Skip to content

02-mcp-access.md

Latest commit

 

History

History
89 lines (66 loc) · 12.3 KB

02-mcp-access.md

File metadata and controls

89 lines (66 loc) · 12.3 KB

02 — MCP Access

Authoritative per-workflow surface: the mcp-servers: + tools: blocks in that workflow's frontmatter. .github/copilot-mcp.json is the local Copilot surface (used by assign_copilot_to_issue / agent files in .github/agents/), not by news workflow runs.

Servers & tool naming

News workflows declare three data MCP servers + the built-in github toolset (via tools.github.toolsets: [all]) + bash + edit + web-fetch (frontmatter key; agent calls it as web_fetch) + agentic-workflows + cache-memory (resilience).

Naming convention reminder: gh-aw frontmatter keys use kebab-case (tools.web-fetch:, tools.cache-memory:, tools.agentic-workflows:); the runtime tool names the agent invokes use snake_case (web_fetch, cache_memory, …). The same split applies to safe outputs (safe-outputs.create-pull-request: in YAML → safeoutputs___create_pull_request at call time).

Server / tool Transport Declared in Tool-name style Example tools
riksdag-regering HTTP (Render) workflow mcp-servers: snake_case get_sync_status, search_dokument, get_voteringar, get_dokument_innehall
scb container (@jarib/pxweb-mcp, node:26-alpine) workflow mcp-servers: snake_case search_tables, get_table_info, query_table
world-bank container (worldbank-mcp, node:26-alpine) workflow mcp-servers: kebab-case get-economic-data (despite the legacy name, this method serves the non-economic indicator IDs documented in analysis/worldbank/indicators-inventory.json: governance WGI, environment, social/education residue, defence historicals, crime/justice; route economic context through scripts/imf-fetch.ts), get-country-info, search-indicators
github HTTP (Copilot MCP) workflow tools.github (toolsets: [all]) standard full GitHub MCP toolset (issues, PRs, repos, code-search, actions, releases, discussions, …)
bash local helper workflow tools.bash: true standard shell execution (also hosts the IMF CLI — see § IMF CLI below)
edit local helper workflow tools.edit: standard filesystem edits inside $GITHUB_WORKSPACE
web-fetch local helper workflow tools.web-fetch: standard HTTP fetch for non-MCP public sources (e.g. www.statskontoret.se, riksdagsmonitor.com) — domain-filtered through the AWF firewall. Agent calls this as web_fetch (snake_case runtime name)
cache-memory GitHub Actions cache workflow tools.cache-memory: (filesystem) persistent file storage at /tmp/gh-aw/cache-memory/ keyed by news-${workflow}-${article_date} (14-day retention). Survives across runs and can restore the most recent prior cache via restore-keys when the exact key is not found → resilience for failed-PR retries. See 07-commit-and-pr.md §Cache-memory recovery.
safeoutputs runner (Streamable HTTP) always available snake_case safeoutputs___create_pull_request, safeoutputs___noop, safeoutputs___dispatch_workflow, safeoutputs___add_comment, safeoutputs___missing_data, safeoutputs___missing_tool, safeoutputs___report_incomplete

filesystem, memory, and sequential-thinking are declared in .github/copilot-mcp.json for the local Copilot / assign_copilot_to_issue channel and are not available to news workflows unless the workflow itself declares them under mcp-servers:.

playwright must be treated separately: in news workflows it is available as the built-in workflow tool tools.playwright when that workflow declares it under tools: (e.g. news-evening-analysis, news-realtime-monitor). In that case it is not an MCP server, so do not infer its availability from mcp-servers: alone and do not skip Playwright/browser validation steps when tools.playwright is present in workflow frontmatter.

Authoritative inventory: .github/copilot-mcp.json for the local Copilot MCP surface, and each workflow's mcp-servers: plus tools: frontmatter for the actual per-run surface.

IMF CLI (primary for all economic data — NOT an MCP server)

IMF data is the primary source for all macro / fiscal / monetary / external-sector / trade / COFOG / commodity / exchange-rate context. IMF is not an MCP server — access is via the TypeScript CLI invoked through the bash tool:

npx tsx scripts/imf-fetch.ts <command> [flags]
Command Purpose Example
weo Single-country WEO time series (annual, incl. T+5 projections) tsx scripts/imf-fetch.ts weo --country SWE --indicator NGDP_RPCH --years 15 --persist
compare Batched WEO compare across the Nordic peer set (1 call, N countries) tsx scripts/imf-fetch.ts compare --indicator GGXWDG_NGDP --countries SWE,DNK,NOR,FIN,DEU --persist
sdmx SDMX 3.0 passthrough (CPI / FM / BOP / GFS_COFOG / MFS_IR / IMTS / PCPS / ER — IFS dissolved into CPI/MFS_IR/ER, DOTS renamed to IMTS, PCPS moved to IMF.RES in the 2026-05 refactor) tsx scripts/imf-fetch.ts sdmx --path "/data/IMF.STA,CPI,5.0.0/SWE.CPI._T.IX.M?startPeriod=2024-01" --indicator _T.IX --country SWE --persist
list-indicators Discovery — list built-in indicator codes tsx scripts/imf-fetch.ts list-indicators

Rules:

  • Rate-limit discipline: IMF advertises ~10 req / 5 s. Prefer compare over parallel weo. sleep 1 between invocations. Target ≤ 10 IMF calls per article.
  • Caching: always pass --persist — writes to analysis/data/imf/{indicator}/{country}.json. Re-use across article types in the same daily run.
  • Vintage discipline: every projection quote MUST include the vintage tag — (WEO Apr-2026, GGXWDG_NGDP).
  • Provider decision: macro / fiscal / monetary / external → IMF; governance (WGI) / environment / social residue → World Bank; Swedish-specific ground truth → SCB.
  • Authoritative inventory: analysis/imf/indicators-inventory.json (machine-readable) · analysis/imf/data-dictionary.md (dataflow reference) · analysis/imf/agentic-integration.md (7-step playbook) · analysis/imf/indicator-policy-mapping.md (committee matrix).
  • Contract: .github/aw/ECONOMIC_DATA_CONTRACT.md v3.2+.
  • Firewall egress: www.imf.org, api.imf.org, data.imf.org, dataservices.imf.org (already in every workflow's network.allowed and safe-outputs.allowed-domains). The TypeScript IMF client sends an explicit Riksdagsmonitor User-Agent; do not replace it with raw node / undici fetch calls because IMF Datamapper can reject those with HTTP 403. Every SDMX request to api.imf.org/external/sdmx/3.0 requires the IMF_SDMX_SUBSCRIPTION_KEY (Azure APIM Ocp-Apim-Subscription-Key header) — already exported to the agent shell by news-prewarm.
  • Statskontoret egress: www.statskontoret.se / statskontoret.se are public non-MCP web sources used for agency capacity, state-governance evaluations, implementation feasibility, administrative burden and public-sector efficiency evidence.
  • Lagrådet egress: www.lagradet.se / lagradet.se are public non-MCP web sources for Council on Legislation referrals and yttranden on government propositions touching constitutional law, fundamental rights, criminal procedure, court organisation, surveillance, and taxation principles. Allow-listed in every news workflow's network.allowed. Required input for major-bill risk-assessment.md, threat-analysis.md and forward-indicators.md per 03-data-download.md §Lagrådet enrichment.

Health gate (in-prompt)

Run once at workflow start, then proceed — do not loop forever.

  1. Call get_sync_status({}). Retry up to 3 times, 20 s apart. Server is pre-warmed by the CI steps: block.
  2. If the third attempt fails, set ANALYSIS_DIR=analysis/daily/$ARTICLE_DATE/$SUBFOLDER and branch on whether prior analysis exists, using a single concrete on-disk test:
    • Prior analysis on disk ([ -s "$ANALYSIS_DIR/synthesis-summary.md" ] returns true): do not exit. Route to improvement-mode in 04-analysis-pipeline.md and continue without live MCP — extend artifacts using on-disk evidence, refresh article.md + rendered HTML, and commit one PR.
    • No prior analysis on disk ([ ! -s "$ANALYSIS_DIR/synthesis-summary.md" ]): apply the MCP-unreachable no-op policy from 07-commit-and-pr.md §No-op policy and exit.
  3. Once get_sync_status succeeds, proceed. Do not spend more than 2 minutes on warm-up.
  4. Read data/imf-context.json from the pre-warm action before making economic claims:
    • status: ok / stale-vintage / degraded: continue IMF-first. For degraded, use WEO/FM Datamapper evidence and avoid SDMX-only claims unless cached data exists.
    • status: unavailable or data/imf-unavailable.flag present: inject the standard warning block and use cached IMF / SCB fallback only; never substitute World Bank for macroeconomic claims.
  5. Pre-warm IMF with one throwaway weo call through the CLI (not raw fetch): npx tsx scripts/imf-fetch.ts weo --country SWE --indicator NGDP_RPCH --years 1 >/dev/null 2>&1 || true ; sleep 1.

Data sourcing rules

Rule
Riksdag tool arguments are documented under .github/skills/riksdag-regering-mcp/.
Economic data is IMF-first. Only use get-economic-data (world-bank MCP) for articles written pre-2026-04-20 or as an explicit legacy fallback — never as a primary source in new articles.
Statskontoret is a public non-MCP source for Swedish agency governance, administrative capacity, implementation feasibility and public-sector efficiency. Use web_fetch / primary URLs where available, cite report title + URL, and record retrieval in data-download-manifest.md.
Lagrådet is a public non-MCP source for Council on Legislation referrals and yttranden. Required for major-bill propositions per 03-data-download.md §Lagrådet enrichment. Cite referral URL + yttrande publication date; tag referral pending when no yttrande yet exists.
Prior-voteringar enrichment is a standard call: search_voteringar (riksdag-regering MCP) keyed by committee bet prefix and the last 4 rm (riksmöten) for every committee-report, motion, or interpellation cycle. Feeds historical-parallels.md, coalition-mathematics.md, and swot-analysis.md evidence rows (see 03-data-download.md §Prior-voteringar enrichment).
Treat MCP failure mid-run as partial data: continue with what you have, document gaps in data-download-manifest.md, never silently drop documents.
Source authority and no-fabrication rule: see 00-base-contract.md rules 1 + 3.

Pre-warm step (CI job, not prompt)

Every news workflow declares a single curl-based pre-warm step with ≤ 6 retries, ≤ 20 s apart. With curl --max-time 30, the worst-case runtime can exceed 4 minutes, so this is a best-effort pre-warm rather than a hard ≤ 2 minute guarantee. If a strict 2 minute cap is required, the workflow's curl timeout and/or retry policy must be reduced accordingly. No background pingers.

MCP gateway session timeout (engine.mcp.session-timeout) — DO NOT SET without re-testing on v0.3.9

🚫 Currently removed from every workflow. MCP Gateway v0.3.1 (ghcr.io/github/gh-aw-mcpg:v0.3.1) rejected the gh-aw v0.71.3 compiled sessionTimeout field as additionalProperties 'sessionTimeout' not allowed (gh-aw #29353). The gh-aw v0.74.3 lock files now ship MCP Gateway v0.3.9 — the field's acceptance has not yet been re-validated on this repo. Do not re-add engine.mcp.session-timeout without first running one news workflow end-to-end against v0.3.9 to confirm the gateway accepts it.

The sandbox.mcp.keepalive-interval setting also remains removed — the MCP gateway default keepalive (now on v0.3.9) is sufficient for the 60-min job window. The PR deadline is governed by Timer A (job timeout-minutes: 60, measured from job start) and Timer B (Copilot API session ~60 min). Call the PR safe-output by agent minute 42.