@buyong/pi-refuseless0.1.2

pi-refuseless

English | 한국어

Adds Refuseless to Pi as a model provider. You enter your Refuseless API key in Pi's /login, and /model lists the models Refuseless offers for that key. While you use a Refuseless model, the status line shows your available credits.

Requirements

• Node.js 22.19.0 or later and Pi (tested with Pi 1.4.0).
• A Refuseless API key from the Refuseless console. Keys start with abl_. Requests use your Refuseless credits.

Installation

Install from npm:

pi install npm:@buyong/pi-refuseless

Or install from a checkout of this repository:

pi install /path/to/pi-packages/apps/pi-refuseless

Then restart Pi, or run /reload in a running Pi. If you installed from a checkout, keep it in place: Pi loads a local package from its directory instead of copying it.

To try the package for one session without installing it, start Pi with pi -e /path/to/pi-packages/apps/pi-refuseless.

Log in and choose a model

In Pi:

1. Run /login, choose Sign in with an API key, then Refuseless.
2. Paste your API key and press Enter. Pi shows Saved API key for Refuseless. Credentials saved to <path>, stores the key in its auth.json (~/.pi/agent/auth.json by default), and fetches the model list with the key. If no model was selected before, Pi also shows Saved API key for Refuseless, but no default model is configured for provider "refuseless". Use /model to select a model. Continue with the next step.
3. Run /model and choose a Refuseless model. Press Enter to use it in this session only, or Ctrl+S to also make it the default model for new sessions.

Pi keeps the key across restarts. If no key is saved, Pi uses the REFUSELESS_API_KEY environment variable. Pressing Escape at the key prompt, or Enter without a key, saves nothing and keeps any key you saved before.

If Refuseless rejects the key, Pi still saves it, shows Saved API key for Refuseless, but its model catalog could not be refreshed; using cached models., and keeps the models it had before (none on a first login). Run /login again with the right key.

Check which model a session is on

/model + Enter changes only the current session. A new session (a new Pi start, /new, or another terminal) begins on Pi's default model, which may belong to another provider, and Pi shows no warning. Check the model before you rely on it:

• In /model, the current model has a ✓ and every line shows its provider, as in ✓ glm-5.3-abliterated [refuseless]. The default model is marked · default.
• The status line at the bottom shows the current model and thinking level, as in glm-5.3-abliterated • low. When more than one provider is set up, it also shows the provider, as in (refuseless) glm-5.3-abliterated • low.

To start new sessions on a Refuseless model, select it in /model with Ctrl+S; Pi shows Default model: refuseless/<model id>. To start a single Pi run on it without changing the default model, use pi --model refuseless/<model id>.

Models

The package has no fixed model list. Pi fetches the list from Refuseless with your key after /login, when /model opens and when interactive Pi starts, and shows every model Refuseless returns, abliterated or not. Refuseless marks each model as abliterated (refusal behavior removed from the weights) or not, and cli models shows that mark. Pi saves the last list and uses it when it cannot fetch, for example in pi -p and pi --list-models. To refresh the saved list, open /model once in interactive Pi.

For each model, Pi uses:

• Context: the context length Refuseless reports, or 128,000 tokens if it reports none.
• Thinking levels: the levels Refuseless reports for the model (see below).
• Max output: 131,072 tokens, or the context length if that is smaller.
• Input: text only.
• Cost: not shown. See refuseless.com/pricing and the Refuseless console.

Thinking levels

Refuseless accepts four thinking levels: off, low, high and max. Each model supports some of them and has a default level. Pi offers only the levels a model supports. For example, the GLM models support low, high and max, so Pi does not offer off for them. A model that supports only off has no thinking levels in Pi. cli models shows each model's levels and default level.

Which level a model starts at:

• Picked in a session with no thinking level yet: the model's default level. This covers a new session that started without a model, where Pi sets the thinking level to off, and a switch from a model without thinking levels.
• Picked after a model with thinking levels: Pi keeps the level you had. If the new model does not support it, Pi moves it to the next supported level, for example off to low and xhigh to max.
• A new session that starts on a Refuseless model: Pi's own setting applies: --thinking, then the default thinking level in /settings, then medium. Refuseless has no medium, so Pi moves it to high.

A per-model or default thinking level in /settings, or a level in --models, takes the place of the level in the first two cases. One exception: in a session with no thinking level yet, a setting of off (or of the lowest level, on a model without off) still gives the model's default level.

Press Shift+Tab to change the thinking level of the current model.

Pi sends each level to Refuseless as follows:

Pi level	Level sent to Refuseless
off	off
minimal, low	low
medium, high	high
xhigh, max	max

Check your credits

While a Refuseless model is active in interactive Pi, the status line at the bottom shows the credits a new request can use, as in Refuseless $41.30. Pi reads the balance when a session starts on a Refuseless model, when you pick one in /model, and after each reply. Reading the balance spends no credits.

Run /pi-refuseless to see the whole balance:

Line	Meaning
Available	What a new request can use: the balance minus what requests in progress hold, capped by the key limit. The status line shows this amount.
Balance	Your paid credits, including what requests in progress hold. Free promo credits and plan allowances are not included, so a request on a promo or plan model can still succeed when Available is $0.00.
Reserved	What requests in progress hold right now. The unused part returns to the balance when each request ends.
Key limit	What this key's spend budget, or its workspace's limit, still allows. none when the key has no budget.

To turn the status line off, select Credits in status line in /pi-refuseless and press Enter or Space; do the same to turn it on again. Press Escape to close. Pi applies the change at once and saves it in refuseless.json in its agent directory (~/.pi/agent/refuseless.json by default), so later sessions keep it.

Without a key, or when Refuseless does not answer, the status line shows no credits. /pi-refuseless then shows the reason, for example No Refuseless API key. Run /login and choose Refuseless. Outside interactive Pi, /pi-refuseless reports the balance as a notification, or on stderr in pi -p.

Errors

When Refuseless rejects a request, Pi stops the reply and shows the Refuseless error, for example:

• 401 invalid_key: the key is wrong. Run /login again.
• 402 insufficient_credits: top up your credits. This can also end a reply partway: before a request starts, Refuseless holds its worst-case cost, and when Available is lower, the reply can stop early. /pi-refuseless shows how much is available.
• 429 model_busy or model_warming: try again a little later.

If models.json also defines refuseless

The Pi setup in the Refuseless SDKs guide adds a refuseless provider to ~/.pi/agent/models.json. Its model entries replace this package's models that have the same IDs, including their thinking levels. Remove that refuseless entry when you use this package.

Logout, update and removal

• Run /logout and choose Refuseless to remove the saved Refuseless key. Pi shows Removed stored API key for Refuseless. Environment variables and models.json config are unchanged. REFUSELESS_API_KEY and other providers' keys stay.
• After updating the package or checkout, run /reload or restart Pi.
• To uninstall, run pi remove npm:@buyong/pi-refuseless for an npm installation, or pi remove /path/to/pi-packages/apps/pi-refuseless for a local installation. This keeps the checkout, the saved key and refuseless.json; to remove the key, run /logout first.

Development and verification

Run these from the repository root after pnpm install:

pnpm --filter @buyong/pi-refuseless typecheck
pnpm --filter @buyong/pi-refuseless test:unit
pnpm --filter @buyong/pi-refuseless cli verify

• typecheck and test:unit exit with 0. The tests send requests only to a local mock of the Refuseless API, so they need no key and spend no credits.
• cli verify also needs Pi installed, with pi on your PATH. It runs the tests, checks the package, and has your installed pi fetch the model list from the mock and answer one pi -p request. It ends with All … checks passed. and exit code 0, and prints the directory that holds its reports.

These commands do not change ~/.pi/agent; each uses its own agent directory.

Try it by hand

To try Pi with the package against the local mock, without a Refuseless key:

pnpm --filter @buyong/pi-refuseless cli pi --mock

In Pi, run /login, choose Sign in with an API key and Refuseless, and enter any key that starts with abl_. Then run /model, choose a Refuseless model, and send a message. The mock lists six models from a saved copy of the real Refuseless list, with their thinking levels and default levels, and replies with Mock Refuseless reply from <model>. For the balance, it returns the example from the Refuseless documentation, so the status line shows Refuseless $41.30. A message that contains tool, credits or busy makes the mock call the bash tool, end the reply with insufficient_credits, or return 429.

When you exit Pi, the command prints Run again in this directory: and a command. Run that command to check that the key and the model list survive a restart.

To try the real service with your key (chat requests spend credits):

pnpm --filter @buyong/pi-refuseless cli pi

Log in with your key in Pi. Try a model if you like, or exit right away if you only want to see the model list. To see the models Refuseless returns for that key, run cli models with the directory after --dir in the command printed when Pi exits:

pnpm --filter @buyong/pi-refuseless cli models --dir <directory>

cli models shows each model as Pi registers it, with its thinking levels, default level and abliterated mark, and then each model entry exactly as Refuseless returned it. It sends no chat request. Without --dir, it uses REFUSELESS_API_KEY; with --mock, it shows the mock's list.

To check a whole conversation with real models (chat requests spend credits):

REFUSELESS_API_KEY=<your key> pnpm --filter @buyong/pi-refuseless cli live --model glm-5.3-abliterated --model deepseek-v4.1-flash-abliterated

cli live runs your installed pi, with this package, once for each model you name, all at the same time, and talks to each model for 10 turns. The turns check short answers, a code word recalled later, two bash tool calls, and the thinking levels high, low and off. It prints pass or FAIL for each turn and each model, and at the end the path of a JSON report.

--turns N stops after N turns. --dir <directory> uses the key saved by cli pi instead of REFUSELESS_API_KEY. cli live does not save the key to disk.

cli mock runs the mock on its own and prints each request it receives. cli --help lists every command and option.

Tested environments

• Pi 1.4.0 (installed CLI and automated tests) with Node.js 24.14.0 on macOS.
• The automated tests run against the local mock, which serves the model list Refuseless returned on 2026-10-27.
• On 2026-10-27, cli live passed all 10 turns with glm-5.3-abliterated, glm-5.3-flash-abliterated and deepseek-v4.1-flash-abliterated on the real service. The other models in the list have not been tried on the real service.
• deepseek-v4.1-flash-abliterated returned no thinking text at any level, including high. A direct request to the Refuseless API gave the same result, so Pi shows no thinking for this model.
• The credits display, /pi-refuseless and the abliterated mark were tested only against the local mock. The mock's abliterated marks are set from the model IDs, not taken from the real service.

verification/acceptance.md records the checks and their results.

License

MIT