[NielsRogge]

Summary

• Paper-reader turns include the first 80 KiB of stored Markdown. A truncated paper's continuation now points at pwc_read_paper with offset set to its included_bytes, not at the title (the old cursor: "next" restarted reading at the beginning).
• Budget trimming now rewrites each paper's included_bytes to the bytes it kept, so an offset never skips text the model did not see.
• pdf_text papers are excluded from offset continuation, because their text is not the stored Markdown.

Deploy order

Deploy the pwc-cli MCP change first (huggingface/pwc-cli#35). Until it is live, the hosted read_paper rejects offset. This is a chat change, so it is not merged or deployed; it waits for explicit authorization.

Verification

• npx --no-install node --test --import tsx test/paper_context_budget.test.ts: 11 pass, including trimmed papers report the Markdown bytes kept for offset continuation.
• test/paper_pagination.test.ts: the new offset test passes. pagination survives session restore... fails, but it also fails on main (the installed @earendil-works/pi-coding-agent differs from the version the test expects).
• Full pi_worker suite: 108 pass, 20 fail, the same 20 that fail on main (createMcpExtension is missing from the installed pi-coding-agent).
• uv run --env-file backend/keys.env pytest backend/tests/chat/test_paper_context.py: 28 pass.
• Full backend/tests/chat: 9 fail here, 8 of which also fail on main. The ninth, test_turn_job_migration_is_guarded_and_preserves_schema_ownership, fails on origin/main's docs/deployment.md, which this change does not touch.
• ruff check and ruff format pass.

🤖 Generated with Claude Code

---

Note

Medium Risk
Changes paper-reader continuation semantics and depends on a coordinated pwc-cli MCP deploy; misaligned offsets could skip or duplicate paper text until both sides are live.

Overview
Truncated in-context paper text can now be continued correctly via pwc_read_paper: models are instructed to pass offset from each paper’s included_bytes (no cursor on the first call, then cursor: "next"), instead of paging from the start. content_origin papers (e.g. PDF-extracted text) are excluded from offset continuation because their body is not stored Markdown.

Context-budget trimming rewrites each paper’s included_bytes to the raw Markdown byte count actually kept in the prompt (after unescaping HTML entities), and the budget notice documents that contract so offset reads do not skip unseen text.

Paper-reader system notices (Python + MCP research prompt) are aligned with the same workflow. Tests cover trimmed included_bytes, escaped Markdown byte accounting, and offset-then-cursor pagination.

^{Reviewed by Cursor Bugbot for commit c361798. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit cf0b240. Configure here.}