Web Manual Maintenance Bundles

Bundle maintenance

Open this when changing the manual bundle or checking whether a fact belongs in a script, asset, or reference. The parent manual owns first-call behavior; operation-contract.md owns exact Web semantics; the tier index owns recovery routing.

Retained procedure support

Scripts describe their own implementation only, not the public Web contract. Assets are historical hints, not current authority: fixed free/unlimited/quota claims, example contacts, package names and endpoints require upstream checking. Inspect current vendor documentation for endpoints, quotas, dependencies, and prices; this bundle grants no install, credential, paid-use, network, or access-control authority. Never put secrets or private paths in examples or diagnostics.

Known legacy limitations

• cached_get.py has no imported JSON module and its json argument shadows that name. A normal successful GET reaches json.dump and raises AttributeError after opening its cache file for writing; an existing file may already have been truncated. Cache reads also fail. Do not recommend it as a working cache or run it over shared/pre-existing data. Its cache key includes URL/method but excludes params, headers and identity. No automatic purge occurs: old claims about one-day/200-entry eviction were not implemented.
• Cache defaults are /tmp/web-browsing-cache, overridden by WEB_BROWSING_CACHE_DIR, and TTL 3600, overridden by WEB_BROWSING_CACHE_TTL at import. clear_cache(url) unlinks that GET entry; clear_cache() unlinks every matching JSON file. Both require explicit deletion authorization; neither is a diagnostic. The CLI prints a preview.
• assets/extraction-pipeline.json is not valid JSON (500/month is an unquoted expression). Read it only as legacy notes, not machine configuration. The other parseable assets are not automatically consumed by public Web.
• extract_page.py is a separate Requests/browser/vendor helper, not the vetted public BrowserEngine. Its preview fields are truncated (including the JSON saved by --json); exit 0 means no error key, not complete text or useful content. CrossRef helper is_oa actually contains a citation count, not an OA boolean; verify OA against the real source. --save/--json overwrite paths.
• --fallback can advance through local tiers to Jina, sending the target URL to an external service. auto_tier is a heuristic; explicit --tier still may do several metadata requests. Authorize the whole selected route, dependencies and data disclosure before execution; no implicit chain. Helper diagnostics may include raw URLs/exceptions: never feed secrets or private URLs. Tier 3 blocks images/styles/fonts/media and returns only a preview, so it is not a visual-validation harness.

These are preserved implementation limitations, not runtime fixes in the manual reduction. For a proposed repair, use the lingtai-issue-report protocol and obtain filing/implementation authority; do not silently patch installed helpers.

Cleanup / Footprint

Public Web leaves complete tmp/tool-results/web-* artifacts; snapshots, cursors and references are bounded process-local state. Optional helpers leave explicit downloads/JSON outputs and the cache described above. None are implicitly disposable: preserve live referenced results, evidence, user data, configuration and secrets, and never blindly clear a shared cache.

After a large retrieval session or before retirement, follow skills-manual → reference/cleanup-footprint-contract.md#shared-footprint-check-recipe. Select only this Agent's tmp/tool-results/web-* plus the exact task-created output paths; inspect a configured helper cache separately only when authorized. The shared recipe reports count/bytes without changing inspected files; its optional logs/cleanup.jsonl append is a separate explicit write. Present a dry-run listing what stays/goes, obtain explicit human consent, then perform only the approved cleanup and record timestamp/tool/dry-run-or-apply/count/bytes/ path summary/approval. If consent is absent, stop at inspection.