fix(development-productivity): Opus 5 fixes for commands, skills, and test-writer

claude-init-plus:
- added an explicit instruction against one-subagent-per-package fan-out. Its
  Hierarchy Deduplication rule requires knowing what was written at every other
  level, and a subagent cannot see what a sibling wrote -- so the fan-out breaks
  it silently, with every individual file still looking correct.
- replaced the trailing 4-item verify checklist (and the block restating the
  same items as failure conditions) with the two conditions that are actually
  checkable after the fact.

update-claude-md:
- fixed a dangling sentence fragment at line 16 left by a bad edit.
- dropped the Step 10 post-write re-read; deduplication and the no-TODO rule are
  constraints on the write, and the 200-line case is already in Error Handling.

generate-tests:
- gen-tests-guide required coverage: number, but its allowed-tools contain no
  Bash, so it cannot run a coverage tool. Now coverage | 'not_measured' with a
  coverageSource field.
- SKILL.md claimed "Write test files with full coverage" with no Write tool and
  no Bash; now returns artifacts and names untested behaviors.
- context-loader/security-analyzer/performance-analyzer/agent-orchestrator all
  live in development-codebase-tools; qualified every reference with the
  cross-plugin prefix, including in allowed-tools where a bare name would not
  resolve.
- added a when-NOT-to-delegate note; the agent list is a menu, not a sequence.

optimize-prompt: "GPT-4 best practices" replaced with provider-neutral guidance
plus the concrete Anthropic parameter constraints.

test-writer: 400-word cap on prose fields (test bodies and enumerated lists are
data, not covered); added a coverage instruction for untested behaviors.

CLAUDE.md: documented the inline-verification posture, the measured-inputs rule,
and the sequential-generation constraint.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Author: wkoutre

packages/plugins/development-productivity/CLAUDE.md

@@ -51,6 +51,16 @@ disclosure best practices for CLAUDE.md content:
 - Skills are the primary interface; agents are invoked via `Task(subagent_type:agent-name)`
 - `claude-init-plus` runs once (bootstraps); `update-claude-md` runs on each significant change
 - The `documentation-agent` consolidates doc-writer, claude-docs-manager, and fact-checker agents
+- **Documentation verification is inline, never a second-agent pass.** Both
+  `documentation-agent` and `claude-docs-initializer-agent` verify claims against the repository
+  as they write, and flag anything they could not confirm on an `unverified_claims` list. The
+  initializer no longer emits `requires_verification: true`, and callers must not invoke
+  `documentation-agent` as a fact-checker over another agent's output.
+- **Numbers in agent output require measured inputs.** No self-assigned quality scores, no
+  coverage percentages against an unenumerable whole, no predicted improvement percentages for
+  something that was never run. Report qualitatively, or emit an explicit `not_measured`.
+- `/claude-init-plus` generation is sequential — never one subagent per package. A subagent
+  cannot see what a sibling wrote, which silently breaks Hierarchy Deduplication.
 - `generate-tests` supports: jest, vitest, pytest, cypress, playwright
 
 ## File Structure

packages/plugins/development-productivity/agents/test-writer.md

@@ -25,7 +25,12 @@ You are **test-writer-agent**, a specialized testing subagent with advanced test
 
 **Output**
 
-- `summary`: comprehensive testing strategy and rationale
+Prose fields (`summary`, `rationale[]`, `recommendations`) are capped at **400 words
+combined**. Test file bodies and enumerated scenario/edge-case lists are data, not prose, and
+are not covered by that cap. Cover the substance; cut restatement and any summary that repeats
+what the test names already say.
+
+- `summary`: testing strategy and rationale
 - `suggestedTests[]`:
   - `file`: destination test path
   - `contents`: complete test file body
@@ -285,6 +290,8 @@ Each test recommendation includes:
 
 ## Quality Assurance
 
+- Report every behavior you identified but did not write a test for, and why. Do not drop the
+  uncertain or low-priority ones — an untested behavior nobody names is one nobody covers.
 - All generated tests must be deterministic and repeatable
 - Tests should fail for the right reasons (not flaky)
 - Mock external dependencies appropriately

packages/plugins/development-productivity/commands/claude-init-plus.md

@@ -192,6 +192,20 @@ already exists in any ancestor CLAUDE.md. If it does, omit it from the subdirect
 When a convention applies only within a specific subtree, put it in the most specific
 CLAUDE.md that covers all relevant code, not in the root.
 
+### Do not fan out one subagent per package
+
+**Write every CLAUDE.md in this session, sequentially.** Do not dispatch a subagent per core
+node, per package, or per language group.
+
+The reason is the deduplication rule directly above: it requires knowing what has already been
+written at every other level. A subagent cannot see what a sibling wrote — each one starts with
+its own context, decides in isolation that a repo-wide convention is worth stating, and writes
+it. The result is the exact duplication this section exists to prevent, and nothing in the run
+reports a failure, because every individual file looks correct on its own.
+
+Discovery (Phase 1) is read-only and has no such constraint; parallelizing a large scan is fine.
+The generation phase is not.
+
 ## `.claude/rules/` Scaffolding
 
 When initializing a project that has clear cross-cutting concerns, offer to create
@@ -248,21 +262,17 @@ content into `.claude/rules/<topic>.md` and reference it from the CLAUDE.md inst
 1. Sections marked: `<!-- AUTO-GENERATED - DO NOT EDIT -->`
 2. Content that exactly matches a generated template with no user modification
 
-## Success Criteria and Verification
-
-**After generation, verify (required):**
+## Failure Conditions
 
-1. **File validity**: File created/updated, not empty, has required sections
-2. **Content quality**: No `[TODO]` placeholders, conventions are specific and actionable
-3. **Length**: Each file is under 200 lines
-4. **Deduplication**: No content duplicated from ancestor CLAUDE.md files
+Report the run as failed, naming the affected files, if:
 
-**Mark as failure if:**
+- Any Write/Edit operation returned an error
+- A generated file exceeds 200 lines (see Length Constraint — factor into
+  `.claude/rules/<topic>.md` instead)
 
-- Any Write/Edit operation failed
-- File size is 0 bytes or exceeds 200 lines
-- Required sections are missing
-- `[TODO]` placeholder entries are present in the output
+Content quality (no `[TODO]` placeholders, specific and actionable conventions) and
+deduplication against ancestor files are constraints on how you write each file, applied as you
+write it — not a re-read pass afterward.
 
 ## Error Handling
 

packages/plugins/development-productivity/commands/update-claude-md.md

@@ -13,7 +13,6 @@ Update CLAUDE.md files based on staged git changes. The goal is to capture **con
 gotchas, and team preferences** that Claude cannot infer by reading code — not to inventory
 files or list dependencies. Run this **before committing** whenever staged changes reveal a non-obvious
 pattern, constraint, or workflow decision.
-constraint, or workflow decision.
 
 ## Usage
 
@@ -199,11 +198,9 @@ In **automated mode**, skip confirmation and apply only if the proposed update p
 Write the updated CLAUDE.md content. For new `.claude/rules/` files, create them at the
 nearest `.claude/` directory in the ancestry tree (or at the repository root if none exists).
 
-After writing, verify:
-
-- Line count is under 200 for each modified CLAUDE.md
-- No content duplicated from ancestor CLAUDE.md files
-- No `[TODO]` placeholder entries
+Deduplication (Step 6) and the no-`[TODO]` rule are constraints on what you write, applied
+before the write — not a re-read pass afterward. The 200-line overflow case is handled under
+Error Handling below.
 
 ### Step 11: Show Completion
 

packages/plugins/development-productivity/skills/generate-tests/SKILL.md

@@ -1,6 +1,6 @@
 ---
 description: Generate comprehensive tests for code. Use when user says "write tests for this function", "add unit tests to this file", "generate integration tests for the API", "I need test coverage for this module", or "create e2e tests for the checkout flow".
-allowed-tools: Read, Grep, Glob, Task(subagent_type:test-writer-agent), Task(subagent_type:context-loader-agent), Task(subagent_type:security-analyzer-agent)
+allowed-tools: Read, Grep, Glob, Task(subagent_type:test-writer-agent), Task(subagent_type:development-codebase-tools:context-loader-agent), Task(subagent_type:development-codebase-tools:security-analyzer-agent)
 model: sonnet
 ---
 
@@ -22,7 +22,9 @@ Generate comprehensive tests with advanced testing strategies, scenario generati
 2. **Select Strategy**: Choose testing approach based on code type
 3. **Generate Tests**: Create tests with appropriate framework
 4. **Identify Edge Cases**: Boundary conditions and error handling
-5. **Output Files**: Write test files with full coverage
+5. **Return Files**: Return the test file contents as artifacts. This skill has no Write tool
+   and no Bash, so it cannot write to disk or measure coverage — do not claim a coverage
+   figure, and name the behaviors you did not write a test for.
 
 ## Options
 
@@ -58,7 +60,11 @@ Invoke **test-writer-agent** agent with:
 - `testType`: Testing strategy
 - `requirements`: User stories (if provided)
 
-For complex scenarios, coordinate with **context-loader-agent** and **security-analyzer-agent**.
+For a single file or a few files in one module, invoke `test-writer-agent` directly — that is
+the common case. Only for work spanning several areas, coordinate with
+**development-codebase-tools:context-loader-agent** and
+**development-codebase-tools:security-analyzer-agent** (both live in another plugin, hence the
+prefix).
 
 ## Examples
 

packages/plugins/development-productivity/skills/generate-tests/gen-tests-guide.md

@@ -1,7 +1,7 @@
 ---
 description: Generate comprehensive tests with advanced testing strategies, scenario generation, and edge case identification using the enhanced test-writer agent.
 argument-hint: [paths...] [--framework jest|vitest|pytest|cypress|playwright] [--type unit|integration|e2e|all] [--strategy standard|scenario|property|mutation|accessibility] [--requirements "user stories"]
-allowed-tools: Read(*), Grep(*), Task(subagent_type:test-writer-agent), Task(subagent_type:context-loader-agent)
+allowed-tools: Read(*), Grep(*), Task(subagent_type:test-writer-agent), Task(subagent_type:development-codebase-tools:context-loader-agent)
 ---
 
 ## Inputs
@@ -52,10 +52,14 @@ For complex test generation (multiple files or --type all), use orchestration:
      - **security-analyzer-agent** for security test cases (if applicable)
 
 3. **Quality Assurance Phase**:
-   - Validate test coverage completeness
+   - Name the behaviors you did NOT write a test for, and why
    - Check for test anti-patterns
    - Ensure proper mocking strategies
 
+**When to skip the orchestration above:** a single file, or a handful of files in one module,
+is faster and more accurate done directly. The agent list below is a menu of what is available,
+not a set of steps to work through — invoke only the agents the specific task needs.
+
 ## Delegation
 
 ### Simple Case (single file, unit tests)
@@ -71,22 +75,25 @@ Invoke **test-writer-agent** with:
 
 ### Complex Case (multiple files or integration/e2e)
 
-**If agent-orchestrator-agent is available** (from development-codebase-tools plugin):
+The orchestrator and the loader/analyzer agents live in the **development-codebase-tools**
+plugin, so dispatch them with the cross-plugin form
+`Task(subagent_type: development-codebase-tools:<agent-name>)`. `test-writer-agent` is in this
+plugin and needs no prefix.
 
-Invoke it to coordinate:
+**If `development-codebase-tools:agent-orchestrator-agent` is available**, invoke it to coordinate:
 
-- **context-loader-agent**: Build comprehensive understanding
+- **development-codebase-tools:context-loader-agent**: Build comprehensive understanding
 - **test-writer-agent**: Generate test scenarios and implementations
-- **security-analyzer-agent**: Add security test cases (for APIs)
-- **performance-analyzer-agent**: Add performance benchmarks (for critical paths)
+- **development-codebase-tools:security-analyzer-agent**: Add security test cases (for APIs)
+- **development-codebase-tools:performance-analyzer-agent**: Add performance benchmarks (for critical paths)
 
-**Fallback (if agent-orchestrator-agent is not available)**:
+**Fallback (if the orchestrator is not available)**:
 
 Execute agents sequentially:
 
-1. First invoke **context-loader-agent** to gather context
+1. First invoke **development-codebase-tools:context-loader-agent** to gather context
 2. Then invoke **test-writer-agent** with the gathered context
-3. Optionally invoke **security-analyzer-agent** for API endpoints
+3. Optionally invoke **development-codebase-tools:security-analyzer-agent** for API endpoints
 4. Aggregate results manually
 
 ## Output Format
@@ -96,9 +103,15 @@ Execute agents sequentially:
   summary: string; // Overview of test generation strategy
   testStrategy: {
     approach: string; // Selected testing approach
-    coverage: number; // Estimated coverage percentage
-    edgeCasesIdentified: number;
-    scenariosGenerated: number;
+    // Coverage is only reported when a coverage tool actually ran and produced a
+    // number. This skill's allowed-tools include no Bash, so it cannot run one --
+    // in that case emit 'not_measured' and say which command the user should run.
+    // Never estimate a coverage percentage; a guessed number reads as a
+    // measurement and is not one.
+    coverage: number | 'not_measured';
+    coverageSource?: string; // Command whose output produced the number
+    edgeCasesIdentified: number; // Count of edge cases you actually enumerated below
+    scenariosGenerated: number; // Count of scenarios you actually generated below
   };
   suggestedTests: Array<{
     file: string; // Test file path

packages/plugins/development-productivity/skills/optimize-prompt/SKILL.md

@@ -44,10 +44,12 @@ Optimize prompts for better AI model performance using engineering techniques.
 
 ### 4. Model-Specific Optimization
 
-- GPT-4 best practices
-- Claude optimization techniques
+- Techniques for the target model family, checked against its current API rather than recalled
 - Prompt chaining strategies
-- Temperature/parameter tuning
+- Sampling parameter tuning — parameters are provider-specific. `frequency_penalty` and
+  `presence_penalty` are OpenAI-only and are rejected by the Anthropic Messages API. On the
+  Claude 5 family, `temperature` is rejected when extended thinking is enabled, and
+  `budget_tokens` is not accepted (use `thinking: {type: "adaptive"}`).
 - Token budget management
 
 ### 5. RAG Integration