itongxiaojun/dsh-openrouter-free-models

dsh-openrouter-free-models

A DSH (DeepSeek Harness) plugin that keeps your OpenRouter model list current and honest:

1. Filters OpenRouter's catalog down to the free models.
2. Measures each one — real time-to-first-token and tokens/second, plus declared context size.
3. Publishes the survivors into the harness OpenRouter model list, with speed and context written into the display name:

MiniMax: MiniMax M3 (free) · ⚡48 tok/s · 1M
inclusionAI: Ling 3.0 Flash Fin (free) · ⚡61 tok/s · 256K

Design documents: docs/PRD.md · docs/CheckList.md

---

One important fact up front

"Free" on OpenRouter means zero-priced, not anonymous. The catalog endpoint is public, but every inference call — including :free models — requires an API key. Without one this plugin still lists and sizes the models; it simply cannot measure their speed, and it says so instead of inventing a number.

Requirements

• DSH with the llm-pi-ai adapter mounted (every standard profile has it).
• An OPENROUTER_API_KEY in the harness credential store or the environment.
• Node 20+ (for the standalone CLI).

Install

First: find the harness home the app actually boots

DSH Desktop does not use ~/.dsh; it sets its own harness home. Installing into the wrong one produces a plugin that never loads. Read it from the log:

grep 'DSH_HOME=' ~/Library/Logs/DSH\ Desktop/harness.log | tail -1
# [stdout] [harness-node] DSH_HOME=/Users/<you>/Library/Application Support/dsh-desktop/harness

The profile name is on the [desktop] profile <name> line just above. The examples below use $DSH_HOME/profiles/web.

Install

dsh plugin --profile web add github:itongxiaojun/dsh-openrouter-free-models

That is the whole install. The package declares dsh.bundle.patch, so dsh adds it to dsh.profile.bundles and its cordis.patch.yml becomes a profile layer — the plugin activates on the next start with no hand-edited entry.

Verified:

dependencies:
+ dsh-openrouter-free-models github:itongxiaojun/dsh-openrouter-free-models
dsh.profile.bundles: ["@deepseek-ai/dsh-base", "dsh-openrouter-free-models"]

It is also listed on the plugin market (设置 → 插件市场), once awesome-dsh-plugin#5625 is merged — the market serves its catalog from that list.

Migrating from a manual install? If you previously inserted id: openrouter-free-models into your profile's own cordis.patch.yml by hand, remove that entry first. The bundle patch inserts the same id, and two entries with one id is a conflict.

A hand-rolled install also had to be a bare package name, because the harness discovers a plugin's browser half by resolving exports["./client"] from the package named in the loader entry, and client-modules only recognises exact package specifiers. A relative .mjs path loads the host half and silently drops the Settings section.

Restart, then confirm

Quit and reopen DSH Desktop. The harness log must show both lines:

openrouter-free-models: mounted (tool=openrouter_free_models)
openrouter-free-models: settings bridge mounted on /api/dsh-openrouter-free-models

grep openrouter-free-models ~/Library/Logs/DSH\ Desktop/harness.log | tail

Nothing after a restart means the entry did not resolve — check the path above and that the symlink target exists.

Use

From the agent

One tool is registered:

"Update my OpenRouter free models, then show me which is fastest."

Useful arguments: policy (either | suffix | zero-price), requireTools, textOnly, excludeIds, concurrency, timeoutMs, maxTokens, samples, nameTemplate, reuseCached, maxAgeMinutes, dryRun, keepUnknown.

From the shell

The CLI runs the same pipeline without a harness:

node lib/cli.mjs list                          # what is free, right now
node lib/cli.mjs probe --limit 8               # measure the first eight
node lib/cli.mjs apply                         # dry run: show the planned write
node lib/cli.mjs rollback --yes                # undo it (CLI writes a plain list)

Settings UI

Once installed as a package, the plugin adds a section to the Settings dialog (设置 → 免费模型), listed after the built-in sections.

It gives you:

• Every tunable in one form — free-model policy, tool/text filters, excluded ids, probe concurrency, timeouts, sample count, display-name template, cache window, and the credential reference. Invalid values are repaired to their defaults on save, and the response names what was repaired.
• Manual runs — 列出免费模型, 测速（不写入）, 测速并写入, and 回滚上次写入, the same four actions the agent tool exposes.
• The last result as a table — model name (with speed and context baked in), tokens/second, context size, and probe status.

The browser half is lib/client.js, a prebuilt bundle in the harness's ModuleLoader format. It holds no state: it reads and writes through the loopback bridge below, so the section and the agent tool always agree.

The config document

The UI edits the plugin's own configuration file, not a settings namespace:

$DSH_HOME/storages/openrouter-free-models.config.json

It is deliberately separate. A harness settings namespace requires a schemastery schema, which would make this plugin depend on resolving @deepseek-ai/schemastery from wherever it happens to be installed; keeping its own document preserves the zero-dependency guarantee. The trade-off is that these values do not appear in settings.yaml — see lib/config.mjs for the complete field list and defaults.

The bridge

lib/client.js talks to /api/dsh-openrouter-free-models/* on the harness web server. Every route is POST-only and loopback-only: the bridge mutates harness state and spends API credits, so it must not be reachable off-machine. Malformed bodies are a 400 and a handler throw is a JSON 500, so the browser form always receives a parseable answer.

How the model list is decided

• suffix — id ends with :free.
• zero-price — both prompt and completion priced 0.
• either (default) — the union. As of this writing that is 24 models from a 446-model catalog: 21 with the suffix, plus openrouter/free and two zero-priced audio models. The audio models are then dropped by the default textOnly filter.

By default only models advertising the tools parameter are kept, because agent use needs function calling.

What gets written

Into llm-pi-ai.providers.openrouter.models, one entry per model:

- { id: "minimax/minimax-m3:free", name: "MiniMax: MiniMax M3 (free) · ⚡48 tok/s · 1M", contextWindow: 1048576 }

Order: fastest first

The free models are written sorted by measured speed, fastest to slowest, so the picker opens on the models most likely to feel responsive.

Only a trustworthy throughput figure earns a rank. Models that were never measured, or whose single sample was too short to time (tps: null), cannot be placed on that scale and follow the measured ones in catalog order — putting them first would advertise a speed nobody observed. The sort is stable, so equal speeds keep their catalog order instead of shuffling between runs.

Your own entries keep their relative order behind the ranked free models. The ordering is applied at write time from the report's measurements, so a report cached before this behaviour existed still produces a correctly ordered list.

Writing is subject to four safeguards:

1. Merge, never clobber. Ranked free models lead; your own entries and any extra fields you set by hand (maxTokens, input, compat) survive.
2. Validate before touching anything. A report the adapter would reject is refused whole, and settings are left exactly as they were.
3. Back up first. The previous list lands in ~/.dsh/storages/openrouter-free-models.backup.json before each write.
4. Revision-checked. The write carries the settings revision it read, so a concurrent edit from the settings page is a conflict, not a silent loss.

Name template

Default: {name} · {speed}{ctx}. Placeholders: {name} {speed} {tps} {ttft} {ctx} {id}. A model that was not measured simply omits its speed segment.

Writing the model list is done through the harness settings service, by the agent tool or the Settings section. The standalone CLI deliberately refuses to write when the settings file already lists models, because it edits YAML textually and would orphan the child fields (maxTokens, input) on entries it cannot faithfully reproduce.

Probe behaviour

Two things the probe does deliberately, both of which the first live run against a real key proved necessary:

• Thinking is disabled per request (reasoning: { enabled: false }). Reasoning models otherwise spend the whole token budget thinking and never emit visible text, which is indistinguishable from a broken model. Models whose provider ignores the directive are reported distinctly — only reasoning tokens arrived — rather than as an empty response.
• The probe asks for a counted list, not a single word. A one-word answer measures first-token latency and nothing else: every model would report tps: null, so the speed column would always be empty. The default budget is 64 tokens, and 探测输出上限 is clamped to a minimum of 8.

A model whose reply is too short to time is marked low-confidence and excluded from the fastest/slowest ranking — reporting it as "slowest" would state a finding that was never measured.

Free-tier rate limiting

Free models are rate limited, and it shows. Probing all 20 at concurrency 4 drew HTTP 429 on 7 of them; at concurrency 2 it drew 4. Speed readings also vary a lot between runs for the same model (one model measured 127 tok/s, then 55).

Treat a failed probe as "try again later", not as a verdict, and lower 并发探测数 if you are seeing many 429s.

Speed numbers: read them with care

Free tiers are rate limited, so absolute figures move around. The value here is relative: it tells you that one free model answers three times faster than another on this machine, right now. A reply too short or too fast to time is marked low-confidence rather than reported as a speed.

Files

Development

npm test                # 93 tests, no network required
node lib/cli.mjs list   # the one command that does hit the network

The suite includes an integration test that feeds the plugin's output through the real @deepseek-ai/dsh-llm-pi-ai configuration schema, so the written document is checked against the adapter that will actually consume it. That test skips cleanly where the harness is absent.

Troubleshooting

License

MIT