[redmaplewww]

Outcome

Rebuilds AutoResearch as a runnable local single-user project evidence workbench. The default chain is Idea → literature → reading/synthesis → research report → human approval → manuscript versions → review.

What changed

• Adds the complete V2 function/design manual, state objects, API contract, button trace matrix, quality gates, and acceptance matrix.
• Adds stable /api/v1 project, capability, Idea, literature, reading, synthesis, report, manuscript, review, file, knowledge, run, resume, cancel, and SSE event routes while preserving legacy endpoints.
• Adds explicit offline degradation: project/local import/read remain available; model/search actions stop with structured errors and never emit template research output.
• Replaces the prototype monolith and demo fixtures with project-level Next.js routes and working workspaces.
• Adds report approval gate, versioned manuscript editing, deterministic findings, idempotency, source-bound claims, and run ledgers.

Verification

• Python Ruff: pass
• Pytest warnings-as-errors: 32 passed
• Doctor: ok, exactly 5 business agents
• Web lint/typecheck: pass
• Vitest: 4 passed
• Next.js production build: pass
• Browser: real project create, offline Idea persistence/reload, metadata import, knowledge/evidence, project files, capability diagnostics
• Responsive: 1440×900 and 390×844, no page-level horizontal overflow
• Project-to-Act validate: valid, zero issues
• repo-task-sync: valid
• secret pattern scan: no matches

Remaining gate

The 20-paper licensed corpus / 30 gold-question human quality Pilot is intentionally still pending. Software-loop acceptance does not substitute for that research-quality gate.