash .go .rs .rb .java .c .cpp .h .hpp
• Test directories: paths containing /test/, /tests/, /spec/, /__tests__/, or files matching test_*.py

Does NOT fire on: .md, .txt, plan files, BACKLOG.md, CLAUDE.md, skill definitions (confirmed by test)

Action: Non-blocking. Injects additionalContext via hookSpecificOutput.additionalContext with an <orchestrator-read-warning> XML block asking: "Will you Edit or Write this file this turn?"

Verified behavior (live test 2026-02-20):

• Read on BACKLOG.md → {} (no output, does not fire)
• Read on .js file → fires, additionalContext present
• Read on .py file → fires
• Read on .toml file → fires
• Grep on .js file → fires
• Grep on directory path (no extension) → {} (does not fire)

Gap identified: Grep with a directory path argument (e.g., Grep pattern /path/to/src) does not fire because the directory path has no extension. This is a false negative — the hook only checks the path field of Grep, which may be a directory. Broad Grep searches across source directories are not caught.

pre-tool-diagnostic-command-gate.cjs

Trigger condition: Bash tool call where command matches any of:

• Python: ty check, ruff check, mypy, pyright, basedpyright, pylint, pytest
• JS/TS: eslint, tsc.*--noEmit
• Rust: cargo check, cargo clippy
• Go: go vet
• Meta: pre-commit run, prek run

Does NOT fire on: git status, ls, wc, uv run alone, any non-diagnostic Bash (confirmed by test)

Action: Non-blocking. Injects additionalContext with <orchestrator-diagnostic-warning> block showing the matched command and instructing delegation.

Verified behavior (live test 2026-02-20):

• uv run ty check . → fires
• uv run pytest tests/ → fires
• uv run prek run --files foo.py → fires
• git status → does not fire

Gap identified: ruff format is not in DIAGNOSTIC_PATTERNS but produces large output. Whether it should be gated is a design question. uv run ruff check IS caught (because ruff check is in patterns). mypy with uv run prefix is also caught.

Rules Content Summary

File: plugins/orchestrator-discipline/rules/CLAUDE.md (6615 bytes)

Sections:

1. Context Window Read Constraints — Explicit permitted/prohibited lists. Falsifiable test: "Will I Edit or Write this file in this turn?"
2. Delegation Constraints — No exemption categories. Never pre-gather data for agents. Never pre-read task files.
3. Investigation Escalation Anti-Pattern — 4-step escalation sequence with mermaid flowchart. Trigger signal: 3+ Read/Grep/Bash on source files without interleaving Edit/Write/Task.
4. Agent Output Polling Anti-Pattern — TaskOutput with block=false on running agents. Prohibition on reading .output files directly. Session 77509a5e cited as observed instance.
5. Diagnostic Commands — Lists commands that must be delegated. Exception for post-edit single-file verification.
6. Epistemic Identity Scope — Scopes the "use tools to verify" imperative: task-routing information only, not source code.

Format: Mermaid flowcharts for both anti-patterns. Hard constraints use NEVER/MUST NOT language.

Skill Summary

File: plugins/orchestrator-discipline/skills/orchestrator-discipline/SKILL.md (4082 bytes)

Frontmatter:

---
description: Orchestrator context window discipline enforcement. Prevents the orchestrator from reading source files it will not edit, running diagnostic commands that waste context, and rationalizing delegation bypasses. Use when setting up orchestrator guardrails, reviewing delegation discipline, or diagnosing context window waste in multi-agent workflows. Activates PreToolUse hooks that surface decision points before source file reads and diagnostic command execution.
---

No name field (correct — avoids Claude Code bug where name field prevents slash command registration). No user-invocable field (skill is auto-loaded via plugin hooks/rules, not invoked manually).

Contents:

• Explains what the plugin provides (hooks, rules, reference material)
• Hook behavior reference — extensions list, command patterns
• "Does NOT trigger on" section for both hooks
• Correct orchestrator workflow mermaid diagram

Trigger phrases in description: "orchestrator guardrails", "delegation discipline", "context window waste", "multi-agent workflows" — usable for semantic activation.

Reference Files

plugins/orchestrator-discipline/skills/orchestrator-discipline/references/investigation-escalation.md (13264 bytes)

Contents:

• Pattern description with 4-step escalation sequence
• Root cause analysis (competing instructions, soft phrasing, no structural enforcement, exemption invention)
• Detection signals flowchart (mermaid)
• Correct workflow examples (diagnostic baselining, code investigation, config changes)
• Anti-pattern examples with token cost estimates (~21,000 chars vs ~500 chars)
• Variant: Agent Output Polling (identical root cause, different surface form)

Reference is linked from SKILL.md: [Investigation Escalation Anti-Pattern](./references/investigation-escalation.md)

Plugin Manifest Completeness

File: plugins/orchestrator-discipline/.claude-plugin/plugin.json

{
  "name": "orchestrator-discipline",
  "description": "Enforces orchestrator context window discipline via PreToolUse hooks and rules. ...",
  "version": "1.3.2",
  "author": { "name": "Jamie Nelson", "url": "https://github.com/bitflight-devops" },
  "skills": ["./skills/orchestrator-discipline"],
  "hooks": "./hooks.json",
  "rules": ["./rules"],
  "commands": ["./skills/orchestrator-discipline"]
}

Declared components:

• skills: points to ./skills/orchestrator-discipline (directory exists, SKILL.md present)
• hooks: points to ./hooks.json (file exists at root)
• rules: points to ["./rules"] (directory exists, CLAUDE.md present)
• commands: points to ["./skills/orchestrator-discipline"] (duplicates skills — appears intentional to register as slash command)

CRITICAL ISSUE — validation failure: "rules" is an unrecognized key per claude plugin validate. No other plugin in the repo uses rules in plugin.json. This is a blocking issue for marketplace distribution.

Marketplace entry: Present in .claude-plugin/marketplace.json as:

{ "name": "orchestrator-discipline", "source": "./plugins/orchestrator-discipline" }

Note: Entry lacks id field. Other entries in marketplace.json also lack id — this appears to be the repo's standard format.

Validation Results

Command run: claude plugin validate plugins/orchestrator-discipline/

Validating plugin manifest: /home/ubuntulinuxqa2/repos/claude_skills/plugins/orchestrator-discipline/.claude-plugin/plugin.json

✘ Found 1 error:

  ❯ root: Unrecognized key: "rules"

✘ Validation failed

Root cause: The "rules" field in plugin.json is not part of the Claude Code plugin.json schema. It is a plugin-specific convention in this repo but not a recognized plugin manifest key.

Impact: Plugin cannot be validated with claude plugin validate. Marketplace users who run validation will get a failure. Cannot be submitted to marketplace as-is.

Fix options:

1. Remove "rules" from plugin.json — rules may still load if Claude Code loads rules/ by convention (unverified)
2. Move rules content into SKILL.md or a CLAUDE.md at a different level
3. Investigate whether Claude Code loads a plugin's rules/ directory automatically without it being declared

Privacy Scan Results

Email search: No matches found in any plugin file.

Personal names found:

• plugin.json line 6: "name": "Jamie Nelson" — in the author field. This is the repo owner's name and appears in other plugin manifests. Not a privacy issue for this repo context.
• plugin.json line 7: "url": "https://github.com/bitflight-devops" — public GitHub org URL. Not private.

API keys, tokens, passwords: No matches found.

Session ID in rules/CLAUDE.md line 78: 77509a5e — references a session ID from the source analysis session. This is an internal reference, not a credential. Appears acceptable.

Conclusion: No PII or private data issues beyond the author attribution, which is standard for this repo.

Related Skills and Agents

Skills directly relevant to this work:

• plugins/plugin-creator/skills/hooks-core-reference/SKILL.md — hook event types, lifecycle
• plugins/plugin-creator/skills/hooks-io-api/SKILL.md — hook input/output schema, hookSpecificOutput structure
• plugins/plugin-creator/skills/hooks-patterns/SKILL.md — hook design patterns
• plugins/plugin-creator/skills/claude-hooks-reference-2026/SKILL.md — comprehensive hooks reference
• plugins/plugin-creator/skills/claude-plugins-reference-2026/SKILL.md — plugin.json schema (authoritative source for what fields are valid)

For validation testing:

• plugins/plugin-creator/scripts/plugin_validator.py — internal validator (separate from claude plugin validate)

Agents relevant to this work:

• @plugin-creator:plugin-assessor — comprehensive plugin structure analysis
• @plugin-creator:refactor-validator — validates refactoring completeness

Prior Work

Commits:

03f4342  feat(orchestrator-discipline): add plugin for orchestrator context window discipline
         (2026-02-19, initial creation — added all 7 files)

1cfefd4  refactor(mypy): update mypy configuration for improved type checking
         (touches orchestrator-discipline? — adjacent commit, may be unrelated)

323e23e  feat(dasel): add data transformation and exploration plugin
         (unrelated)

908f73e  style(frontmatter): normalize YAML quoting across all component files
         (touches orchestrator-discipline files — YAML quoting normalization)

The 908f73e commit updated SKILL.md frontmatter quoting. The initial commit 03f4342 created all files. No other commits modify this plugin.

Commit message (03f4342) notes:

• "Updates marketplace.json (version 2.13.0) and BACKLOG.md with plugin validation task"
• Confirms the validation task was added to BACKLOG at time of plugin creation

BACKLOG.md: Backlog item at lines 23-28 is this exact item — correctly describes what needs to be done.

Session artifacts: The investigation-escalation.md reference file mentions session 77509a5e (dasel plugin creation session 2026-02-19) as the observed instance of the agent output polling anti-pattern. This is a narrative reference, not a file artifact.

Planning docs: No separate planning documents found for this plugin. Plugin was created directly from anti-pattern analysis.

Gaps and Blockers

BLOCKER — validation failure: claude plugin validate fails with: root: Unrecognized key: "rules". This is the primary issue to resolve before the plugin can be considered production-ready or marketplace-distributable.

Investigation needed:

• Does Claude Code load a plugin's rules/ directory automatically (by convention) even without a rules key in plugin.json?
• If yes: Remove rules from plugin.json and verify rules/CLAUDE.md still loads in sessions.
• If no: Find the correct mechanism to ship rules with a plugin — possibly by moving rules content into the SKILL.md or using a different field.

Hook gap — Grep on directory paths: When the orchestrator calls Grep with a directory as the path argument (no file extension), pre-tool-orchestrator-read-warning.cjs does not fire. The hook checks if toolInput.path matches SOURCE_FILE_EXTENSIONS, but directories have no extension. This is a false negative for broad codebase searches.

Unverified — rules loading behavior: Whether rules/CLAUDE.md actually loads into the session context when the plugin is installed has not been tested. The hook behavior was verified (live tests). The rules loading depends on Claude Code's plugin rules mechanism, which is separate from the hook mechanism.

Unverified — slash command registration: The SKILL.md has no user-invocable: true field. Without this, the skill may not register as a /orchestrator-discipline:orchestrator-discipline slash command. The SKILL.md description mentions "Manual activation is useful when..." — implying it should be manually invocable. Whether it actually appears as a slash command is not confirmed.

hooks.json location note: hooks.json is at the plugin root, not inside hooks/. This is intentional (plugin.json "hooks": "./hooks.json" references it correctly). No confusion for implementation, but worth noting for reviewers.

Suggested First Steps

Ordered by priority:

1. Resolve the rules key validation failure (BLOCKER)

   • Read plugins/plugin-creator/skills/claude-plugins-reference-2026/SKILL.md to find the canonical plugin.json schema
   • Test whether rules load without being declared: install plugin, open session, check if rules/CLAUDE.md content appears in system context
   • If auto-loaded: remove "rules" from plugin.json and re-run claude plugin validate — expect clean pass
   • If not auto-loaded: research correct declaration method or move rules content into SKILL.md

2. Verify rules load into session context

   • Install plugin locally: claude plugin install orchestrator-discipline@jamie-bitflight-skills --scope local
   • Open a Claude Code session
   • Confirm rules/CLAUDE.md content is active (ask Claude to describe its orchestrator read constraints)
   • OR test with claude --plugin-dir ./plugins/orchestrator-discipline session

3. Verify hooks fire in real Claude Code session

   • Session test with plugin installed: attempt to Read a .py file and confirm <orchestrator-read-warning> appears in context
   • Session test: attempt uv run pytest and confirm <orchestrator-diagnostic-warning> appears
   • Session test: Read on BACKLOG.md and confirm no warning fires

4. Evaluate Grep-on-directory gap

   • Assess whether this is a meaningful gap: orchestrators searching broad directories are the high-risk behavior
   • If yes: change hook to fire when tool is Grep regardless of path extension (fire on all Grep calls, or check if path is a directory)
   • Document design decision

5. Add user-invocable: true to SKILL.md if manual activation is intended

   • SKILL.md says "Manual activation is useful when..." but has no user-invocable field
   • If slash command registration is desired, add user-invocable: true
   • Run uv run prek run --files plugins/orchestrator-discipline/skills/orchestrator-discipline/SKILL.md after

6. Verify slash command appears after installation

   • With plugin installed, confirm /orchestrator-discipline:orchestrator-discipline appears in autocomplete
   • If not: check for name field interference (plugin-creator docs note this is a known bug)

7. Lint all plugin files

   • Run uv run prek run --files plugins/orchestrator-discipline/rules/CLAUDE.md
   • Run uv run prek run --files plugins/orchestrator-discipline/skills/orchestrator-discipline/SKILL.md
   • Run uv run prek run --files plugins/orchestrator-discipline/skills/orchestrator-discipline/references/investigation-escalation.md