[Sev7eNup]

What & why

Weekly documentation audit for 2026-10-25, run against origin/main 4f83f89. Week index 2964, documentation block 4 of 0–7 (44 files: grafana/README.md through src/NodePilot.Ai/Prompts/*.md), code slice 0 (activities, execution, triggers, scheduling). Every edit below is a documentation correction proven by the code. No code, tests or config were changed.

How (corrections and evidence)

AI prompts (src/NodePilot.Ai/Prompts/). These are sent to the LLM, so drift here changes runtime behaviour.

• workflow-system.md: the prompt told the model to reply with a bare {nodes, edges} object and "no other top-level keys". WorkflowGenerationService.GenerateAsync appends an output envelope {name, description, definition} to this same prompt, and ParseEnvelope throws "Envelope is missing 'definition' object" for the bare shape. The rule now points at the envelope.
• activity-reference.md: the edge example used sourceHandle: "out" / targetHandle: "in". Those are not valid ports. lib/edgePorts.ts accepts only top/right/bottom/left, with defaults right→left, and normalizes anything else. The example now uses right/left.
• activity-reference.md: added the missing <= and >= operators (EdgeConditionValidator.cs:18, ConditionEvaluator.cs:53,155-156).
• knowledge-system.md: "three kinds of questions" → "five". The list below it has five groups, and KnowledgeChatToolRegistry registers five gated tool groups.
• script-system.md: Write-Error does not set a non-zero exit code. Step success is based on errors, and the exit code comes only from $LASTEXITCODE (RunScriptActivity.cs:68,167; runScript promptNotes).

Other docs

• docs/testing/e2e/71-80.md: the restart and hot-reload section lists were incomplete. Added Agents and Performance (restart) and AiKnowledge and WaitForCondition (hot-reload), per SettingsSchema.cs. That gives 13 + 10, which matches CLAUDE.md.
• scripts/release-lab/README.md: 164 → 167 computer-use cases (computer-use/test_gate.py:23, computer-use README).
• scripts/tech-demo/README.md:
  • main.json has 39 nodes, not 37.
  • It uses 17 distinct types and contains no folderOperation.
  • The runScript capture now describes the wrapper's two nested scopes (PowerShellScriptWrapper.cs, PowerShellReservedVariables).
• scripts/tech-demo/xl.README.md: xl.json covers 26 types (20 activities + 6 triggers), not 25/19.
• scripts/README.md: the master workflow covers 27 of the 29 activities. It does not cover aiAgent or aiAgentTeam; the catalog has 29 per activity-config-reference.json.
• scripts/stress-test/README.md:
  • engine: auto runs as a Windows PowerShell 5.1 process, not as in-process runspaces (PowerShellEngineFactory.GetEngine).
  • The per-pass workloads now match main.json: √-sum 5·10⁶, SHA in blocks of 50 000, 25 000 regex lines, 200×200 matrix.
• scripts/ad-sso-labtest/README.md: the test-directory table and its cross-references now use the names Setup-LabDirectory.ps1 creates: svc-nptest-dir, nptest.*, NPTest-Access / NPTest-Admins. The README's own results table already used these names.

Testing

• No .NET SDK is available in the audit environment, so no dotnet test was run. The relevant guards, PromptCatalogTests and DocumentationCountsTests, are left to CI.
  • The PromptCatalogTests assertions ("nodes", "runScript", "Output rules", "PowerShell") are still satisfied by the edited prompts. I checked this by reading the test, not by running it.
  • None of the changed counts sit in a file covered by DocumentationCountsTests.
• I re-checked every corrected claim against its source a second time after editing.

Related issues (category B, not auto-fixed)

#642, #643, #644, #645, #646, #647

---

Generated by Claude Code