* Add MCP activation + usage PostHog events and North Star dashboard prompt
- mcp:authorize_success (server) on OAuth consent completion
- mcp:tool_call (server) on every MCP tool invocation, with tool/success/
error_code/client_id and mcp_client vs in_app_agent source
- mcp:consent_viewed / mcp:consent_denied on the OAuth consent page
- mcp:setup_url_copy / mcp:setup_command_copy intent events on /ai
- docs/posthog-north-star-dashboard.md: prompt to configure the PostHog
core-actions dashboard (activation rate, MCP setup funnel, daily MCP users)
* Fix codex review findings: count schema-rejected MCP calls as failures, wrap audit tools
- mcp:tool_call now fires after output validation; schema mismatches report
success:false with MCP_OUTPUT_VALIDATION (the SDK surfaces them to the
client as JSON-RPC errors)
- run_site_audit/get_audit_status/get_audit_issues/get_audit_pages were
registered without instrumentMcpToolHandler, so their usage and failures
were invisible
* Remove dashboard prompt doc
* Site audit P0 (1/3): issue engine, incremental persistence, block detection
Server-side foundation of the P0 feature set from docs/site-audit-pm-research.md:
- Issue engine: shared registry of issue types (severity, explanation,
how-to-fix). Per-page reporters run inside crawl steps; cross-page checks
(duplicate titles/descriptions/content, broken internal links, redirect
chains/loops, orphan pages) run at finalize as SQL over the persisted crawl.
- New audit_links + audit_issues tables, audit_pages columns (depth, content
hash, header signals, fetch class, sitemap flag); audit tables moved to
src/db/{,pg/}audit.schema.ts; migrations 0029 (D1) / 0006 (PG).
- Incremental persistence: pages/links/issues written inside each crawl-batch
step with deterministic row ids + upserts (retry idempotent); slim step
state; robots.txt checkpointed as step state; merged progress steps keep a
10k-page crawl within the Workflows step budget.
- Crawler: manual redirect handling with inline follow of normalization-
equivalent redirects, response header capture (X-Robots-Tag, Link
rel=canonical), BFS depth, sitemap-last seeding, SSRF check on discovered
links, honest 'we were blocked' classification (403/429/cf-mitigated/
challenge).
- MCP: run_site_audit, get_audit_status, get_audit_issues, get_audit_pages;
limitTier resolved via shared AuditService.resolveAuditLimitTier.
- Lighthouse strategies reduced to auto/none (legacy all/manual map on read).
- Self-healing: getStatus reconciles audits whose workflow instance errored/
terminated without reaching mark-failed.
The Issues UI and the badseo.dev e2e fixture site stack on top of this PR.
Deploy notes: run db:migrate:prod (additive); terminate running audits before
deploying — the workflow step structure changed and in-flight instances cannot
replay under the new code (a finalize guard fails them loudly instead of
completing empty).
* Store only internal link edges in audit_links
Both consumers (broken-internal-link and orphan checks) filter on
isInternal; per-page external counts already live on audit_pages.
Dropping external rows cuts stored edges on outbound-heavy sites.
Column stays so P1 external-link checks can re-add rows without a
migration.
* Review fixes: failAudit CAS guard, dedupe hash helpers, cheaper checks
- failAudit only transitions running audits, so the getStatus reconciler
can't flip a just-completed audit to failed when it races finalize
- collapse the duplicate SHA-256 helper into audit/ids.ts
- finalize integrity guard uses a limit-1 existence probe instead of
fetching every page row
- get_audit_status MCP tool no longer reads the audit row twice when an
explicit auditId is given
The stateless MCP server returns JSON on POST (enableJsonResponse) and
pushes no server-initiated messages, so the optional standalone GET SSE
stream serves no purpose. Left enabled, each GET holds an SSE stream open
indefinitely (25s keepalive, no eventStore) and pins a fresh per-request
McpServer (~5MB of tools + Zod schemas); a few dozen concurrent connected
clients exceed the 128MB isolate limit. This was 100% of the /mcp
exceededMemory OOMs (GET only; POST never OOMed).
Return 405 (spec-compliant 'no standalone stream') before building the
server, so GET allocates nothing. Also removes the bulk of the elevated
GET canceled / responseStreamDisconnected outcomes.
* fix(mcp): render full row data in tool text output, not just counts
MCP clients that surface only the text content block (not structuredContent)
saw count/truncated summaries instead of the actual rows. Add a shared table
renderer (server/mcp/table.ts) and use it so every row ships in the text block
across research_keywords, get_ranked_keywords, get_keyword_metrics,
search_local_businesses, get_local_serp_results, get_google_business_questions,
find_serp_competitors, get_backlinks_profile, get_backlinks_overview,
get_domain_keyword_suggestions, get_rank_tracker, get_serp_results, and
get_search_console_performance.
* test(mcp): assert tool text output renders full row tables
Cover the shared table renderer (table.test.ts) and add per-tool assertions
that each tool's text content block contains the actual row data (exact
rendered rows, null->em-dash), guarding against a column wired to the wrong
field. Adds tool-text-output.test.ts for the service-backed tools and text
assertions to the DataForSEO and Search Console tool tests.
* release: v0.0.23
* style(mcp): prettier-format tool table columns
* feat: add paginated get_backlinks_profile MCP tool
Exposes detailed per-link backlink rows over MCP, reusing BacklinksService
paginated row fetching. Closes#36.
* fix(test): use type-narrowing text assertion to satisfy oxlint
expect.stringContaining inside toMatchObject tripped
typescript-eslint(no-unsafe-assignment), breaking ci:check. Match the
.toContain() pattern used elsewhere in the MCP tool tests.
* refactor(mcp): simplify get_backlinks_profile handler + fix cost estimate
- Drop the redundant backlinksRowsPageRequestSchema.parse re-validation in
the handler; the MCP SDK already validates args against inputSchema. Build
the service request straight from args, removing the duplicated defaults
that could silently diverge.
- Preserve the target length cap by adding .max(2048) to the input schema
(previously enforced only via the re-parse).
- Correct the credit estimate in the tool description: measured cost is
~30 credits/page, not the ~200-500 copied from get_backlinks_overview.
- Minor: type the statuses accumulator.
* fix(mcp): correct get_backlinks_overview credit estimate
Measured real cost: summary (~26 credits) + history for domain scope
(~29 credits) = ~50 per domain, ~25 for a single page. The previous
~200-500 figure was a ~5-10x overestimate (same one get_backlinks_profile
inherited).
---------
Co-authored-by: Matt Van Horn <455140+mvanhorn@users.noreply.github.com>
Co-authored-by: Ben Senescu <bensenescu@gmail.com>
* Add OpenSEO skills for keyword and link workflows
* Remove credit scare copy from MCP and skills
* Remove local skills test sandbox
* Format OpenSEO skill docs
* Disclose OpenSEO credit threshold for MCP research
* Tighten OpenSEO skill workflows
