Skip to content

news-breaking.md

Latest commit

 

History

History
500 lines (439 loc) · 24 KB

news-breaking.md

File metadata and controls

500 lines (439 loc) · 24 KB
name News: EU Parliament Breaking News — Unified
description Generates a single PR containing analysis artifacts and the rendered breaking-news article (Stages A → B → C → D → E in one workflow).
strict false
true
schedule workflow_dispatch
cron
0 */6 * * *
inputs
force_generation
description type required default
Force generation even if recent analysis exists
boolean
false
true
permissions
contents issues pull-requests actions discussions security-events
read
read
read
read
read
read
timeout-minutes 60
features
mcp-gateway
true
sandbox
agent mcp
awf
port keepalive-interval
8080
300
imports
.github/agents/news-generation.agent.md
shared/mcp/news-mcp-servers.md
concurrency
group cancel-in-progress
news-breaking
false
runtimes
node
version
25
network
allowed
defaults
github
node
data.europarl.europa.eu
dataservices.imf.org
api.worldbank.org
*.europa.eu
hack23.com
riksdagsmonitor.com
euparliamentmonitor.com
tools
timeout startup-timeout github bash edit web-fetch agentic-workflows cache-memory repo-memory
300
90
toolsets
all
true
true
key retention-days allowed-extensions
news-breaking-${{ github.repository_owner }}
7
.md
.json
.jsonl
.txt
.html
branch-name allowed-extensions max-file-size max-file-count max-patch-size
memory/news-generation
.md
.json
51200
50
51200
safe-outputs
threat-detection max-patch-size allowed-domains create-pull-request dispatch-workflow
continue-on-error
true
10240
github
data.europarl.europa.eu
hack23.com
riksdagsmonitor.com
euparliamentmonitor.com
title-prefix labels draft expires allowed-base-branches max if-no-changes fallback-as-issue auto-close-issue excluded-files
[news]
agentic-news
analysis-data
type:breaking
false
14d
main
1
warn
true
false
**/*.lock
node_modules/**
.github/workflows/*.lock.yml
workflows max
news-translate
1
steps
name uses with
Setup Node.js
actions/setup-node@53b83947a5a98c8d113130e565377fae1a50d02f
node-version
25
name run
Install dependencies
npm ci --prefer-offline --no-audit
name run
Build TypeScript
npm run build
name run
Copy mermaid + chart vendor assets
npm run copy-vendor
post-steps
name if continue-on-error run
Capture agent recovery patch
always()
true
bash scripts/gh-aw-capture-agent-patch.sh
jobs
pat-pr-fallback
name needs if runs-on permissions steps
Host-side PAT PR fallback
agent
always() && needs.agent.result != 'skipped'
ubuntu-latest
contents pull-requests
write
write
name uses with
Checkout repository
actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd
ref token persist-credentials fetch-depth
${{ github.base_ref || github.event.pull_request.base.ref || github.ref_name || github.event.repository.default_branch }}
${{ secrets.COPILOT_MCP_GITHUB_PERSONAL_ACCESS_TOKEN || secrets.GITHUB_TOKEN }}
false
1
name uses with
Download agent artifact
actions/download-artifact@3e5f45b2cfb9172054b4087a40e8e0b5a5461e7c
name path
agent
/tmp/gh-aw/
name env run
Run host-side PAT PR fallback
GH_TOKEN GH_AW_PAT_PR_FALLBACK_TOKEN GH_AW_PAT_FALLBACK_SLUG GH_AW_PAT_FALLBACK_WORKFLOW_NAME GH_AW_PAT_FALLBACK_RUN_URL
${{ secrets.COPILOT_MCP_GITHUB_PERSONAL_ACCESS_TOKEN || secrets.GITHUB_TOKEN }}
${{ secrets.COPILOT_MCP_GITHUB_PERSONAL_ACCESS_TOKEN || secrets.GITHUB_TOKEN }}
breaking
News: EU Parliament Breaking News — Unified
${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
bash scripts/gh-aw-pat-pr-fallback.sh
engine
id model mcp max-continuations
copilot
claude-sonnet-4.6
session-timeout
65m
1

📰 EU Parliament Breaking News — Unified Workflow

You are the Analysis Agent for EU Parliament Monitor. This workflow runs Stages A → B → C → D → E in a single agent session and ships one PR containing both the analysis artifacts and the rendered article(s) for this article type. There is no paired article workflow — Stage D is a deterministic CLI invocation (npm run generate-article), not an agent prose pass.

📚 Required Reading (read in this order, once per run)

  1. .github/prompts/00-scope-and-ground-rules.md — workspace scope, forbidden/allowed edits, neutrality, agent never writes article prose
  2. .github/prompts/08-infrastructure.md — frontmatter, MCP gateway, stable folder layout, --analysis-only flag, npm run generate-article
  3. .github/prompts/01-data-collection.md — Stage A
  4. .github/prompts/07-mcp-reference.md — canonical tool tables
  5. .github/prompts/02-analysis-protocol.md — Stage B (2 passes; §2 re-run merge rule; §3 time budgets)
  6. .github/prompts/03-analysis-completeness-gate.md — Stage C (blocking); §6b resuming a same-day folder
  7. .github/prompts/04-article-generation.md — Stage D (deterministic CLI; metadata/SEO contract; agents do not author prose)
  8. .github/prompts/05-analysis-to-article-contract.md — artifact-to-article contract and read-before-render duties
  9. Article-Generation.md — end-to-end article pipeline reference, UI/UX export contract, and article.md provenance
  10. .github/prompts/06-pr-and-safe-outputs.md — single-PR rule, unified-workflow PR contract
  11. On error → .github/prompts/09-troubleshooting.md

🔖 Workflow Parameters

Parameter Value
ARTICLE_TYPE_SLUG breaking
Family Unified (Stages A → B → C → D → E in one workflow)
Data window today (last 12 h); fallback one-week
Primary feeds get_adopted_texts_feed, get_events_feed, get_procedures_feed, get_meps_feed with timeframe: "today"; fall back to "one-week" per endpoint if empty/error.
Stage A budget ≤ 4–5 min (per article-horizons.ts)
Stage B budget (2 passes) 22–28 min — HARD CEILING per article-horizons.ts (do not exceed the per-slug ceiling on Stage B even if Pass 2 still has shallow sections; force GATE_RESULT=ANALYSIS_ONLY instead)
Stage C budget (gate + optional Pass 3) ≤ 4 min
Stage D budget ≤ 2 min (deterministic)
Stage E budget (commit + single PR) ≤ 2 min
Stage C exit tripwire minute 36 elapsed (long-horizon prospective: 39; long-horizon retrospective: 38; electoral: 42) — the decision threshold for forcing GATE_RESULT=ANALYSIS_ONLY and (if late) skipping Stage D so the run can still reach the PR call. Per-slug stage ceilings live in src/config/article-horizons.ts; the tripwire backstops any per-stage overrun. Note: Stage D + E run after this tripwire, between the Stage C exit and the PR-call deadline.
Hard PR-call deadline minute ≤ 45 elapsed (target ≤ 42) — deadline for the single safe-outputs create_pull_request call. Backed by engine.mcp.session-timeout: 65m (gh-aw v0.71.3+) which keeps the safeoutputs HTTP session alive for the full 60-min cap.
Hard safety cap 60-min timeout-minutes
PR rule Exactly one [news] PR at end of run

⏱️ MCP session lifetime (gh-aw v0.71.3+): This workflow sets engine.mcp.session-timeout: 65m, which keeps the safeoutputs HTTP session on localhost:3001 alive for 65 minutes — outlasting the 60-min timeout-minutes cap with a 5-min margin. The previous ~28–30 min hard TTL (which capped the old 45-min schedule and bit run #24963129839) no longer applies. The Stage C exit tripwire still fires at the slug-specific elapsed-minute mark in src/config/article-horizons.ts so Stage D + E retain enough budget to land the single PR call by minute ≤ 45. See 09-troubleshooting.md §5 for the historical context.

🎯 Article-Type Specifics

  • Long-horizon stage helpers: see .github/prompts/10-horizon-stage-helpers.md for the registry-driven Stage-A/B/C contract.
  • Every document/event reference MUST include its publish date. Items without a recent date are not breaking news.
  • If no new developments exist today, still ship a PR marked gateResult=ANALYSIS_ONLY in manifest.json.history[] — the deterministic renderer will then emit a noop placeholder rather than a full article.
  • Include intelligence/coalition-dynamics.md and intelligence/mcp-reliability-audit.md in the analysis set (per reference-quality-thresholds.json).

🗓️ Date Context + Stable Folder Resolution (MANDATORY — first bash block)

TODAY=$(date -u +%Y-%m-%d)
LAST_WEEK=$(date -u -d '7 days ago' +%Y-%m-%d)
LAST_MONTH=$(date -u -d '30 days ago' +%Y-%m-%d)
RUN_EPOCH=$(date -u +%s)
RUN_ID="breaking-run$$-$RUN_EPOCH"
# Canonical stable folder — no -run<NN> suffix. Repeated runs on the same
# date share this dir and append to manifest.json.history[].
ANALYSIS_DIR=$(scripts/resolve-analysis-dir.sh "$TODAY" breaking)
WORKFLOW_START_EPOCH=$RUN_EPOCH
echo "ARTICLE_TYPE_SLUG=breaking"                  >> "$GITHUB_ENV"
echo "TODAY=$TODAY"                               >> "$GITHUB_ENV"
echo "RUN_ID=$RUN_ID"                             >> "$GITHUB_ENV"
echo "ANALYSIS_DIR=$ANALYSIS_DIR"                 >> "$GITHUB_ENV"
echo "WORKFLOW_START_EPOCH=$WORKFLOW_START_EPOCH" >> "$GITHUB_ENV"

⚠️ DATE GUARD: When passing dateFrom/dateTo to any MCP tool, always derive dates from $TODAY / $LAST_WEEK / $LAST_MONTH. Never hard-code a year.

🔁 Stage Order (absolute)

Stage A · Data Collection (per-slug budget — see article-horizons.ts)
  → Stage B · Analysis (Pass 1 + Pass 2, hard ceiling per article-horizons.ts)
    → Stage C · Completeness Gate (≤ 4 min) — BLOCKING; elapsed-time
      tripwire (per-slug) forces ANALYSIS_ONLY before Stage D
      → Stage D · Article Render (npm run generate-article — deterministic, ≤ 2 min)
        → Stage E · Single PR (≤ 2 min — by minute ≤ 45; exactly once)

Per-slug minute boundaries are derived from src/config/article-horizons.ts (stageBudgets) and surfaced in the Workflow-Parameters table above. The full table of per-family tripwires lives in .github/prompts/02-analysis-protocol.md §3.

Stage A — Data Collection (Ref: 01, 07)

Run the canonical gateway block from 08-infrastructure.md §4. Source scripts/mcp-setup.sh, then scripts/wb-mcp-probe.sh and scripts/imf-mcp-probe.sh. Collect EP feed data first; fall back to direct endpoints on failure. Deep-fetch up to 10 procedures / voting records / meeting decisions into ${ANALYSIS_DIR}/data/. Target ≤ 4 min.

Stage B — Analysis (Ref: 02 §"Re-run improve/extend rule" — never no-op)

If ${ANALYSIS_DIR}/manifest.json already exists with non-empty history[] from a prior run today, the re-run is never a no-op — apply the re-run improve/extend rule from 02-analysis-protocol.md:

  1. Always run npm run prior-run-diff -- "${ANALYSIS_DIR}" (always-on; the helper no longer reads ENABLE_PRIOR_RUN_MERGE). Persist the plan to ${ANALYSIS_DIR}/runs/prior-run-diff.json. The plan classifies every prior artifact as a must-extend target (carryForward[]) with priorLines + extendFloor exposed, or a below-floor target (rewrite[]).
  2. For every carryForward[] entry, extend & deepen the artifact — skip-writes are forbidden. Each must reach its extendFloor (= max(threshold floor, priorLines + 20)) AND add at least one of: a new section, ≥3 new evidence citations, or ≥1 new chart/diagram. Log one line per extended artifact: [EXTEND-FROM-PRIOR: <relativePath> prior=<priorLines>L → new=<newLines>L (+<delta>)].
  3. For every rewrite[] entry, write a stronger version (overwriting the prior file) sized to the catalog floor.
  4. Append a new entry to manifest.json.history[] (written automatically by runAnalysisStage via mergeManifestHistory). On a re-run, manifest.pass2.rewriteCount MUST equal the total artifact count — rewriteCount === 0 on a re-run is a Stage-C hard RED.

Pass 1 (~60% of analysis time): Apply every methodology and template (analysis/methodologies/ + analysis/templates/) to every downloaded data file. Write every mandatory artifact. Populate manifest.json with top-level articleType: breaking and every produced file under files.*.

Pass 2 (~40% of analysis time): Read every file you wrote, end to end. Expand shallow sections, add evidence citations, add 🟢/🟡/🔴 confidence labels, add cross-references. No [AI_ANALYSIS_REQUIRED] markers may remain.

Emit before Stage C:

PREFLIGHT_ATTESTATION: read N/N artifacts from ${ANALYSIS_DIR} (LINES lines, FRAMEWORKS frameworks)

Stage C — Completeness Gate (Ref: 03) — BLOCKING

Read every manifest-listed artifact and compare it with reference-quality-thresholds.json, the artifact catalog, and the IMF/SEO rules in prompts 01, 03, and 04. Emit exactly one gate line:

STAGE_C_GATE: GREEN articleType=${ARTICLE_TYPE_SLUG} artifacts=<N> lines=<L> imf=<pass|not_required>
STAGE_C_GATE: ANALYSIS_ONLY articleType=${ARTICLE_TYPE_SLUG} reason="<why no article render>"
STAGE_C_GATE: RED articleType=${ARTICLE_TYPE_SLUG} missing=<N> short=<N> placeholders=<N>
  • GREEN → set GATE_RESULT=GREEN and proceed to Stage D.
  • RED (first) → run Pass 3 on the named artifacts, re-run Stage C.
  • RED (second) → set GATE_RESULT=ANALYSIS_ONLY, skip full article render, and ship analysis-only in the single PR.

⏱️ Elapsed-Time Tripwire: At the top of every Stage C iteration, compute the elapsed minutes (mirror the safe two-step pattern from news-translate.md — no nested expansions):

NOW_EPOCH=$(date -u +%s)
ELAPSED_MIN=$(( (NOW_EPOCH - WORKFLOW_START_EPOCH) / 60 ))

Look up the slug-specific Stage C exit tripwire in src/config/article-horizons.ts — short/mid prospective & retrospective slugs trip at minute 36, long-horizon prospective slugs at minute 39, long-horizon retrospective slugs at minute 38, electoral-overlay slugs at minute 42. If ELAPSED_MIN ≥ tripwire, immediately set GATE_RESULT=ANALYSIS_ONLY — even if Stage C has just emitted GREEN. Stage D + E need ≥ 4 min of budget before the PR-call deadline at minute ≤ 45. Emit the gate line as a single unbroken record (note the mandatory articleType= field — required by the contract above and by scripts/validate-analysis-completeness.js):

STAGE_C_GATE: ANALYSIS_ONLY articleType=${ARTICLE_TYPE_SLUG} reason="elapsed-time tripwire at minute ${ELAPSED_MIN}; reserve remaining budget for Stage D+E PR-call deadline"

Then skip Pass 3 and all Stage D render attempts and proceed straight to Stage E. Shipping ANALYSIS_ONLY at the tripwire is strictly better than blowing the 60-min timeout-minutes cap. The 65-min engine.mcp.session-timeout (gh-aw v0.71.3+) keeps the safeoutputs HTTP session alive for the full run, but the workflow still hard-caps at 60 minutes. See 09-troubleshooting.md §5 for run #24963129839 (session not found at minute 29 under the old 45-min schedule) — historical context for the new design.

Stage D — Deterministic Article Render (Refs: 04-article-generation + Article-Generation.md)

Agents do not write article prose. Stage D is a deterministic CLI that aggregates the committed analysis/** artifacts into one canonical markdown document, then renders it to localized HTML article(s).

source scripts/mcp-setup.sh
export USE_EP_MCP=true


# Deterministic article rendering from the committed analysis artifacts.
#    Emits news/<TODAY>-breaking.en.md (canonical aggregated markdown) plus
#    news/<TODAY>-breaking-en.html (rendered article). Always
#    regenerates article.md + HTML on every run (never no-op): the renderer
#    overwrites byte-for-byte from the freshly extended/rewritten Stage-B
#    analysis artifacts.
npm run generate-article -- --run "${ANALYSIS_DIR}"

The renderer is bounded to ≤ 2 min on a typical run. If the gate result is ANALYSIS_ONLY, the renderer emits a short placeholder article documenting the missing artifacts rather than a full prose article.

Stage E — Single PR (Ref: 06)

Emit to stdout immediately before the single PR call:

SINGLE_PR_ATTESTATION: about to emit the only PR of this run at elapsed=<N>m with <X> analysis files + <Y> news files staged (gateResult=<GREEN|ANALYSIS_ONLY>)

Then invoke safeoutputs___create_pull_request exactly once with:

  • base: "main"
  • head: "news/${TODAY}-breaking-${RUN_ID}"
  • title: "[news] breaking — ${TODAY} (run ${RUN_ID})" (append (analysis-only) suffix when GATE_RESULT=ANALYSIS_ONLY)
  • body: summarize the completeness-gate result, list produced analysis artifacts and emitted news files, diff vs. the prior same-day run if manifest.json.history[].length > 1. Cite every analysis file the rendered article consumed via the manifest. Do not author article prose in the PR body — that's the renderer's job.

Banned patterns: see 06-pr-and-safe-outputs.md §4. The repo CI lint (scripts/lint-prompts.js) fails the build if any banned pattern appears in this workflow. Never call the single-PR tool more than once per run.

🚫 Never

  • Never invoke the single-PR tool more than once per run.
  • Never author article prose — the deterministic renderer owns Stage D. Agent edits to news/<date>-breaking.en.md after npm run generate-article exits are forbidden.
  • Never edit anything under news/** outside of npm run generate-article.
  • Never dispatch another workflow except via the dispatch-workflow: safe output (news-translate, exactly once after merge).
  • See 00-scope-and-ground-rules.md for the full list.