大肥鱼少吃点我的token……节省一下！（DeepSeek Harness Token Saver）
⚠️ 维护状态 / Maintenance Status
这是一个个人实验项目，由 AI 协助整理发布。作者是编程新手，没有时间或能力持续维护。
代码按“原样”提供，不保证兼容未来版本，不提供技术支持。请自行阅读代码、评估风险后使用。
This is a personal experimental project, published with AI assistance. The author is a beginner and cannot provide ongoing maintenance or support. The code is provided "as is", without warranty of any kind. Use at your own risk.
不懂命令行、想直接装上用？ 把这份文件整个复制给 AI 助手，让它照着装：
docs/install-with-ai.zh.md
这份英文 README 是给英文读者的；中国人请从上面两个链接进。
其余文档（小白版/技术版介绍、架构、实测数据、兼容性、排障）索引在 docs/。
Cut the token bill of a DSH agent session by keeping bulky text out of the transcript.
A DeepSeek Harness plugin plus the measured configuration and the measurement tools that go with it. No build step, no harness-internal imports, two public dependencies.
New here? Read the plain-language introduction or the technical one. To hand installation to an AI assistant, give it docs/install-with-ai.zh.md (Chinese).
The problem is not the model's price
An agent loop does not make one request per conversation. It makes one request per step, and every request carries the whole transcript built so far. From this project's own session logs:
| Measure | Value | Command |
|---|
| Requests in one session | 672 | test/turns-report.mjs |
| Prompt tokens in that session | 287,481,671 | test/cost-report.mjs |
| Median prompt size per request | 252,044 tokens | test/cost-report.mjs |
| Cache hit rate, all sessions | 99.3% | test/cost-report.mjs |
Read that third row again. The median request re-sends a quarter of a million tokens, and almost all of it is history.
The cache does not save you. 99.3% of prompt tokens are cache hits, and the measured bill is still 50.0% cache-hit input — because cached input is billed. A high hit rate means the prefix is stable, not that replay is free.
So the lever is not "find a cheaper model". It is reduce what gets re-sent:
cost ≈ (average context size) × (number of requests)
Both factors are under your control, and this project attacks both.
What it actually does, with the measured effect
| Mechanism | Measured effect | Where it lives |
|---|
| Spill oversized tool output | The shipped 50 KB cap fired on 1 of 1,918 results (2.0% of tool text). At 12 KB it captures 39 results carrying 33.0%. A verified 60,562-byte output put ~11,600 bytes into the transcript. | configuration |
| Delegate bulk text to a free model | A verified delegation turned a 14,366-character payload into a 122-character answer — and the answer is what gets re-sent from then on. | this plugin |
| Rotate when a free model throttles | The free per-minute allowance is per model, not per account (verified). A throttled model is side-stepped in about a second instead of waited out for a minute. | this plugin |
| Measure the bill from the logs | Four scripts report where the tokens actually went, so the next decision is evidence-based. | this repo |
One more number, because it is the cheapest win of all: measured requests per human turn ranged from 1.0 to 84.0. One session here spent 84 requests on a single user turn. Fewer, better-aimed tool calls cost nothing to adopt.
Honest accounting
- The biggest single saving is a configuration value, not this plugin. Lowering
maxInlineBytes from the shipped 50,000 to 12,000 is worth more than everything the plugin does. If you only do one thing, do that.
- Routing session titles off the paid route is a small win, configured and not oversold. The bill is replay, not titles.
- Compaction was deliberately left on the paid route. Its summariser replays the conversation prefix verbatim, which would be slow on a local model, and a session containing images would fail outright because the local models declare text-only input.
- Cache hits are 50% of the bill and this project does not reduce that share. It reduces the volume being cached. The distinction matters if you are choosing where to spend effort.
- Numbers here come from one machine's logs. The scripts are included so you can check yours before believing any of it. See docs/measurements.md.
Quick start
Requires DSH 0.1.5-rc.2 or later (see compatibility) and Node 22.19+.
# 1. install the plugin into a profile (pnpm writes the profile manifest)
dsh plugin --profile web add ./dsh-token-saver
# 2. apply the configuration that does the real work
# copy ONE template over $DSH_HOME/profiles/web/cordis.patch.yml
# templates/profile-patch.hosted.yml hosted free tier + local floor
# templates/profile-patch.local-only.yml local llama.cpp only
# 3. confirm the composition, then restart DSH
dsh --profile web --dump-config
# 4. check that everything still applies
node tools/doctor.mjs
Step 2 is not optional. Installing the plugin alone will not noticeably change your bill — the spill cap and the free-model routes are what move it.
The plugin loads when the DSH server restarts, not on a page reload: a profile's bundle list is read once at composition time.
Editing the source afterwards: link: versus file:
dsh plugin add <path> installs a link: dependency, which is a directory junction (or symlink) to your checkout. Verified by installing this plugin into a throwaway profile and inspecting the result:
dependencies: { "dsh-plugin-local-offload": "link:C:/…/dsh-token-saver" }
node_modules/dsh-plugin-local-offload → Junction → C:\…\dsh-token-saver
So with the documented install command, editing this checkout needs only a DSH restart — no reinstall. tools/doctor.mjs confirms it by comparing the two paths and reports installed copy is byte-identical to this source checkout.
A file: dependency, by contrast, is a copy. That is what an older hand-rolled installer produced, and with it every source edit needs a reinstall and a restart. If doctor.mjs reports the installed copy as differing, check which spec your profile uses:
grep dsh-plugin-local-offload "$DSH_HOME/profiles/web/package.json"
To switch a file: install to link::
dsh plugin --profile web remove dsh-plugin-local-offload
dsh plugin --profile web add /path/to/dsh-token-saver
dsh plugin add <path> also works with nothing but this checkout on disk — it does not need a registry, and it writes the profile manifest itself, so no BOM can be introduced by hand.
Configuration
The tool is configured from the profile's cordis.patch.yml, so it can be retuned without touching the plugin. Full values and their reasoning are in templates/README.md.
| Field | Default | Meaning |
|---|
baseUrl | http://127.0.0.1:18080 | Origin of the OpenAI-compatible endpoint, without a trailing /v1 |
model | local-qwen3-8b | Model id sent to the endpoint |
apiKey / apiKeyEnv | local-no-key / unset | Literal bearer token, or the name of an environment variable holding it. Prefer apiKeyEnv so no secret sits in configuration |
chain | '' | Ordered rotation, as a JSON array of `"baseUrl |
chainApiKeyEnv | follows apiKeyEnv | Separate credential for chain entries on other hosts |
cooldownSeconds | 65 | How long a rate-limited endpoint is parked. The allowance is per minute |
maxContextTokens | 0 | Context window to assume when the endpoint cannot report one |
maxInputChars | 400000 | Character ceiling on one payload before truncation with a warning |
maxOutputTokens | 1024 | Default output ceiling per call |
timeoutMs | 300000 | Per-call deadline |
temperature | 0.2 | Low, to keep extraction faithful |
extraBody | '' | JSON object merged into every request body, as an escape hatch |
credentialsFile | <DSH_HOME>/.credentials.yaml | Where to read a saved credential when the environment variable is not set. Empty disables the lookup |
maxInputChars is a coarse guard; the effective ceiling is derived from the endpoint's real token window (75% of it, minus the output budget), capped at 7,000 tokens per call so that one delegation cannot spend a whole minute's quota.
Where the API key comes from
Save it once in the DSH Models page, or add it under refs: in <DSH_HOME>/.credentials.yaml:
refs:
SILICONFLOW_API_KEY: sk-...
Then name it in the config with apiKeyEnv: SILICONFLOW_API_KEY.
Do not expect process.env to contain it. The harness resolves apiKeyEnv references on demand for its own routes; it does not export them into the process environment. Verified against the installed harness: the only two writers of process.env are dsh-app-boot, which materializes .env files (a different store), and dsh-http-proxy. This plugin's first release read process.env[apiKeyEnv], sent the placeholder, and got HTTP 401 on every delegation — which is why the plugin now resolves the credential itself, per request, falling back from the environment to credentialsFile and only then to the literal placeholder.
If you inject the key as a real environment variable, that still wins, and setting credentialsFile: '' forbids the file lookup — the right choice in a container.
The tool
local_delegate(task, instruction, text, maxTokens?)
task | Use for |
|---|
summarize | Dense factual digest of a long document or log |
extract | Pull specific fields, identifiers, or findings out of bulk text |
classify | Assign a label from the instruction |
translate | Translate while preserving code and identifiers |
rewrite | Restructure without losing facts |
answer | Answer a question using only the payload as evidence |
code | Analyse code; name files, symbols, and line-level issues |
Every result ends with a line naming the model and its token counts, so the saving is visible rather than assumed. A rotation adds a second line naming the chain state.
Version compatibility
| Harness version | Status |
|---|
0.1.5-rc.2 | Verified — written and measured on it |
master (current checkout) | API verified — plugin API and all four config row ids checked against the official repository |
| later versions | Untested — not known-broken, simply not verified |
0.1.2-alpha.*, 0.1.3-alpha.*, 0.1.5-alpha.* | Not supported — 0.1.2-alpha was an architecture refactor with a plugin-compatibility notice |
Run node tools/doctor.mjs after any upgrade. It needs no dependencies and reports which of the four config rows still exists, whether defineTool's signature still matches, whether the plugin's dependency surface is still clean, and whether ~/.dsh has BOM damage. Exit code is non-zero when something broke.
Failure is designed to be quiet and survivable: if a config row is renamed, that patch becomes a no-op and DSH falls back to shipped behaviour. Nothing crashes. The plugin itself fails loudly (it will not load) if its own API changes.
docs/compatibility.md explains what to check after an upgrade, how to repair each failure mode, and — deliberately — how to rebuild all four mechanisms from scratch if this plugin never works again. That last section is written to outlive this repository.
What this repo does not do
- It does not lower the price per token, and it does not switch your model.
- It does not implement spill. Spill happens inside the harness; the plugin cannot influence it. This repo measures and configures it.
- It does not require a GPU for the hosted configuration. The local
llama.cpp server is optional and is the floor of the chain, not a prerequisite.
- It does not turn a paid model into a free one. The paid model stays primary; the goal is to spend fewer of its tokens.
Repository layout
| Path | Role |
|---|
lib/index.js | Plugin entry: Config schema and apply |
lib/tools.js | The local_delegate tool: task contracts, payload clamping, cross-chain attempts |
lib/rotation.js | EndpointChain: per-endpoint cooldowns, automatic recovery. No harness imports |
lib/http.js | OpenAI-compatible client, reasoning stripping, URL normalisation, local-server self-heal |
lib/embeddings.js, lib/indexer.js, lib/vector-store.js | Semantic-retrieval prototype, measured and parked (no corpus that gets queried repeatedly) |
tools/doctor.mjs | Post-upgrade self-check. Zero dependencies |
tools/gh-tunnel.mjs | Local CONNECT proxy for GitHub when DNS is blocked. See troubleshooting |
templates/ | Profile patch templates, with the reasoning beside each value |
test/ | Measurement and verification scripts |
.github/workflows/ci.yml | Regression net: syntax, credential rules, rotation, encoding, self-check |
docs/ | architecture, measurements, compatibility, troubleshooting, sponsorship |
Design constraints worth knowing
- Two public imports only:
@deepseek-ai/dsh-tools (defineTool) and @deepseek-ai/schemastery. No harness-internal module. node tools/doctor.mjs re-checks this against the shipped source rather than taking it on trust.
- The tool definition matches the official tutorial field for field, so it is not relying on an undocumented seam.
- No build step. Pure ESM JavaScript. This matters for distribution: a TypeScript package installed from a git host arrives without its
lib/ output and fails to load, and pnpm ≥ 10 additionally refuses to run a git dependency's prepare script until the user allowlists it — which is permission to execute the package's code at install time. Prebuilt or tarball distribution sidesteps all of it, and this package needs neither.
- No hardcoded tunables. Every deployment-varying choice is a validated
Config field changeable from cordis.patch.yml, which is the convention the harness's own AGENTS.md requires of plugins.
Tests
There are no third-party dependencies anywhere in this repository — not in the plugin, and not in the tests. Every script below runs with nothing installed.
node tools/doctor.mjs # post-upgrade self-check; runs from anywhere
node tools/doctor.mjs --strict # warnings are fatal too; this is what CI runs
node tools/verify-live.mjs # end-to-end acceptance (see below)
node test/test-credentials.mjs # credential resolution order and edge cases
node test/test-rotation.mjs # rotation unit tests: park, skip, recover, multi-park
node test/test-yaml-min.mjs # the YAML reader, diffed against js-yaml when available
node test/tunnel-resilience.mjs # the GitHub tunnel survives reset storms
node test/check-bom.mjs # BOM / JSON damage scan under $DSH_HOME
node test/cost-report.mjs # billing attribution from your own session logs
node test/profile-report.mjs # tool-output volume and per-request prompt sizes
node test/turns-report.mjs # requests per human turn
node test/cache-report.mjs # per-session cache and compaction counts
node tools/check-links.mjs # every relative documentation link resolves
Setup for the scripts that import the plugin by name
A few scripts deliberately import the plugin by its package name (dsh-plugin-local-offload) so they exercise the same resolution DSH uses, rather than reaching into lib/ by path. Those, and the ones that import @deepseek-ai/dsh-llm-pi-ai, need a node_modules entry. One command creates them by pointing at the harness already on your machine — no install, and no administrator rights:
node tools/setup-dev.mjs # create the links
node tools/setup-dev.mjs --status # report only, change nothing
It links node_modules/@deepseek-ai to the harness's own copy and node_modules/dsh-plugin-local-offload to this checkout. node_modules/ is gitignored, so nothing reaches the repository. If linking is blocked on Windows, the script prints the two mklink /J lines that do the same job without needing a privilege.
Everything else runs without that step. The nine scripts that need it are verify-live.mjs, verify-all.mjs, verify-route.mjs, test-credentials.mjs, test-plugin.mjs, test-defer.mjs, test-race.mjs, test-selfheal.mjs, test-payload-guard.mjs, test-rotation-live.mjs and test-siliconflow.mjs. Each of them prints node tools/setup-dev.mjs when the import fails, so a fresh clone tells you what to do rather than failing with a bare ERR_MODULE_NOT_FOUND.
Why there are two verification scripts
test/verify-all.mjs is the richer one: it validates the route against the harness's own Config schema and compares the installed plugin copy against the source. It imports harness packages by bare specifier, and ES module resolution for a bare specifier is relative to the importing file's directory — never the current working directory. A script living outside the DSH installation's own node_modules tree therefore cannot resolve them at all, and even dynamic import() from the profile directory fails, because the specifiers inside the imported file are still resolved against that file:
$ cd ~/.dsh/profiles/web
$ node <repo>/test/verify-all.mjs
Error [ERR_MODULE_NOT_FOUND]: Cannot find package '@deepseek-ai/dsh-llm-pi-ai'
node tools/setup-dev.mjs fixes that by creating the link. tools/verify-live.mjs was written as the check that needs no setup at all, so there is always a way to verify the plugin on a fresh clone.
Tests that skip rather than fail
test/test-client.mjs and test/verify-route.mjs need the local llama.cpp server, which is deliberately not auto-started (it occupies VRAM while running, and it is only the fallback). When nothing is listening they print SKIP and exit 0, because "the endpoint is absent" and "the endpoint is broken" are different facts and must not look the same. Start it with powershell -File F:\bonsai\start-local-ai.ps1 and they run for real.
.github/workflows/ci.yml runs the offline subset on Node 22.19 and 24, on Linux and Windows, against an empty DSH_HOME so a developer's real harness home can never influence the result and no check can read a real credential by accident. It is named as code rather than linked because, as the note below explains, that file may legitimately be absent from a published copy — and a link to a missing file is worse than a plain name.
CI deliberately does not install the harness. The harness is a moving pre-stable target, and pinning it into CI would produce failures that say nothing about this repository; tools/doctor.mjs is the tool for that question, and it is meant to be run by hand against a real installation. CI also does not run verify-live.mjs, because that would mean putting an API key into CI secrets for a project whose entire premise is that a free tier and a local model are enough.
Why the workflow file may be missing from a published copy. GitHub refuses to let a personal access token create or update anything under .github/workflows/ unless that token carries the workflow scope on top of repo. A push including the file is rejected as a whole, and removing it from the newest commit does not help, because a push sends history. If you clone this repository and .github/workflows/ci.yml is absent, that is why: add the token scope and push it. docs/troubleshooting.md records the exact error text and the fix.
tools/verify-live.mjs is the acceptance run. It exercises the source checkout rather than the profile's installed copy, and it asserts one behavioural property in particular: that an extract delegation returns all four distinct error codes from a synthetic payload. A vague instruction once made an 8B model return two of four and stop, so the exhaustiveness clause in the system prompt is load-bearing and this test keeps it that way.
node tools/verify-live.mjs # live, against the configured endpoint
node tools/verify-live.mjs --offline # registration and contract checks only
Contributing
The most useful contribution is a measurement that contradicts one published here, with the command that produced it. Second most useful: your harness version and doctor.mjs output, which is what keeps the compatibility table honest.
License
MIT.
Sponsorship
Optional and non-commercial. If this saved you money, the tip jar is here — including a checklist of platform rules to verify before publishing a payment code.
