ChatHaikuCLI
BY ROOTCOMPUTER
GitHub
ChatHaikuCLI is a lightweight Python client for interactive model use. Switch between the whole Rootcomputer lineup by name, connect key-protected providers, download training corpora, run plugin workflows, tune sampling, and collect local preference data.
python chathaiku.py
$ python chathaiku_dev.py
Endpoint: https://chathaiku.com/api/haiku.php
You: /endpoint epos3
Probing Epos 3... ok
You: Explain quantum computing
epos3: Quantum computing uses qubits to represent and process information in ways classical bits cannot.
You: /endpoint silas
karotte requires an API key.
karotte API key (input hidden): ********
Probing Silas... ok
You: /rc list datasets
9 dataset(s) available
The repository ships with a public client for everyday terminal use and a developer client for endpoint testing, sampling control, and preference-data collection.
Uses the default public ChatHaiku endpoint, supports basic commands, and keeps the interface intentionally small.
python chathaiku.pySwitch models by name, manage provider API keys, download training corpora, inspect health, tune sampling, reload plugins, and collect structured SFT/DPO records locally.
python chathaiku_dev.py --server epos3The developer client keeps a registry of every Rootcomputer model plus partner providers. Type a key and it resolves the endpoint, probes it, and switches. Endpoint locations are published as a manifest, so new models appear without updating the client.
/models
Registered models (probing 7 endpoints)...
* haiku Haiku H2 ok params=217,515,008
tanka Tanka 3.5 ok
epos Epos ok
epos2 Epos 2 ok
epos3 Epos 3 ok
koda Koda ok
silas Silas ok [no sampling]
/endpoint epos3
/endpoint silas
/karotte list models
Rootcomputer models use the chathaiku shape with full sampling control.
Partner models such as Silas use the karotte shape, which is stateless
per request and exposes no sampling parameters. The client detects the shape from the
route structure and handles each protocol correctly, so switching between them is a
single command.
/endpoint koda
/endpoint karotte:silas
Haiku H2, Tanka 3.5, Epos, Epos 2, Epos 3, and Koda are registered out of the box. No key required.
/endpoint haiku · /endpoint epos3
A partner model reachable through the same interface. Requires an API key, which the client prompts for and stores securely.
/endpoint silas
Probe every registered endpoint at once and see which are reachable, which model each reports, and their parameter counts.
/models · /rc list models
Move between registered models, a local /api/chat route, or any self-hosted base server without editing the script.
/endpoint epos3 · /endpoint localhost:8000
Adjust temperature, top-p, top-k, max tokens, repetition penalty, and no-repeat n-gram settings during a session.
/temp 0.7 · /top-p 0.9 · /max-new 200
Probe any endpoint without switching to it. Failures report the exact route, HTTP status, and reason instead of a bare error.
/ping silas · /info
Clear context, view history, save transcripts, retry the last generation, or undo the last exchange.
/history · /save FILE · /retry · /undo
Mark good replies and save positive examples for supervised fine-tuning datasets.
/good → data/sft_positive.jsonl
Rewrite poor answers into chosen/rejected preference pairs for alignment experiments.
/bad · /rewrite → data/dpo_pairs.jsonl
Keys are prompted for without echo, stored outside the endpoint config with restricted permissions, and always displayed masked.
/apikey set karotte · /apikey test
List the training corpora published on rootcomputer.dev and download any of them with a progress bar and resume-safe writes.
/rc list datasets · /rc get dataset ID
Model locations are fetched from a published manifest and cached locally, so the lineup can change without a client update.
/endpoints · /endpoints refresh
Clear the terminal the way a shell does, keeping the session header in place while wiping scrollback.
/cls
Drop plugin files into a /plugins folder, reload from the terminal, and extend the developer client without editing the core script.
/plugin · /plugin reload · /plugin help
ChatHaikuCLI can load Python plugin files from a local plugins/ folder.
Plugins can register their own slash commands, inspect or adjust sampling state,
read conversation history, write preference data, and run one-off chat calls
through the active endpoint.
$ python chathaiku_dev.py --plugins-dir plugins
/plugin
autodpo: /autodpo
duel: /duel, /vs
evaluator: /evaluator, /eval
/plugin help duel
/duel run epos koda --max-turns 20
Each plugin defines a Plugin subclass with a name, optional slash commands,
and a handler. The client provides a session context so plugins can work with the current
endpoint, sampling profile, conversation, and local SFT/DPO collectors — and can
resolve and call any registered model, not just the active one.
plugins/my_tool.py
/plugin reload
Regenerates DPO preference data using your current model's actual output. Point it at an
existing dataset of prompt and chosen rewrites, and AutoDPO queries
your model fresh for every prompt, captures the response as the new rejected,
and writes a clean training-ready JSONL — clearing conversation history between every record
to keep pairs context-independent.
--resume to continue after a Ctrl-C or crash. Output is appended live, so partial progress is never lost.autodpo run input.jsonl output.jsonl --resume
Runs multi-choice benchmark tests against the currently selected model endpoint.
Drop JSONL test files into plugins/evaluator/tests/, then run one test
or the full suite directly from ChatHaikuCLI. Evaluator formats each question,
queries the model with deterministic sampling, parses the answer letter, and saves
structured JSON results to plugins/evaluator/results/.
.jsonl benchmark files to the tests folder with no registration step./evaluator run or every available test with /evaluator run-all.category field, results include per-category accuracy summaries.evaluator run example.jsonl --verbose
Puts two model endpoints into conversation with each other. Each model believes it is talking to a user — in reality its "user" is the other model. Useful for cross-model behavioural evaluation, adversarial probing, generating self-play datasets, and watching two small models try to hold a conversation.
family:model pairs, or raw URLs — including across different providers.duel run epos koda --max-turns 20 --a-temp 0.9
The developer client is designed for iterative model review. Use it to compare models side by side, pull training corpora, adjust generation behavior, load optional plugins, save good answers, and convert poor replies into useful preference examples.
Switch by key, or point at a direct chat route or local base URL.
Chat normally while the CLI manages JSON payloads and history.
Tune generation parameters without restarting the process.
Write positive SFT examples and DPO pairs as local JSONL files.
Download training corpora from rootcomputer.dev without leaving the terminal.
ChatHaikuCLI speaks two endpoint shapes. Rootcomputer servers expose POST /api/chat returning a JSON reply field, with an optional /api/health route. Partner endpoints such as Karotte use a stateless shape the client detects automatically from the route structure.
{
"history": [
{ "role": "user", "content": "Hello" }
],
"temperature": 0.85,
"top_p": 0.92,
"max_new_tokens": 200
}{
"reply": "Hello. How can I help?"
}POST {base}/ai/chat
{
"model": "silas",
"message_history": { },
"message": "Hello"
}{
"answer": "Hello! How can I help you today?",
"message_history": { }
}Use the Python standard-library client for the primary CLI experience, or choose the Windows portable build for a no-Python setup. The current release is v1.4.