* feat(transport): remote Streamable HTTP transport + demo-playground mode
Adds an HTTP transport for the hosted public demo playground while leaving the
default stdio entrypoint unchanged:
- LOG10X_MCP_HTTP_PORT starts a stateful Streamable HTTP server (one McpServer
per session, keyed by Mcp-Session-Id) at POST/GET/DELETE /mcp, plus a GET
/health liveness probe for the ALB. No auth in this layer — the deployment
fronts it with a TLS-terminating proxy and serves only the read-only demo.
- Refactors server construction into a factory (configureServer) so tool
registration + telemetry wrapping + the manifest registry work per-session;
stdio path is functionally identical.
- LOG10X_MCP_DEMO_PLAYGROUND=1 bypasses the index.ts metric-tool not_configured
gate so pure-demo mode serves the public demo data (isDemoMode stays true).
- Recategorize resolve_batch + extract_templates ('identify' -> 'paste') so the
demo category allowlist excludes the two tools that spawn a local tenx CLI.
- Dockerfile (minimal, non-root, no cloud CLIs) for the Fargate image.
Verified: tsc --noEmit clean; live smoke test confirmed /health, MCP initialize
(session id), tools/list = the 20 allowlisted read-only tools only, and a
metric tool served data.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* feat(http): expose build/version info on /health + boot log
/health now returns JSON {status, version, commit, builtAt} (still HTTP 200, so
the ALB check is unaffected) so anyone can confirm which build is live and
verify a redeploy actually rolled. commit/builtAt are baked into the image via
Docker build-args (GIT_SHA / BUILD_TIME) and default to 'unknown' for local runs.
Also added to the mcp.boot log line.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* chore: bump to 1.12.0
Remote Streamable HTTP transport + demo-playground mode + /health build-info.
Version surfaces via serverInfo and the /health JSON.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(build): list vendor files explicitly (brace expansion breaks on sh/dash)
npm runs scripts via /bin/sh on Linux CI; dash doesn't brace-expand
{promql.js,package.json}, so `cp` got the literal path and the build failed
(breaking build-and-test + campaign-rescore). Pre-existing — same script is on
main. List the two files explicitly (behaviour-identical on bash).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* test: fix 16 pre-existing failures inherited from recent main commits
These tests broke on main when recent feature commits changed behavior/output
without updating the tests; they were masked because the build step failed
first. All are stale-test or test-harness fixes — no production code changed:
- manifest: cost_drivers renamed to cost_options.
- metrics-backend: amp/gcp/grafana_cloud adapters are implemented now (not
phase-1 stubs) — assert the real credential/auth behavior.
- test harness: `test` script now copies vendor/promql-parser into test-build
(the build script already did) so the datadog promql->datadog require resolves.
- next-actions: renderer emits a NEXT_STEPS_FOR_USER block before NEXT_ACTIONS.
- cloudwatch: query format changed to quoted-term FilterLogEvents search.
- pattern-extraction: resolveTenxMode prefers docker when unset — pin local;
DevCliNotInstalledError wording changed ("not available").
- poc-status: executePocStatus returns a typed input_invalid envelope, not throw.
- metrics-that-moved: backend-fetch now retries 5xx — adjust failure rails.
- poc-report: section 6 renamed "Compaction Potential" -> "Compact-byte Ratio (Measured)".
- siem/views: receiver YAML quotes the identity; identity is the raw hash.
Full suite: 851 pass / 0 fail / 5 skipped; tsc clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* test(eval): drop registry refs to removed tools (fix campaign-rescore build)
The eval harness (eval/src/tool-registry.ts) still imported + registered three
tools that recent main commits removed from src/tools/: correlate-cross-pillar,
translate-metric-to-patterns, configure-compact. tsc failed with TS2307 at the
"Build eval harness" step. Removed their imports, handler entries, and schema-map
entries. eval `tsc` is now clean (0 errors). Pre-existing — surfaced once the
build-script fix let campaign-rescore get past "Build MCP".
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* test(conformance): tolerate tenx-absent not_configured for extract_templates
extract_templates always runs the local tenx pipeline; with no tenx on the box
(CI) it returns a well-formed not_configured envelope, not a success one — so
the new envelope-conformance test could never pass on CI (it required the unified
fields, which live on the success path). Branch the assertion: accept the
graceful not_configured envelope as conformant, keep the unified-fields check for
the success path (tenx present). Test-only; no product/envelope-framework change.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* test(eval): anchor the 4 mcp-only/paste hero fixtures (spec-anchor lint)
campaign-rescore's spec-anchor-lint failed: 4 fixtures had an empty must_mention
(no anchor). Populate each, mirroring sibling fixtures + the scenario's explicit
ask: cost-audit→["opentelemetry","service"], env-health→["freshness","service"],
error-investigation→["service_instance_id","ERROR"] (same as error-top-error-pattern),
paste-no-match→["no matching pattern"]. Best-effort anchors matching Tal's style —
worth an eval-owner sanity check for rubric rigor.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Dor Levi <11015255+dor-levi@users.noreply.github.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
