[cu1ch3n]

What and why

Papilot rejected PDFs above 180,000 characters, restricted AI review to three clients, and capped or truncated instructions, skill content, comments and findings. Reviews now use any connected agent with its normal tools and approval flow; the agent chooses the scope, organization and amount of feedback.

• Remove review input/output length limits, annotation count limits, and silent truncation throughout the composer, API, saved notes and review engine.
• Remove the PDF viewing page quota and preserve complete context sent to agent helpers.
• Send Antigravity prompts through stdin and Claude system instructions through temporary files, avoiding OS command-line limits. Process launch failures fail the task without terminating Papilot.
• Persist compact request fingerprints and IDs for both new tasks and follow-up messages, keeping multi-megabyte retries stable across reloads and preserving pending IDs from older versions.
• Keep tool-permission diagnostics separate from the final review JSON so valid findings are retained.
• Keep annotation shape and PDF anchor validation, task isolation, and the author's explicit choice to apply suggestions.

How it was checked

• npm run typecheck
• npm test: 647 passed, 1 skipped.
• npm run i18n -- check and npm run i18n:check-app
• Main and review-engine production builds.
• npm test -- --run tests/ts/review.test.ts tests/ts/view.test.ts in engine/: 22 passed.
• node scripts/check-ai-review.mjs: browser review flow, including a 3-million-character request, durable retry ID, and retry after reload.
• Regression cases cover a 3,000-line PDF above the old size limit, 600 findings/notes, instructions above 1 MB, long skills/comments/quotes, and viewing 10 pages.
• Both CLI adapters tested with multi-megabyte inputs through their real default process launchers and local fixture executables; startup failures and temporary-file cleanup covered.
• Independent agent reviews verified adapter transport, translations and Claude cancellation cleanup. Follow-up review verified compact retry records, legacy migration and synchronous hashing for LAN HTTP; no outstanding findings.

The complete UI suite and live provider runs have not been performed.