AI Docs Autopilot

Autopilot: Inventory → Generate → Submit

Run the full DDI documentation pipeline end-to-end with no user interaction. The agent inventories APIs from a CSV, generates documentation pages, and submits them as a PR — stopping only on error.

No local repo clones required. All repo interactions use the Azure DevOps REST API. OS source lookups use the substrate-mcp MCP server.

Parameters

Prerequisites

• The user provides a header name and a path to a CSV file. That is the only user requirement.
• Azure CLI (az) should be available for auth token acquisition. If not, the agent will prompt for an ADO Personal Access Token (PAT) once per session.
• The substrate-mcp MCP server must be accessible for source code retrieval.
• The microsoft.docs.mcp MCP server should be accessible for supplemental info (non-fatal if unavailable).

Autopilot Mode

This skill runs all three phases (inventory, generate, submit) in sequence with no user interaction between them. The process only stops if an error is encountered. Progress is reported to the console after each phase.

Progress Tracking

Track elapsed time and documents written across all three phases. Because each phase may run in a separate PowerShell process (variables don't survive between invocations), persist the start timestamp in a file so it can be read by later phases.

Start the timer (Phase 1 inventory script)

At the very beginning of the inventory.ps1 script, record the pipeline start time to a file in the working directory:

$pipelineStartTime = Get-Date
$pipelineStartTime.ToString('o') | Out-File -FilePath (Join-Path $workingDir '.pipeline-start') -Encoding utf8 -Force
$docsWrittenCount = 0

Log per-document progress (Phase 2)

Phase 2 is orchestrated by the agent (not a single script). After writing each documentation file, the agent must:

1. Read the start time from {workingDir}\.pipeline-start.
2. Increment a running document counter.
3. Report progress to the user with elapsed time.

Run this in the terminal after each file is written:

$start = [DateTime]::Parse((Get-Content '{workingDir}\.pipeline-start' -Raw).Trim())
$elapsed = (Get-Date) - $start
Write-Host "[$($elapsed.ToString('hh\:mm\:ss'))] Wrote doc {N}: {filename.md}"

Alternatively, the agent can compute elapsed time from the stored timestamp itself and include it in its console message — the key requirement is that each doc written produces a visible [HH:MM:SS] Wrote doc N: filename progress line.

Report totals (Phase 3 submit script)

At the end of the submit.ps1 script, read the start time back and compute the total elapsed time:

$startFile = Join-Path $workingDir '.pipeline-start'
if (Test-Path $startFile) {
    $pipelineStartTime = [DateTime]::Parse((Get-Content $startFile -Raw).Trim())
    $totalElapsed = (Get-Date) - $pipelineStartTime
} else {
    $totalElapsed = [TimeSpan]::Zero
}

Include the elapsed time and document count in the final summary banner (see Phase 3, step 15).

Phase 1: Inventory

Read a pre-provided CSV of target API filenames, cross-reference each entry against existing docs, stubs, and published content via the ADO REST API, classify their status, and finalize the CSV.

File Naming Conventions

Map each API entity to a filename using these prefixes:

The filename pattern is: {prefix}-{header}-{api_name_lowercase}.md

Where {header} is the header name without the .h extension.

Header Landing Page

On the stubs branch, the header landing page is always named na-{header}.md. When it is copied to wdk-ddi-src/content/{header}/ on the main branch, it is renamed to index.md. In the published docs repo, it also appears as index.md. When cross-referencing, check for na-{header}.md on the stubs branch and index.md on main and in the published docs repo.

Legacy Filename Exceptions

Some existing files in the published docs repo or on main may have non-standard filenames (e.g. an extra underscore like ns-header-_struct_name.md instead of ns-header-struct_name.md). These are historical naming errors. Do not rename existing files — doing so would break published links. When cross-referencing, also check for these variant filenames. For new APIs, always use the approved naming convention.

Execution Strategy

Write all PowerShell logic into a single self-contained .ps1 script file, then execute it in one terminal call. Do NOT run ADO REST calls or variable assignments as separate interactive terminal commands — PowerShell variables are lost between terminal invocations and long-running commands may time out and get moved to the background, breaking the workflow.

The pattern is:

1. Create a script file at {workingDir}\inventory.ps1 containing all the logic from the Inventory Procedure below. Use the create_file tool to write the script — this auto-saves the file to disk immediately, avoiding any unsaved-buffer issues.
2. Execute it in a single terminal call: powershell -ExecutionPolicy Bypass -File "{workingDir}\inventory.ps1"
3. Parse the script's console output to present results to the user.

The script should accept no parameters — hardcode the {header}, {csvPath}, {workingDir}, and {userAlias} values directly into the generated script.

Inventory Procedure

Write a single inventory.ps1 script that performs all of the following steps, then execute it.

1. Strip the .h extension from the user-provided header name to get {header} (e.g. soundwireclass.h → soundwireclass).

2. Resolve the user alias for branch naming ({user-alias}). The alias identifies who is submitting the PR, not who owns the APIs. Try these sources in order and use the first non-empty value: a. The Windows username: $env:USERNAME. b. The Azure CLI identity: az account show --query user.name -o tsv, extracting the alias portion before @.

   The resolved alias is used later for the branch name {user-alias}/{header}-update. Hardcode it into the generated scripts.

3. Resolve paths. The user provides a CSV path. Derive the working directory from it. The script should validate the CSV exists and read it:

   $ErrorActionPreference = "Stop"
   $header = "{header}"
   $csvPath = "{user-provided CSV path}"
   $workingDir = Split-Path $csvPath -Parent
   
   # Start the pipeline timer and persist it for later phases
   $pipelineStartTime = Get-Date
   $pipelineStartTime.ToString('o') | Out-File -FilePath (Join-Path $workingDir '.pipeline-start') -Encoding utf8 -Force
   $docsWrittenCount = 0
   
   if (-not (Test-Path $csvPath)) {
       Write-Error "CSV not found at $csvPath."
       exit 1
   }

   Read the CSV:

   $csvData = Import-Csv $csvPath

   The CSV has a header row. Look for a column containing file paths (commonly FilePath or filename) with values like wdk-ddi-src/content/{header}/{filename}.md. Filter to rows with valid .md paths.

4. Obtain ADO auth token. Try Azure CLI first, then fail with a clear message (PAT prompting cannot work inside a non-interactive script):

   $token = (az account get-access-token --resource 499b84ac-1321-427f-aa17-267ca6975798 --query accessToken -o tsv 2>$null)
   if (-not $token) {
       Write-Error "Failed to get ADO token. Run 'az login' first, or set `$env:ADO_PAT before running."
       exit 1
   }
   $h = @{ Authorization = "Bearer $token" }
   $adoBase = "https://dev.azure.com/cpubwin/drivers/_apis/git/repositories"

5. For each entry in the CSV, extract the target filename from the path (e.g. wdk-ddi-src/content/{header}/nf-soundwireclass-somefunc.md → nf-soundwireclass-somefunc.md).

6. List all files in the directory on each branch (one API call per branch, not per file). This detects legacy filename variants (see Legacy Filename Exceptions):

   # List files on main branch of wdk-ddi
   try {
       $r1 = Invoke-RestMethod -Uri "$adoBase/wdk-ddi/items?scopePath=wdk-ddi-src/content/{header}/&recursionLevel=OneLevel&versionDescriptor.version=main&versionDescriptor.versionType=branch&api-version=7.0" -Headers $h
       $mainFiles = @($r1.value | Where-Object { $_.path -like "*.md" } | ForEach-Object { Split-Path $_.path -Leaf })
   } catch { $mainFiles = @() }
   
   # List files on stubs branch of wdk-ddi
   try {
       $r2 = Invoke-RestMethod -Uri "$adoBase/wdk-ddi/items?scopePath=wdk-ddi-src/content/{header}/&recursionLevel=OneLevel&versionDescriptor.version=stubs/main&versionDescriptor.versionType=branch&api-version=7.0" -Headers $h
       $stubFiles = @($r2.value | Where-Object { $_.path -like "*.md" } | ForEach-Object { Split-Path $_.path -Leaf })
   } catch { $stubFiles = @() }
   
   # List files on live branch of wdk-ddi-build (published docs)
   try {
       $r3 = Invoke-RestMethod -Uri "$adoBase/wdk-ddi-build/items?scopePath=wdk-ddi-src/content/{header}/&recursionLevel=OneLevel&versionDescriptor.version=live&versionDescriptor.versionType=branch&api-version=7.0" -Headers $h
       $pubFiles = @($r3.value | Where-Object { $_.path -like "*.md" } | ForEach-Object { Split-Path $_.path -Leaf })
   } catch { $pubFiles = @() }
   
   Write-Host "Main: $($mainFiles.Count) files, Stubs: $($stubFiles.Count) files, Published: $($pubFiles.Count) files"

   Compare each expected filename against the directory listings. If a file exists with a minor variant (e.g. ns-header-_struct_name.md instead of ns-header-struct_name.md), treat it as a match — the doc exists under the legacy name. Use the legacy filename in the CSV (do not rename it).

   a. Check if the file exists on main of wdk-ddi (look for the filename in $mainFiles).

   b. Check if a stub exists on the stubs branch (look for the filename in $stubFiles).

   c. Check if the file exists on live of the published docs repo wdk-ddi-build (look for the filename in $pubFiles).

7. Classify each API into one of three categories:

   • new: Stub exists on stubs/main but no completed doc on main
   • update: File already exists on main — needs to be compared with the latest stub to identify changes
   • stub not found: No stub and no existing doc

8. Output grouped results to the console for the agent to parse and present:

   === New (ready to document) ===
   nf-soundwireclass-somefunc.md          (function)
   ns-soundwireclass-some_struct.md       (structure)
   
   === Update (existing doc, needs diff) ===
   nf-soundwireclass-otherfunc.md         (function)  [+2 params, -1 param, ~1 param]
   
   === Stub not found ===
   ne-soundwireclass-some_enum.md         (enumeration)

9. Diff existing docs against stubs. For each "update" entry, fetch the existing doc from main and the latest stub from stubs/main via the ADO REST API. Compare the parameter/field/value sections to identify:

   • Added parameters, fields, or enum values (present in stub but not in existing doc)
   • Removed parameters, fields, or enum values (present in existing doc but not in stub)
   • Modified parameters, fields, or enum values (present in both but with different types or descriptions)

   Output a per-file change summary (e.g. [+2 params, -1 param, ~1 param]) in the console results.

10. Write the updated CSV at the original CSV path. The format must be compatible with Phase 2:

    filename,status
    wdk-ddi-src/content/{header}/{filename1}.md,new
    wdk-ddi-src/content/{header}/{filename2}.md,update

    The script writes both "new" and "update" entries. "Stub not found" entries are excluded.

11. Output a summary to the console:

    • Number of APIs in the CSV by type
    • Number selected for documentation
    • Working directory: {workingDir}
    • CSV file location: {csvPath}

After the script finishes, the agent should verify the CSV was written successfully. If the inventory found zero entries for documentation (all "stub not found"), stop with an error. Otherwise, proceed immediately to Phase 2 with no user prompt.

Phase 2: Generate API Docs

Generate complete API reference documentation pages for WDK DDI entities by combining stub files, OS source code declarations, and supplemental information from published docs.

Generate Procedure

1. Strip the .h extension from the user-provided header name to get {header} (e.g. soundwireclass.h → soundwireclass).

2. Resolve paths. The user provides a CSV path. Derive the working directory from it:

   $csvPath = "{user-provided CSV path}"
   if (-not (Test-Path $csvPath)) {
       Write-Error "CSV not found at $csvPath. Please provide the CSV file first."
       return
   }
   $workingDir = Split-Path $csvPath -Parent
   $outputDir = Join-Path $workingDir "output"
   Import-Csv $csvPath

   If the CSV does not exist, stop with an error.

3. Obtain ADO auth token. Try Azure CLI first, then fall back to prompting for a PAT:

   try {
       $token = (az account get-access-token --resource 499b84ac-1321-427f-aa17-267ca6975798 --query accessToken -o tsv 2>$null)
       if (-not $token) { throw "No token" }
       $headers = @{ Authorization = "Bearer $token" }
   } catch {
       $pat = Read-Host "Enter ADO PAT (scope: Code Read)"
       $base64 = [Convert]::ToBase64String([Text.Encoding]::ASCII.GetBytes(":$pat"))
       $headers = @{ Authorization = "Basic $base64" }
   }
   $adoBase = "https://dev.azure.com/cpubwin/drivers/_apis/git/repositories"

4. Read the style guide remotely from the wdk-ddi repo to ensure all formatting rules are followed:

   $styleGuide = Invoke-RestMethod -Uri "$adoBase/wdk-ddi/items?path=.github/copilot-instructions.md&versionDescriptor.version=main&versionDescriptor.versionType=branch&api-version=7.0" -Headers $headers

5. Parse the CSV file to get the list of target filenames. The CSV has a header row and one filename column with paths like wdk-ddi-src/content/{header}/{filename}.md.

6. Create the output directory if it doesn't exist; if it does, clear it:

   if (Test-Path $outputDir) {
       Remove-Item "$outputDir\*" -Force -ErrorAction SilentlyContinue
   } else {
       New-Item -ItemType Directory -Path $outputDir -Force | Out-Null
   }

7. For each target file listed in the CSV, skip index.md (the header landing page is handled separately in step 8).

   For each remaining target file:

   a. Get the stub or existing file

   Use the ADO REST API to check if the file already exists on main:

   # List files on main branch
   $mainFiles = Invoke-RestMethod -Uri "$adoBase/wdk-ddi/items?scopePath=wdk-ddi-src/content/{header}/&recursionLevel=OneLevel&versionDescriptor.version=main&versionDescriptor.versionType=branch&api-version=7.0" -Headers $headers

   If the target filename appears in the listing, retrieve it from main and use its content as the starting point — preserve existing text, but always update ms.date to today's date when making any content changes. Only write the file to the output folder if you make changes. Do not copy unchanged files.

   $content = Invoke-RestMethod -Uri "$adoBase/wdk-ddi/items?path=wdk-ddi-src/content/{header}/{filename.md}&versionDescriptor.version=main&versionDescriptor.versionType=branch&api-version=7.0" -Headers $headers
   $content | Out-File -FilePath (Join-Path $outputDir "{filename.md}") -Encoding utf8

   If it does not exist on main, retrieve the stub from the stubs branch:

   $content = Invoke-RestMethod -Uri "$adoBase/wdk-ddi/items?path=wdk-ddi-src/content/{header}/{filename.md}&versionDescriptor.version=stubs/main&versionDescriptor.versionType=branch&api-version=7.0" -Headers $headers
   $content | Out-File -FilePath (Join-Path $outputDir "{filename.md}") -Encoding utf8

   b. Retrieve source declaration from the OS source tree

   Use the substrate-mcp MCP server to find the API's source declaration:

   • Extract the API name from the filename (e.g. nf-soundwireclass-somefunc.md → SomeFunc)
   • Search for the definition:

     search_code(query="def:{ApiName}", ext="h")

   • If def: doesn't find it, try keyword search:

     search_code(query="{ApiName}", ext="h", path="**/{header}.h")

   • Once found, use get_file_content with a line range to retrieve the full declaration including:
     • SAL annotations (_In_, _Out_, _Inout_, _In_opt_, _Out_opt_, _Reserved_)
     • Return type
     • All parameter names and types
     • Any adjacent comments describing the function
   • For structures/enums, retrieve the full type definition with all fields/values

   c. Gather supplemental information

   Use the microsoft.docs.mcp server's microsoft_docs_search tool to search for any existing published docs or related conceptual content for the API:

   microsoft_docs_search(query="{ApiName} WDK driver")

   d. Find formatting models

   Find 1-2 completed reference pages in the same header folder to use as formatting models. Use the ADO REST API to list files, then retrieve them:

   # List existing completed docs of the same entity type in the {header} folder
   $mainFiles = Invoke-RestMethod -Uri "$adoBase/wdk-ddi/items?scopePath=wdk-ddi-src/content/{header}/&recursionLevel=OneLevel&versionDescriptor.version=main&versionDescriptor.versionType=branch&api-version=7.0" -Headers $headers
   # Filter for files matching the same prefix (e.g. "nf-") and pick 1-2

   Then read the model files:

   $modelContent = Invoke-RestMethod -Uri "$adoBase/wdk-ddi/items?path=wdk-ddi-src/content/{header}/{model-filename.md}&versionDescriptor.version=main&versionDescriptor.versionType=branch&api-version=7.0" -Headers $headers

   If no completed pages exist in the same folder, look in a related header folder. Read these model files to match their style.

   e. Write the completed documentation page

   Write the completed file to {outputDir}\{filename.md}. Follow these rules:

   YAML Frontmatter

   • Preserve all existing metadata from the stub
   • Add ai-usage: ai-assisted if not already present
   • Set tech.root to the same value as existing files in the same header folder. If no existing files in the folder (i.e. a brand-new API set), make an educated guess based on the header's source location, API naming patterns, and functionality (e.g. headers in minkernel/ are typically kernel, networking headers are netvista, storage headers are storage, etc.). Log the guessed value and add a note for the user: "NOTE: tech.root guessed as '{value}' — please verify this is correct."
   • Always set ms.date to today's date in MM/DD/YYYY format — both for new files and when updating existing files with changes
   • Verify req.header, f1_keywords, api_name, topic_type are correct
   • req.header uses lowercase (e.g. classpnp.h). req.include-header uses sentence capitalization (e.g. Classpnp.h). Note that req.include-header does not always require a value — omit it or leave it empty when an include header is not applicable.
   • req.construct-type for macros: If the source declaration is a preprocessor macro (#define), set req.construct-type: macro, not function. The file prefix remains nf- and the title should say "macro" (e.g. CLEAR_FLAG_NOFENCE macro (classpnp.h)).
   • Leading underscores in UIDs: If the stub file's UID uses a leading underscore (e.g. NE:wdbgexts._POOL_HEADER_FIELD_NAME), preserve that underscore in the UID of the generated document. However, all customer-facing content (title, description, headings, prose) must use the version without the leading underscore (e.g. POOL_HEADER_FIELD_NAME). For api_name: and f1_keywords:, include both versions — the underscored name and the public name — so the page is discoverable by either form.

   Content Sections (in order)

   ## -description

   • Write a concise one-paragraph summary of what the API does
   • Focus on the public API behavior, not internal implementation
   • Use bold for API names referenced in text

   ## -parameters (functions/callbacks/macros)

   • One ### -param {Name} [{direction}] subsection per parameter
   • Map SAL annotations to direction labels:
     • _In_ → [in]
     • _Out_ → [out]
     • _Inout_ → [in, out]
     • _In_opt_ → [in, optional]
     • _Out_opt_ → [out, optional]
     • _Reserved_ → [in] with note "Reserved. Must be zero/NULL."
   • Describe each parameter concisely
   • Macro parameters requiring address-of (&): If the macro implementation takes the address of a parameter (e.g. &(Flags)), describe the parameter as "An addressable {type} storage location (such as a variable or structure field)" rather than "A {type} value." A macro that takes &(param) requires an lvalue, not an arbitrary expression.

   ## -struct-fields (structures only)

   • One ### -field {FieldName} subsection per field
   • Union types use the structure classification. If the C source declares a union, it is still documented with the ns- file prefix and req.construct-type: structure. All prose (description, field descriptions, remarks, and cross-references from other pages) must call it a "structure", never a "union." The heading should read # {NAME} structure, not # {NAME} union.

   ## -enum-fields (enumerations only)

   • One ### -field {ValueName}:{NumericValue} subsection per value (if numeric value is known)

   ## -ioctlparameters (IOCTLs only)

   • Document input buffer, output buffer, and status block

   ## -returns (functions with return values)

   • Document the return type and possible return values
   • For NTSTATUS returns, list common status codes

   ## -remarks

   • Add usage notes, IRQL requirements, important caveats
   • Include code examples when helpful
   • If the entity is an extended version (ends in Ex, or a number like 2, 3, or V2), reference the base entity and document only the extensions
   • Don't generate repetitive information
   • Don't include internal implementation details

   ## -see-also

   • List related APIs as relative Markdown links, one per line
   • Write entries in alphabetical order
   • Use relative paths: [**OtherFunc**](nf-{header}-otherfunc.md) for same-header APIs
   • Use ../ paths for cross-header: [**CrossFunc**](../otherheader/nf-otherheader-crossfunc.md)

   After writing each file, log progress with elapsed time. Run the following in the terminal (or compute the elapsed time from the stored timestamp and include it in the agent's console output):

   $start = [DateTime]::Parse((Get-Content '{workingDir}\.pipeline-start' -Raw).Trim())
   $elapsed = (Get-Date) - $start
   Write-Host "[$($elapsed.ToString('hh\:mm\:ss'))] Wrote doc {N}: {filename.md}"

   The agent must track the running document count itself (incrementing after each file) and substitute {N} with the current count and {filename.md} with the actual filename.

   Formatting Rules

   • Use bold for API names in prose: **IoCreateFile**
   • Use italic for parameter name references in prose: *DesiredAccess*
   • Use site-relative links for learn.microsoft.com content: [title](/windows-hardware/drivers/...)
   • Code blocks use cpp language tag
   • Ensure exactly one blank line at the end of the file

8. Handle the header landing page (index.md). The landing page is automatically updated during the build to reflect the current API set, so it almost never needs manual editing once published.

   • Check if index.md already exists on main of wdk-ddi or staging of the published docs repo using the ADO REST API:

     # Check wdk-ddi main
     $mainItems = Invoke-RestMethod -Uri "$adoBase/wdk-ddi/items?scopePath=wdk-ddi-src/content/{header}/&recursionLevel=OneLevel&versionDescriptor.version=main&versionDescriptor.versionType=branch&api-version=7.0" -Headers $headers
     
     # Check windows-driver-docs-ddi staging
     $pubItems = Invoke-RestMethod -Uri "$adoBase/windows-driver-docs-ddi/items?scopePath=wdk-ddi-src/content/{header}/&recursionLevel=OneLevel&versionDescriptor.version=staging&versionDescriptor.versionType=branch&api-version=7.0" -Headers $headers

   • If index.md appears in either listing, skip it — no action needed.
   • If no published index.md exists, retrieve the stub from the stubs branch and save it as index.md:

     $stubContent = Invoke-RestMethod -Uri "$adoBase/wdk-ddi/items?path=wdk-ddi-src/content/{header}/na-{header}.md&versionDescriptor.version=stubs/main&versionDescriptor.versionType=branch&api-version=7.0" -Headers $headers
     $stubContent | Out-File -FilePath (Join-Path $outputDir "index.md") -Encoding utf8

   • Content rules for a new index.md: After retrieving the stub, ensure the ## -description section follows this strict format:
     1. Write no more than one sentence describing the header.
     2. Immediately follow it with: This header is used by {tech.root display name}. For more information, see: [{tech.root display name}](../{tech.root}/index.md). Replace {tech.root display name} with the human-readable technology area name and {tech.root} with the tech.root value from the frontmatter (e.g. _netvista → [Networking](/windows-hardware/drivers/ddi/_netvista/)).
     3. Do not add any sections after -description. There must be no -remarks or -see-also sections in an index.md file.

9. Close editor tabs and log progress. After all files have been written to {outputDir}, close all open editor tabs so the workspace is not cluttered with generated files. Run the VS Code command workbench.action.closeAllEditors to close them. Then display:

   • List of all generated/updated files in {outputDir}
   • For each file, show the first 20 lines (frontmatter + description) as a sample
   • Highlight any metadata fields marked as TODO that the user should fill in manually (e.g. req.lib, req.dll, req.irql if not determinable from source)

   Then proceed immediately to Phase 3 with no user prompt.