Doc Sync

Purpose

Scan documentation drift against the implementation and produce focused GitHub PRs. Process these phases in order:

1. Canonical documentation: keep the six website sections and newcomer path accurate.
2. Architecture context and traceability: maintain ADRs and evidence status without upgrading incomplete evidence into a compliance claim.
3. Compatibility redirects: preserve retired routes without restoring a second source of truth.

Do not guess based on filenames. Read the implementation, tests, manifests, and live command help before changing behavioral prose or command examples.

Keep evidence tiers distinct:

1. a path, symbol, option, or command exists;
2. static and documentation checks pass;
3. CPU tests pass;
4. target-hardware execution passes; and
5. model parity, performance, or qualification is proven.

Canonical Information Architecture

The maintained website sources are:

The navigation in website/sidebars.js and website/docusaurus.config.js must expose those sources consistently. website/docs/wiki/, website/docs/operations/, website/docs/unit-design/, and retired architecture pages are compatibility routes, not maintained content sources.

Change Set

Fetch current GitHub main:

git fetch github main
DOC_SYNC_CURRENT_SHA=$(git rev-parse github/main)

If website/docs/context/.last_scan_sha exists, use it only after proving that it names a commit and is an ancestor of the current main SHA:

DOC_SYNC_LAST_SHA=$(cat website/docs/context/.last_scan_sha)
git cat-file -e "${DOC_SYNC_LAST_SHA}^{commit}"
git merge-base --is-ancestor \
  "$DOC_SYNC_LAST_SHA" "$DOC_SYNC_CURRENT_SHA"
git diff --name-status \
  "$DOC_SYNC_LAST_SHA".."$DOC_SYNC_CURRENT_SHA"

When the marker is missing or invalid, choose and report an explicit baseline. Do not silently substitute a date heuristic. Update the marker only when that state change was requested and the scan completed. If there is no change set, report that result instead of inventing documentation churn.

Route Changed Surfaces

Current model ownership spans:

• python/tensorrt_model_connect/families/<family>/;
• src/runtime/models/<family>/; and
• tests/e2e/models/<family>/.

Do not resurrect removed Source workflows, retired task-evaluation commands, root graph helpers, or paths copied from another branch.

CI Documentation Contract

Source contains Community CPU, PR Metadata, the Internal CI Bridge, and Pages workflows. Premerge, including legal compliance, and nightly orchestration are private Internal CI.

• run-internal-ci is the one-shot trusted trigger; run-ci is retired.
• The bridge verifies the label event, current PR metadata head, and successful Community CPU result for that head before dispatching it.
• Source receives only TRTMC Internal CI / Automated premerge gate as PENDING, PASS, or FAIL on that exact head.
• A successful bridge dispatch is not a successful premerge result.
• tools/public_failure/ is the closed public-report validator and plain-text renderer used by the trusted bridge. It never publishes protected raw logs or arbitrary diagnostic text.
• Never copy private logs, artifacts, package coordinates, runner details, or internal URLs into Source docs, Actions, Pages, or PR comments.

Read the workflow and tools/ci/README.md at the target SHA rather than hard-coding private job names or topology.

Scan Classes

For paths, symbols, counts, and commands:

• verify repository-relative paths on disk;
• verify symbols, manifest fields, runtime strategies, and CLI options in code or executable help;
• recompute counts and label snapshot values as such;
• execute safe --help, --list, --dry-run, and parser checks; and
• search modified docs and skills for removed names and paths.

python3 tools/check_doc_file_references.py --strict <changed-doc-directory>

For model-support claims, cross-check all three ownership descriptors and the exact bundle/runtime path. Registration, build success, dry-run planning, parity, performance, and qualification are different claims.

The supported Dev/QA entry point is tools/trtmc_validate.py. Workloads live in tests/validation/workloads.yaml, bindings in tests/validation/model_workloads.yaml, and implementation in tools/validation/. The persisted task_eval artifact key is intentional; do not describe it as a live legacy CLI.

Branch And PR Flow

Create one short-lived branch per coherent documentation owner or review concern. Split phases when their ownership or validation differs; keep them together when one focused review can cover the change:

DOC_SYNC_PHASE_NAME="canonical"  # or "context" or "redirects"
DOC_SYNC_DATE=$(date +%Y-%m-%d)
DOC_SYNC_BRANCH="doc-sync/${DOC_SYNC_PHASE_NAME}-${DOC_SYNC_DATE}"

git fetch github main
git switch -c "$DOC_SYNC_BRANCH" github/main

Skip a phase if the remote branch already exists:

git ls-remote --heads github "$DOC_SYNC_BRANCH"

Push and open a GitHub PR:

git push -u github HEAD
gh pr create \
  --repo NVIDIA/TensorRT-Model-Connect \
  --base main \
  --head "$DOC_SYNC_BRANCH" \
  --title "docs(${DOC_SYNC_PHASE_NAME}): automated doc sync $(date +%Y-%m-%d)" \
  --body-file pr-body.md

Do not attach labels, update scan markers, create issues, or publish private evidence unless the user requested that state change.

Creating or pushing the PR does not start premerge. After confirming that the PR's headRefOid equals the pushed SHA, follow $submit-github-pr and $pr-babysitter: an authorized collaborator applies the one-shot run-internal-ci label, and the exact head must receive a successful TRTMC Internal CI / Automated premerge gate status. Never use the retired run-ci label or publish private Internal CI logs, artifacts, runner details, or URLs in Source.

Use $write-git-messages for the commit and PR text.

Phase 1: Canonical Documentation

Start from current repository ground truth:

find python/tensorrt_model_connect/families/ -name "MODEL.toml" | sort
find src/runtime/models/ -name "MODEL.toml" | sort
find tests/e2e/models/ -path "*/manifests/*.json" | sort
find src/runtime/registry/ -type f \( -name "*.cpp" -o -name "*.h" \) | sort
find src/runtime/domains/ -type f \( -name "*.cpp" -o -name "*.h" \) | sort
find include/trtmc/ -type f -name "*.h" | sort
python3 tools/model_ci.py validate
python3 tools/test_impact.py --validate

Check the newcomer journey in order:

1. getting-started/environment-and-repro.md states supported host, container, Python, CUDA, TensorRT, and hardware assumptions.
2. getting-started/installation.md installs the prerequisites used by the first inference.
3. getting-started/quick-start.md contains one canonical NLP build, inspection, and inference flow.
4. learning-path.md continues from that exact result instead of restarting with a competing quick start.

Run every changed command locally when the required dependencies and hardware are available. Otherwise validate the parser or --help, inspect the implementation that consumes the argument, and record the missing runtime or hardware proof explicitly.

Review the remaining canonical sections:

Behavioral descriptions require reading the implementation first. A parser accepting an option is not evidence that every family supports or qualifies the feature.

Phase 2: Architecture Context And Traceability

ADR maintenance

Scan website/docs/context/adr/*.md excluding README.md.

• Parse frontmatter: number, title, status, date, source_commits, and superseded_by.
• Verify referenced paths, symbols, strategies, and counts against the current tree.
• For a moved path, set a concrete value such as OLD_PATH=src/runtime/old_file.cpp, then run git log --follow --diff-filter=R -- "$OLD_PATH".
• Promote Proposed ADRs only when the decision is clearly implemented.
• Mark an ADR Superseded only when a maintained ADR names the replacement.
• If a source_commits SHA is not an ancestor of GitHub main, document the provenance problem rather than silently rewriting history.
• Rebuild website/docs/context/adr/README.md from frontmatter, sorted by number.

Traceability and safety status

website/docs/context/traceability-and-safety.md is the canonical status and gap report. It is not a normative traceability matrix or a safety case.

Recompute its snapshot from the manifests and test annotations:

find tests/e2e/models/ -path "*/manifests/*.json" | sort
rg -n "Trace:|Trace ID:" tests/builder tests/tools --glob "*.py"
rg -n "Trace:|Trace ID:" tests/cpp --glob "*.cpp"

Verify manifest count, testcase count, empty trace IDs, duplicate IDs, source revision, and date. Keep evidence tiers separate: an implemented check, a passing CPU/static test, a GPU E2E result, qualified performance, and a certification claim are not interchangeable.

Do not add or reassign source/test trace IDs as part of a documentation-only PR. Report those implementation and test gaps for a separately reviewed change.

Phase 3: Compatibility Redirects

Audit every Markdown file below:

website/docs/wiki/
website/docs/operations/
website/docs/unit-design/

Also audit retired canonical paths such as website/docs/architecture/runtime-plugins.md.

Each compatibility page must:

• set unlisted: true;
• render a real Docusaurus <Redirect> to an existing canonical route;
• avoid navigation, pagination, diagrams, or maintained explanatory content;
• contain no unique factual or policy source that is absent from its destination; and
• build successfully so old external links do not become 404s.

If a test or tool still consumes a retired page as normative input, migrate that consumer to the canonical page in the same logical change. Until that consumer is migrated, report the retired route as a compatibility blocker rather than treating its duplicated text as canonical.

Do not add new pages under the retired directories.

Validation

Run at minimum:

git diff --check
python3 tools/check_doc_file_references.py --strict website/docs
python3 tools/check_doc_file_references.py \
  --strict plugins/trtmc-agent-skills/skills/doc-sync
python3 tools/model_ci.py validate
python3 tools/test_impact.py --validate
npm --prefix website ci
npm --prefix website run test:model-support
npm --prefix website run build

For a broad documentation contract change, the representative focused Python set is:

PYTHONPATH=python:. python3 -m pytest \
  tests/tools/test_check_doc_file_references.py \
  tests/tools/test_github_actions_ci.py \
  tests/tools/test_runtime_strategy_matrix_checker.py \
  tests/tools/test_model_owned_validation_scripts.py \
  tests/tools/test_validation_engine.py \
  tests/tools/test_test_impact.py \
  tests/tools/test_trtmc_validate.py \
  tests/tools/test_perf_matrix.py::test_release_suite_covers_every_non_l0_ready_model_profile \
  -q

Run focused Python or C++ tests for commands, contracts, tests, and tooling touched by the documentation change. Use the Node version and GitHub Pages environment documented in website/docs/reference/testing.md when reproducing the production website build.

A green Docusaurus build proves that the site compiles; it does not prove GPU inference, parity, output quality, performance, or model qualification. Record what ran, the exact revision, and every environment boundary in the PR body. npm ci needs network access and changes the local dependency tree; report missing dependencies, baseline failures, and unrun evidence tiers explicitly.

Summary Report

End each run with:

Doc-Sync Summary (YYYY-MM-DD)
Commits scanned: N (<last>..<current>)
Canonical docs: <changes or no changes> <PR URL or skipped>
Architecture context: <changes or no changes> <PR URL or skipped>
Redirect compatibility: <changes or no changes> <PR URL or skipped>
Command evidence: <tests and runtime boundaries>
Blockers: <none or details>

State the baseline and target SHA, corrected facts and source paths, exact checks and outcomes, unrun evidence tiers, known baseline gaps, and PR URLs. “No change” is valid when the current documentation matches current evidence.