场记 / Continuity

A DeepSeek Harness plugin that gives an agent local image / speech / music / SFX generation and remembers what it made — the same character stays the same character across every call, and a failed generation is never allowed to pass as a success.

Runs locally. Models are lazy-loaded per request and released when idle, so when you are not using it the GPU is untouched — 0.21 GiB resident, measured. You can play a game on the same card.

场记 is the continuity supervisor on a film set. Their entire job is two things: make sure the costume, hair and props match between takes, and catch the mistake on set before it is cut into the film. That is exactly this plugin's job.

Install

uvx --from dsh-continuity continuity-setup

That one command does the whole backend: preflight → build the engines → fetch only the weights this machine can use → start them.

The PyPI distribution is dsh-continuity (the import name stays continuity_mcp). It is not continuity-mcp — that name on PyPI belongs to an unrelated project, so do not uvx continuity-mcp.

To run from source instead: uvx --from git+https://github.com/linxuhao/Deepseek-Continuity continuity-setup

Then wire it into your dsh profile's cordis.patch.yml. The npm bundle is not published yet, so add the row by hand — continuity-setup prints this block, filled in for your machine, when it finishes:

- insert:
    - id: continuity
      name: '@deepseek-ai/dsh-mcp-client'
      config:
        serverName: continuity
        transport: stdio
        command: uvx
        args: ['--from', 'dsh-continuity', 'continuity-mcp']
        env:
          CONTINUITY_STATE_DIR: !!js process.env.CONTINUITY_STATE_DIR ?? ''
          SD_SERVER: !!js process.env.CONTINUITY_SD_SERVER ?? ''
          AUDIO_SERVER: !!js process.env.CONTINUITY_AUDIO_SERVER ?? ''

(the complete row, with every passthrough documented: bundle/cordis.patch.yml)

continuity-setup checks the machine before it downloads anything, and sizes the install to what it finds. Run continuity-setup --check first to see what it would do — that reads hardware and changes nothing:

体检结果:
  GPU     AMD Radeon RX 7800 XT (RADV NAVI32)  (16.0 GiB, 此刻可用 15.8 GiB, DISCRETE_GPU, vulkan device 1)
          未选 AMD Radeon RX 7900 XTX (RADV NAVI31) (24.0 GiB, 此刻可用 1.4 GiB)
          跳过 llvmpipe —— 软件渲染, 不是真显卡
  内存    30.9 GiB
  磁盘    3118.4 GiB 可用 / 需要 30 GiB
  生图    启用
  音频    启用
  抠图默认档  best

Two details in there that exist because the naive version is wrong:

• It skips llvmpipe. The software rasterizer advertises 30.9 GiB of "VRAM" (it is your system RAM) and would win any "pick the biggest card" contest. Everything would then run on the CPU — working, looking completely normal, and unusably slow.
• It picks by free VRAM, gates by total VRAM. On the machine above the 24 GiB card has 1.4 GiB actually free because another process holds it; picking by size would select it and then OOM. But "is this card good enough" is a hardware question, so that one uses the total — otherwise a 16 GiB card would be rejected for having a game open.

Minimum requirements

Audio-only installs (see below) need 20 GiB during install, 9.5 GiB after.

All VRAM/RAM figures on this page are GiB (2³⁰ bytes), which is what rocm-smi and vulkaninfo report. An earlier version of this README labelled them GB; that was wrong and made the headroom look tighter than it is.

Vulkan instead of CUDA is not a preference — it is why this runs at all. ROCm miscomputes VAE decode on this GPU class (ROCm#6633): five decodes of identical input returned five mutually uncorrelated results. Vulkan/RADV compiles SPIR-V at runtime instead of looking up a per-arch kernel table, and is correct and faster here. The side effect is portability across all three vendors.

GPU vendors

I only have AMD cards, so I will not claim more than that. Nothing in the code is AMD-specific — no CUDA, no ROCm, no HIP, no /dev/kfd, no gfx targets — and ggml's Vulkan backend is widely run on NVIDIA. But "widely run" is not "I verified it".

The NVIDIA path is a genuinely different wiring, not just a different card: NVIDIA's Vulkan ICD lives in the host driver and must be injected by nvidia-container-toolkit, with NVIDIA_DRIVER_CAPABILITIES including graphics — the default compute,utility gives you working CUDA and an empty device list in Vulkan. continuity-setup detects NVIDIA, uses the right compose overlay, and tells you the path is unverified. Reports either way are welcome.

Host RAM in detail

Idle is negligible; the peaks are what sizes the machine.

Background removal is the ceiling, and its cost is independent of input size — 256 / 512 / 1024 px all peak at ~6.8 GiB, because BiRefNet runs at a fixed internal resolution.

On 16 GiB everything works. Below 12 GiB, continuity-setup sets the default to quality="fast" (u2netp): peak drops to 1.33 GiB and it runs in 0.6 s instead of 7.2 s. On a typical game sprite the two are hard to tell apart by eye — checked side by side over a magenta backdrop with the edges zoomed. best remains the default where there is room, because the models do differ in principle on fine edges (hair, semi-transparent fringes), but treat fast as a legitimate choice rather than a degraded fallback.

One rule, not a tier list

Jobs are serialized, so at any moment exactly one model is needed. Everything else is released before the job starts. That is the whole VRAM policy. (The one exception is a split deployment: if the image backend is on a different host from the audio one, they are not competing for a card, so nothing is released — freeing local VRAM for a remote job buys nothing and costs a reload.)

It buys a property worth more than a few saved seconds: peak VRAM is a constant 6.80 GiB regardless of what you call, in what order. Measured over an alternating speech→image→speech→image sequence:

Keeping them resident is 16% faster and does not fit an 8 GiB card — and "voice a line, then draw something" is the most ordinary sequence there is. An earlier version of this README quoted 7.84 GiB for that overlap; that came from a lighter sequence I happened to test, and using it as the ceiling was wrong. A cloned voice keeps its reference audio resident too, which is where the rest comes from.

What the reload actually costs: 4.8 s instead of 1.2 s, and only on the first call after switching away. Ten dialogue lines in a row pay it once:

第 1 句 4.63s   之后九句平均 1.19s   十句合计 15.4s

So there is no VRAM tier list, and no 12 GiB threshold. Above 8 GiB every card behaves identically. Below 8 GiB the installer explains why image generation will not fit and asks whether to install the audio half alone — it does not quietly substitute a different product:

  生图    显存不足
          Fake GTX 1060 只有 6.0 GiB, 而生图实测峰值 6.80 GiB, 需要 8 GiB。
          换更小的生图模型省不下这部分 (Q4 与 Q8 峰值相同 6.60 / 6.59), 降分辨率也不行
          —— 瓶颈是那个 8 GiB 不量化的文本编码器。
          音频那半仍然可以装: 铸声/配音/音乐/音效/抠图都能用, 4 GiB 就够。

  ⚠️ 这张卡装不了生图那半。
     只装音频那半 (铸声/配音/音乐/音效/抠图)? [y/N]

The audio-only install is a real product, not a consolation prize: casting voices, dialogue, music, SFX and cutout all work in 4 GiB.

What does not adapt at all: the image model. Quantizing it does not move VRAM — Q4_0 (2.29 GiB of weights) peaks at 6.60 GiB, Q8_0 (4.01 GiB) at 6.59 GiB, identical. Lowering resolution does not help either (512 / 768 / 1024 all peak the same; only time changes). The bottleneck is the 8 GiB unquantized 4B text encoder. So there is no "medium" image tier to offer, only installed or not. (Q4_0 ships anyway — same VRAM, 1.7 GiB less disk.)

Going below 8 GiB for images means changing the text encoder or the model family. That is possible, but it moves identity pinning from native ref_images to IP-Adapter, which is not verified here — and identity pinning is the whole point.

The one thing that does still key off a resource is host RAM, and it is a different resource: below 12 GiB RAM the cutout default drops to quality="fast" (see above).

Zero residency

Measured on an RX 7800 XT with nothing else on the card:

Images are free: the engine streams weights per request and never keeps them resident. Audio is released by an idle timer (AUDIO_IDLE_UNLOAD_S, default 120 s) — not immediately, because someone voicing ten lines in a row should not pay a reload each time. Reload costs nothing measurable: the same TTS request took 3.0 s both cold and warm, because weights are mmap'd and sit in page cache.

Requests are serialized and everything unneeded is released first, so peak = the single largest model, always. The idle timer covers the one case the rule cannot: after the last job there is no next job to trigger a release, so the timer does it. Closing the agent releases the VRAM too — the MCP server unloads on exit rather than leaving the engines holding it.

Two things it actually does

1. Identity survives across calls. Generation backends are stateless: ask for the same character twice and you get two people who merely resemble each other. Measured on Qwen3-TTS as pitch spread across four lines of one character — same voice description, same lines, the only variable being whether a reference was pinned:

Two different voices, two different magnitudes, same direction. Read the ratio, not the headline number — how far a description drifts depends on the description. And treat f0 spread as a proxy, not the verdict: autocorrelation pitch tracking makes octave errors on low gravelly voices (an earlier run of the table above reported 76 Hz where the octave-corrected figure is 29), so the numbers above anchor each line's search range to the reference. The real acceptance test is listening to the audition clip, which is why create_actor hands you one.

What the number cannot show is the part that matters most: the drift is not random.

Under greedy decoding the seed is provably inert — seeds 5 / 99 / 777 produced one identical sha256 — so randomness was fully eliminated, and it still drifted 242 Hz. Identity is a function of the input text, not of the random draw. temperature=0 and top_k=1 cannot fix it. Only pinning to a reference artifact can.

create_actor(name, voice)          -> audition clip; listen before you commit
actor_tts(actor, text)             -> same timbre every line

create_character / create_animal / create_object (name, appearance)
subject_image(subject, scene)      -> same look, new scene / angle / outfit

Identity and wardrobe are separate: pin the face and build, then change clothes in the scene prompt. A reference in an indigo robe, asked for wearing heavy red armor, comes back in armor with the same face.

Already cast your character somewhere else? import_actor and import_subject pin an artifact you supply — a real voice recording, an ElevenLabs clip, a character sheet from another tool — and everything downstream behaves identically. Audio is normalized to 24 kHz mono for you (44.1 kHz stereo in, verified: reference f0 identical, and an imported actor tracks a natively-cast one to 11 Hz).

2. Degenerate output is refused. A backend that miscomputes returns a perfectly well-formed all-zero WAV, or a flat grey PNG, with HTTP 200. Every artifact is checked (image standard deviation, audio RMS, non-finite samples) and the call fails loudly rather than reporting success over garbage. Cutouts additionally get a quality report — mostly transparent, nothing removed, subject shattered into fragments, holes eaten through the subject — each with a specific warning instead of a silent pass.

Plus remove_bg: diffusion models draw "transparent background" as an opaque checkerboard; this turns it into a real RGBA cutout, which sprites require. And gen_sfx, which synthesizes sfxr-style game SFX procedurally — bit-identical for a given seed, milliseconds, no GPU — because a diffusion model is the wrong instrument for a 40 ms coin pickup.

Tools

19 tools. Everything returns absolute local file paths, not URLs — the agent and the engines are on the same machine, so a path can go straight into your game project without a download step, and there is no file server to run or misconfigure.

generate_image and generate_speech exist for one-offs and say so in their own descriptions: they explicitly tell the agent that what they produce will not come back on the next call, and point at the pinning tools for anything recurring.

Limits, and why each one exists

Every number here is a measured failure boundary, not a policy.

Imported audio below 24 kHz is accepted but flagged: upsampling cannot restore the octave that was thrown away, so the clone comes out duller than the file you gave it. That is worth a warning rather than a silent pass — it is the same failure shape as everything else this plugin exists to catch.

Oversized inputs are handled differently by type, on purpose. An image that is too large is resized and the result is reported back to you (原图 2400x1600 → 存为 1024x682) — a scaled picture still depicts the same thing. Reference audio that is too long is rejected, not trimmed: cutting the tail off the audio would leave the transcript describing something the audio no longer says, and that alignment is exactly what the cloning depends on. Trimming it silently would hand you an actor that imported successfully and sounds like someone else.

Bring your own backend (optional)

There are two independent backend URLs, so you can move one capability off-box and keep the other local:

Tell the installer which half is yours and it skips that half entirely — no weights, no engine, no VRAM gate — while the tools stay registered:

continuity-setup --sd-server http://your-box:9020      # 生图你自己供; 本地只装音频
continuity-setup --audio-server http://your-box:9021   # 反过来

That matters more than it sounds: without it, BYO-ing the image half still downloaded 10.1 GiB of image weights and started a local sd-server nobody would ever call, then refused to enable image tools because the local card was too small. On a 5.3 GiB integrated GPU, --sd-server turns "生图 显存不足" into "生图(BYO)" and downloads nothing.

Note the deliberate split from --no-image: that one means "I don't want this capability" (tools unregistered); --sd-server means "I supply this capability" (tools work normally).

Either half also works when set purely at runtime via the two env vars, whether you installed from PyPI or wired the dsh plugin — the cordis row passes them straight through. continuity_status names whichever side is unreachable. gen_sfx needs no backend at all.

Be clear about what "your own backend" means here: the same engine, elsewhere. It is not a provider abstraction. The client speaks sd.cpp's and audio.cpp's specific HTTP shapes, so you cannot point SD_SERVER at an OpenAI-compatible endpoint, a ComfyUI instance, or a bare IP-Adapter server and expect it to work. What it is good for: running the engines on a beefier box, or sharing one backend between several agents. (An earlier version of this README implied any reference-image-capable backend would do. That was never true of the code.)

One constraint if you go remote: identity pinning needs the image backend to accept a reference image. sd.cpp's ref_images is what the code uses; without it there is no pinning, which is the whole point.

Reference audio used to be a second constraint — the engine was handed a filesystem path and opened the file itself, so a remote audio backend meant casting succeeded and every line after it failed. That was not an engine limitation, it was the wrong endpoint: /v1/audio/speech takes the reference inline as base64 (5 MiB cap; a 15 s reference is ~720 KB), exactly the way the image path had always passed ref_images. Both halves are symmetric now and nothing has to share a directory.

Prior art

A survey of the current MCP ecosystem — MiniMax-MCP, openrouter-mcp-multimodal, AtlasCloud, the dsh vision/draw plugins, and four game-asset servers — found voice cloning in several, visual subject pinning in none, and output verification in none.

Layout

bundle/   dsh bundle (npm) — one plugin row; dsh spawns and supervises the MCP server
src/      the MCP server: pinning, guardrails, verification, cutout, VRAM lifecycle
src/continuity_mcp/deploy/   compose + engine Dockerfile + weight manifest

License

MIT