Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Current Project Status

Last updated: 2026-11-17 (Content Studio track Y0–Y10 closed, plus the daily productivity & keynote studio suite — native pure-Rust OOXML .pptx writer, 16:9 widescreen Typst .pdf export, art-directed illustration node, 15 Studio catalogue workflows, and Pingora production gateway; see RoadMap.md and TODOs.md)

This is the current authoritative status entrypoint for Yanshi. Historical evaluations, roadmap notes, and TODO queues may explain how the project arrived here, but this document is the shortest maintained summary of what exists now and what remains active.

Summary

Yanshi is a multimodal content-production runtime: sources in (idea / URL / files / recording), deliverables out (report, mind map, audio overview, slides, flashcards, quiz, infographic, data table, video overview, domain engineering diagrams). The contracts are in RFC_CONTENT_STUDIO.md; the evaluation behind the reset is PROJECT_EVALUATION_2026-11-04.md. It is not a coding-agent framework.

Underneath it is a Rust workspace for deterministic DAG workflows, agent-native runtime loops, Skills, MCP tools, RAG, memory, tracing, plugins, distributed worker foundations, and a Web UI run console.

The current architecture is organized into five layers, with a narrow-waist contract kernel (L0) extracted by the P-A track so the runtimes depend only on shared contracts (enforced by cargo xtask check-arch; see docs/RFC_CRATE_ARCHITECTURE.md and docs/ARCHITECTURE.md):

  • L0 contract kernel: yanshi-value (FlowValue), yanshi-graph (the Flow IR), yanshi-store-spi, yanshi-agent-spi, yanshi-async-util, and yanshi-tool (the Tool contract, split out of yanshi-tools in T3.3).
  • L1 execution core (the executor): yanshi-core runs the L0 Flow IR via the FlowExt trait (flow.run()).
  • L2 capability adapters: yanshi-nodes (tool-tier + deterministic renderers including render_slides, pure-Rust pptx_writer, render_video, render_audio, render_mindmap, render_study, render_table, render_svg, render_markdown, render_document, plus web_search, repo_snapshot, pdf_extract, video_ingest), yanshi-nodes-ai (capability-backed node adapters: llm, figure, illustration, asr, tts, text_to_image, image_to_image, image_edit, image_understand, text_to_video, mcp, rag), yanshi-llm, yanshi-tools (builtin tool implementations + OS-sandbox backends), yanshi-tools-ai (TTS / ASR / image / video as Tools), yanshi-mcp, yanshi-rag, yanshi-memory.
  • L3 agent and orchestration: yanshi-agents, yanshi-skills, yanshi-harness, yanshi-config (shared config-first workflow assembly
    • diagnostics, consumed by both the CLI and server), yanshi-cli.
  • L4 operations and productization: yanshi-tracing, yanshi-server, yanshi-db, yanshi-worker, yanshi-ui, plus the standalone Cloudflare Pingora TLS/HTTP2 reverse proxy in crates/gateway.

Implemented Surfaces

  • DAG workflow execution through yanshi-core::Flow (run via the FlowExt trait).
  • Config-first workflow validation and execution through yanshi-cli.
  • Agent-native runtimes through AgentRuntime, ReAct, Plan-Execute, reflection, memory, and supervisor patterns.
  • Dynamic workflow: yanshi_agents::dynamic::compile_plan_to_flow compiles a declarative WorkflowPlan into a parallel Flow of tool calls, and DynamicWorkflowAgent makes the LLM planning call then compiles + executes. Exposed on the CLI as yanshi workflow dynamic --goal ... --model ..., where the LLM-authored plan runs against a restrictive built-in tool sandbox (--allow-path / --allow-domain grant access; shell is never registered), --dry-run prints the plan without executing, and --approve routes every tool call through the Harness approval pipeline.
  • Harness governance shell (yanshi-harness): hooks, interactive approval, sandbox, audit, run limits, background tasks, and the HarnessEvent envelope.
  • Skills through SKILL.md and skill.toml, including tiered [[knowledge]] retrieval (inline files tier or indexed rag tier) with pluggable chunking strategies — fixed-size, sentence, recursive, paragraph, heading, code-AST, and embedding-based semantic chunking (requires OPENAI_API_KEY).
  • Tool abstraction through Tool, ToolRegistry, policy, permissions, and typed output parts.
  • MCP client, server scaffolding, workflow nodes, CLI calls, and Skill tool integration.
  • RAG search/index/eval foundations behind the rag feature.
  • Trace persistence, replay, TUI, redaction, and OpenTelemetry mapping.
  • Server run APIs, event history, SSE streaming, cancellation, and embedded Web UI run console.
  • Subprocess plugin runtime, plugin.toml, workflow plugin nodes, plugin CLI, and marketplace schema support.
  • Distributed scheduler foundation, gRPC worker protocol, worker runtime, and stitched worker trace events.
  • Official offline-first ecosystem samples under examples/ecosystem/.
  • Sandbox and code-execution hardening (the S track, closed 2026-07-24 through 2026-07-27): file+script combination-chain and skill-script integrity fixes; a real Linux 6.12 VM (via Apple’s container CLI) used to compile- and kernel-enforce the Landlock + cgroup v2 backend; OS sandbox now defaults on (security.os_sandbox = true); code_exec — a mandatory, strongly-isolated ContainerBackend tool for running LLM-generated Python, separate from the OS-sandbox tier, with zero network access, hardcoded resource limits, and automatic Harness approval escalation (ToolIdempotency::NonIdempotent). See docs/RFC_LLM_CODE_EXECUTION.md.
  • Long-horizon tasks and retrieval hardening (the L track, closed 2026-07-2x): replan-loop closure for stalled/failed plan steps, task-summary recovery so a resumed session doesn’t re-derive prior progress from scratch, project-level memory persisted across sessions, RAG retrieval strengthening, and a delegation contract (yanshi-agent-spi::delegation
    • aggregation) for sub-agent hand-off with schema-validated answers and conflict-flagged result aggregation.

LLM providers

The full per-provider capability matrix, ProviderRequest contract, ToolChoice modes, ModelCapabilities flags, model families / context windows, and rate-limit handling all live in LLM_PROVIDERS_MATRIX.md. That document is the single source of truth for what each provider supports; entries are verified by yanshi-llm/tests/provider_consistency.rs (offline) and provider_consistency_live.rs (opt-in live).

Stability

The v1 stability inventory lives in:

These documents define stable, beta, experimental, and internal surfaces for Rust traits, manifests, trace schemas, server envelopes, and plugin/marketplace contracts.

Active Work

The short-term execution queue in TODOs.md (maintainer-local, gitignored) is currently empty — every segment through X (2026-10-07 evaluation fixes) and Y0–Y10 (the Content Studio reset specified by RFC_CONTENT_STUDIO.md) has closed:

  • Y0–Y2 — RFC_CONTENT_STUDIO.md, first-class FlowValue::File artifacts with DB registration + /v1/runs/{id}/artifacts streaming download + CLI --output-dir, streaming token usage, modality CostMeter + cost_limit_usd, and DAG checkpoint/resume across CLI and server.
  • Y3 — embedded pure-Rust renderers (render_table, render_svg, render_markdown, render_mindmap, render_study, render_slides + native OOXML pptx_writer), sandboxed media/document renderers (render_audio, render_video, render_document over pandoc + typst), figure node (svg / d2 / tikz / typst / python container with domain checks + compile-error feedback loop), and illustration node (art-directed prompt refinement, 6 visual style presets, dark/light palette harmonization, bounded concurrent SlideDeck / batch generation).
  • Y4–Y6 — 15 Studio catalogue workflows under examples/generators/ (illustrated_slides, tech_slides_video, slides, video_overview, report, infographic, mindmap, circuit_diagram, audio_overview, flashcards, quiz, repo_architecture, backend_project_planner, paper_reader, video_summarizer), examples/applications/content-studio/, persistent notebooks / sources / source_chunks / notebook_artifacts with unified citations and POST /v1/notebooks/{id}/generate, plus inline workflow agent HITL ask_user interrupt/resume in gateway and Studio UI.
  • Y7–Y10 — cost-budget approvals, ToolEffect::Publish, gateway reach ceilings, input precedence fixes, streaming thinking & reply images, Anthropic adaptive thinking + native structured output, and live multimodal provider verification.

Tracks that this product does not need yet (distributed workers, remaining multi-replica gaps, coding-session Harness providers, seccomp refinement, MCP follow-ups) are listed in RoadMap.md § Deferred, each with the condition that reopens it.

The previously closed V (2026-08-05 production-readiness remediation) segment came from findings by six independent parallel sub-agents that each read one architecture layer, cross-checked by the orchestrator running cargo test/clippy/fmt/tree locally:

  • V0 — blocking correctness/security bugs (panics, unsafe defaults, terminal-state bugs, path traversal, an auth gap).
  • V1 — core execution/agent robustness (DAG terminal-state + resume semantics, cancellation races, reliability-stack wiring, LLM retry, memory runtime integration, log normalization).
  • V2 — structured output, token-level streaming, a generic HITL interrupt/resume protocol (ask_user), and agent-loop checkpointing.
  • V3 — fail-closed defaults on non-loopback binds, execution-side sandbox enforcement for DAG shell nodes, two SSRF gaps closed in HttpTool, per-tenant run admission control, supply-chain tightening (Ed25519 marketplace/plugin signatures, bounded MCP stdio reads), and an expr parser recursion-depth limit.
  • V4 — AGENTS.md regenerated from current sources, yanshi-cli’s 2700-line main.rs split into per-domain commands/*/cli.rs modules, cargo-audit wired into CI (11 of 13 pre-existing CVEs resolved via targeted dependency bumps), and completeness follow-ups: an input_mapping whitespace-parsing bug, a Gemini base_url bug, DB-free ~/.yanshi/runs retention, and (as a follow-up once the rest of V4 closed) chunk_strategy = "semantic" now reachable from a skill’s [[knowledge]] manifest entries, backed by yanshi-rag’s embedding-based SemanticChunker.

The ongoing documentation-convergence convention:

  • keep this file as the current status entrypoint;
  • keep RoadMap.md focused on future direction;
  • keep TODOs.md focused on short-term execution;
  • leave historical evaluations marked as historical references.

Historical References