fix(local-pages): show native Claude record rates and resilient probes - #951

Scope

Fleet now shows exported Claude API-request and tool-result record rates from native Loki, with the exact query window and separate datasource links. Protocol errors preserve dated workstation readings, and the vLLM probe uses the documented model-list API. The remaining review corrections update the retirement wording, source references and two previously non-discriminating tests.

• Base: a7bb4a51eea9cc321dd06dd6243a5b2654a5f805, with feat(local-pages): show native lanes, rates and model services #941 landed.
• Head: e1ef838a8ed835ace75e09644ce41b1049e7b591.
• Lane: lane:foundation.
• Exactly ONE follow-up commit; no rebase.
• Owned paths: Fleet data/view and their tracking tests; workstation collector/test; local-pages README; Fleet decision; retirement decision and its registered evidence hash.

The six changes are:

1. Use native numeric Loki rate(...[5m]) queries on current port 21300 for Claude API-request and tool-result records. Keep exported records separate from all API attempts, completed turns and MCP counts. Preserve Loki-only lanes; missing, invalid, stale and unqualified zero rates remain UNKNOWN.
2. Probe http://127.0.0.1:28231/v1/models without reading the response body. This observes model-list API reachability, independently of scrape and unit state.
3. Describe the archived item as “a private repository with its stash”, leaving one consistent host placeholder in the retirement decision; refresh its evidence hash.
4. Correct the Fleet decision and README to the implemented design and immutable Loki/Grafana/vLLM sources. Per-lane Explore links retain the row's datasource and actual observation window, including delayed API/tool evaluation times.
5. Replace the weak Explore passthrough and service-only zero tests with actual collector/render and Codex token/MCP-zero regressions. Production mutants demonstrate that broken datasource/value/window/zero behavior fails assertions.
6. Catch HTTPException and its protocol subclasses in workstation collection, preserving exact CC values/read dates and publishing exception categories without payloads.

SOTA sources

• Native repository at a7bb4a51eea9cc321dd06dd6243a5b2654a5f805: observability/lanes_dashboard.py Claude api_request selection and native per-lane tool_result rate; observability/grand-dashboard/README.md explicitly selects NativeStack2604 Loki21300 and warns that 13100 can reach the other distribution. Existing bounded urllib/CLI/projection interfaces are reused; no dependency or exporter is added.
• Loki v3.7.8, commit 09e6ce2ff1bdc19763a10265b870c86f51c98655: numeric instant API, metric queries, native log rate. The native operator counts records over its window; it is distinct from Prometheus counter startup/extrapolation behavior.
• Grafana v13.2.3, peeled commit 6193dc03311b631b9727b560d24369e683dc396e: Explore URL contract, native serializer, state types: panes, datasource-specific queries, schemaVersion=1 and epoch-millisecond ranges.
• vLLM reference v0.25.0, commit 702f4814fe54fabff350d43cb753ae3e47c0c276: model-list handler. This is a protocol reference, not a claim about the deployed server version or model identity.
• Installed CPython3.13.16, source cbc944f4bc59639a444dd971c737788ba2283a91: HTTPException hierarchy. HTTPException is separate from OSError.
• Existing native evidence-manifest writer and repository private-name hooks; Gitleaks v8.30.1, explicit PR history range only.

Evidence-class table

Local commands run

python3 -B -m unittest discover -s tests -p 'test_local_pages*.py'
exit 0; 233 tests passed in 16.926s

python3 scripts/evidence_manifest.py --write
exit 0; 11294 registered files

python3 scripts/validate.py
FULL exit 0; 70 components, 11294 hashes, 239 receipts, 4 profiles

gitleaks git --redact --no-banner --config .gitleaks.toml --log-opts=a7bb4a51eea9cc321dd06dd6243a5b2654a5f805..HEAD .
exit 0; one commit, about 36.56KB, no leaks

python3 tools/local-pages/build_pages.py --root <task-checkout> --state-root <native-state> --output-dir <isolated-site> --receipt <native-receipt>
exit 0 at e1ef838a

python3 host-proof.py --root <task-checkout> --site <isolated-site> --receipt <native-receipt>
exit 0; receipt SHA-256 e0dd7393583aa6482af5d57a161a4949c698b340892fdb48f719210bdf7e4ea3

Decision record

Native Fleet tracking observations. Existing native interfaces are adopted rather than adding a client/exporter or guessing counters. The source window and evaluation time do not claim last-event freshness or complete invocation coverage. A reproduced counterexample against the actual transport/render regressions would overturn the selected seam.

Host evidence

No evidence/hosts or platform-status changes. Proof uses isolated page outputs; no live site, provider/model readiness or platform gate is promoted. Raw log streams, arbitrary labels, model-list bodies, client configuration and credentials are excluded.

Checklist

• One commit in the co-op-provided fresh worktree.
• Required SOTA sources named and pinned.
• Fail-before and controlled broken-behavior evidence retained.
• Full local-pages suite and FULL validator pass.
• No new dependency, paid hosting or billing surface.
• Gitleaks history scan limited to the explicit actual merge-base range.
• Peer-owned files and frozen task trees preserved.
• CC deep read, new-head CI and landing remain pending.