Trim news-translate per-turn MCP/tool schema to lift the 25M token wall#2828
Merged
Conversation
…n limit (run #26641603577) Co-authored-by: pethers <1726836+pethers@users.noreply.github.com>
…e2e article-class assertion Co-authored-by: pethers <1726836+pethers@users.noreply.github.com>
…o 13 Co-authored-by: pethers <1726836+pethers@users.noreply.github.com>
Copilot created this pull request from a session on behalf of
pethers
May 29, 2026 17:22
View session
github-actions
Bot
added
documentation
Contributor
🏷️ Automatic Labeling SummaryThis PR has been automatically labeled based on the files changed and PR metadata. Applied Labels: documentation,workflow,ci-cd,testing,size-l,news,agentic-workflow Label Categories
For more information, see |
Contributor
🔍 Lighthouse Performance Audit
📥 Download full Lighthouse report Budget Compliance: Performance budgets enforced via |
Contributor
🔍 Lighthouse Performance Audit
📥 Download full Lighthouse report Budget Compliance: Performance budgets enforced via |
Contributor
There was a problem hiding this comment.
Pull request overview
This PR reduces the news-translate workflow’s per-turn MCP/tool schema footprint so translation runs can process larger language batches without hitting the Copilot weighted-token session cap.
Changes:
- Trims unused MCP servers/tools from
news-translateand its compiled lock file. - Adds
max_langsbatching and scopes validator calls to the selected language batch. - Adds regression tests for the trimmed MCP surface and updated worklist contract.
Reviewed changes
Copilot reviewed 5 out of 5 changed files in this pull request and generated 2 comments.
Show a summary per file
| File | Description |
|---|---|
.github/workflows/news-translate.md |
Updates workflow inputs, MCP/tool surface, language batching, and prompt instructions. |
.github/workflows/news-translate.lock.yml |
Recompiled lock file reflecting the reduced MCP/tool surface and new input. |
tests/news-translate-mcp-surface.test.ts |
Adds guard tests to prevent reintroducing heavy unused MCP/tool schemas. |
tests/news-translate-worklist-contract.test.ts |
Extends worklist contract tests for max_langs and scoped validation. |
tests/render-lib.test.ts |
Loosens article wrapper assertion to allow the TOC modifier class. |
Comment on lines
+557
to
+558
| 4. **Title post-processing**: run `npx tsx scripts/postprocess-translated-brief.ts analysis/daily/$DATE/$SUB/executive-brief_*.md` to re-apply the renderer's `cleanArticleTitle` pipeline on every translated H1. This strips locale-specific boilerplate prefixes (`Exekutiv sammanfattning — `, `Zusammenfassung — `, `执行摘要：…`) and trailing date suffixes that the Pass 1 translation may have left in place. The helper only rewrites when the cleaned title differs from the source; it never falls back to a BLUF synthesis. Verify the script reports `✓` or `✏️` for every file (no `✗`). | ||
| 5. **Final validate**: re-run `npx tsx scripts/validate-executive-brief-translations.ts --source <source>` (when available) to confirm the title post-processing in sub-step 4 did not break parity, or fall back to the structural sanity checks listed in TRANSLATION_GUIDE §Acceptance checklist. Fix any failures by re-translating only the offending language (read back only that file). Do not commit a file that fails validation. | ||
| 5. **Final validate**: re-run `npx tsx scripts/validate-executive-brief-translations.ts --source <source> --lang "$TRANSLATION_LANGS"` (when available) to confirm the title post-processing in sub-step 4 did not break parity, or fall back to the structural sanity checks listed in TRANSLATION_GUIDE §Acceptance checklist. Fix any failures by re-translating only the offending language (read back only that file). Do not commit a file that fails validation. |
| 1. **Pass 1 — translate**: Read the source `executive-brief.md` in full **once**. For every language in `TRANSLATION_LANGS`, produce `analysis/daily/$DATE/$SUB/executive-brief_<lang>.md` following the TRANSLATION_GUIDE rules — **write each file with one `edit` tool call per language** (never via `python3`, `bash` heredocs, or shell redirection — see `01-bash-and-shell-safety.md §File creation & overwrite strategy`). Apply the per-language tone register from TRANSLATION_GUIDE **at write time** (so Pass 2 never needs a full re-read just to adjust register), and for `ar`/`he` start the file with `<!-- dir: rtl -->`. Preserve every verbatim block (YAML, HTML comments except `source-sha`, `dok_id` codes, Mermaid DSL bodies, code fences, URLs, file paths, evidence-anchor canonical column values). Translate every always-translate block (prose, headings, list items, table cell text, image alt-text, BLUF, decisions, link text). **Do not** read a freshly written translation back into model context to "confirm" it — the validator in Pass 2 is the authoritative gate. | ||
| 2. **Pass 2 — validate-first & targeted refine**: This is the **primary token-efficiency control**. Run `npx tsx scripts/validate-executive-brief-translations.ts --source <source>` in the runtime shell. That validator already performs every structural-parity check at near-zero model-token cost: heading / table-row / code-fence / Mermaid-block count parity, `dok_id` and URL set equality, banned-English-phrase detection (`Executive Brief`, `Decisions`, `Confidence`, `BLUF`, …), the `<!-- dir: rtl -->` marker for `ar`/`he`, and the `<!-- source-sha: -->` trailer. **Do not read passing translations back into model context** — reading all 13 files back in full is exactly what drove run [#26633644372](https://github.com/Hack23/riksdagsmonitor/actions/runs/26633644372) to `7.6M` effective tokens and a `429 Maximum effective tokens exceeded` abort. Only read back (and re-translate / fix) the specific `executive-brief_<lang>.md` files the validator reports as failing, then re-run the validator until it is clean for the requested languages. | ||
| 2. **Pass 2 — validate-first & targeted refine**: This is the **primary token-efficiency control**. Run `npx tsx scripts/validate-executive-brief-translations.ts --source <source> --lang "$TRANSLATION_LANGS"` in the runtime shell (the `--lang` scope restricts validation to this run's language batch so deferred languages are not reported as missing). That validator already performs every structural-parity check at near-zero model-token cost: heading / table-row / code-fence / Mermaid-block count parity, `dok_id` and URL set equality, banned-English-phrase detection (`Executive Brief`, `Decisions`, `Confidence`, `BLUF`, …), the `<!-- dir: rtl -->` marker for `ar`/`he`, and the `<!-- source-sha: -->` trailer. **Do not read passing translations back into model context** — reading every target file back in full is exactly what drove run [#26633644372](https://github.com/Hack23/riksdagsmonitor/actions/runs/26633644372) to `7.6M` effective tokens and a `429 Maximum effective tokens exceeded` abort. Only read back (and re-translate / fix) the specific `executive-brief_<lang>.md` files the validator reports as failing, then re-run the validator until it is clean for the requested languages. | ||
| 3. **Append the source-revision marker**: compute `SRC_SHA=$(git log -1 --format=%H -- <source>)` in the runtime shell. Use the `edit` tool to add or replace the final `<!-- source-sha: $SRC_SHA -->` line at the end of every translation file (consistent with the file-write contract — no `>>` or `echo`). This is the drift signal future runs use to decide whether to retranslate. |
…N_LANGS batch Addresses review feedback: - source-sha trailer is now only written to files whose language is in $TRANSLATION_LANGS, preventing deferred languages from being marked as current without actual retranslation. - Title post-processing now iterates per-language instead of using executive-brief_*.md glob, avoiding rewriting deferred-language files without subsequent validation. Co-authored-by: pethers <1726836+pethers@users.noreply.github.com>
Contributor
🔍 Lighthouse Performance Audit
📥 Download full Lighthouse report Budget Compliance: Performance budgets enforced via |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
news-translatekept hitting the 25M weighted effective-token Copilot per-session cap (run #26641603577: 13 langs → 27.0M weighted, aborting before its PR shipped). The team had been defending the cap by shrinking work (max_briefs3→1,max_langscapped at 7); this instead lowers the per-unit cost so each run does more.Root cause
Every model turn re-bills the full JSON schema of every declared MCP server and tool. This workflow loaded
github: toolsets:[all], three data-MCP servers (riksdag-regering/scb/world-bankwithallowed:["*"]),web-fetch, andagentic-workflows— yet the translate agent only ever callsbash,edit, andsafeoutputs___create_pull_request. That dead-weight schema, re-sent each turn, is the ~3.5× weighting multiplier (7.67M raw → 27.0M weighted).Changes
.github/workflows/news-translate.md— trimmed the per-turn surface:scbandworld-bankMCP servers.riksdag-regeringfromallowed:["*"](~32 tools) to["get_sync_status"].github: toolsets:[all],web-fetch, andagentic-workflowsfromtools.max_langsdefault 7→13 (input + worklist-step default/clamp) so one run can translate a full source..github/workflows/news-translate.lock.yml— recompiled (gh aw compile, gh-aw v0.77.1); lock shrank ~146 lines. The agent'smcpServersnow holds only a read-only github MCP (context,repos,issues,pull_requests, auto-injected by the compiler for the safe-outputs PR),riksdag-regeringlimited toget_sync_status, andsafeoutputs.tests/news-translate-mcp-surface.test.ts(new) — guards against regression: noscb/world-bankservers, notoolsets:[all], github stays read-only, riksdag narrowed,max_langsdefault 13.tests/news-translate-worklist-contract.test.ts— updated the pinnedMAX_LANGSdefault to 13.Deliberately retained
network.allowed,safe-outputs.allowed-domains, theriksdag-regeringMCP URL, and the02/07prompt imports are contract-locked bytests/network-diagnostics.test.tsand carry zero per-turn token cost (firewall egress / output sanitizer / system prompt — not re-sent each turn). Trimming them would break CI for no token benefit, so the diet targets only the re-billed tool schemas.Effect
The fixed per-turn overhead collapses toward the raw token count, so a full 13-language source fits comfortably under 25M. Wall-clock (Timer A/B, 60 min) and the 100-file safe-outputs cap become the governing limits — the intended state.