xonecas.io
7814
Ta

tiny-c-agent / commit bdc2437

chore: remove old docs

xonecas2026-10-03 bdc2437aac7eea1b23f70221d9bfbd8fe21d2d78 parent 3865700

1 file changed

initial.md+0−104
@@ -1,104 +0,0 @@
1-# tiny-c-agent (`tiny` for short)
2-
3-A fast, minimal and config-first agent for your terminal.
4-
5-## Config
6-
7-The config lives at the user's configured config folder (usually ~/.config).
8-
9-A custom config can be provided via `--config/-c [path to file]` as a CLI argument.
10-If a custom config was specified, and there is a global config, they are merged with the local
11-config overriding the global.
12-
13-Config schema:
14-
15-```jsonc
16-{
17- "theme": "tokyo-night", // string | undefined, the color palette to use.
18- "provider": "", // string | undefined
19- "model": "", // string | undefined
20- "effort": "", // string| undefined; off, minimal, low, medium, high, xhigh, max, depends on model, needs to be clamped on model selection. Default is we don't send one, let the provider decide.
21- "systemPrompt": "", // path | undefined, the path to a markdown or text file to use as system prompt
22- "discoverAgentFiles": true, // boolean | true, if we should discover and append AGENTS.md to the prompt
23- "skillsDirs": [], // [path, ...] | [], list of directories to search for SKILL.md files.
24- "sessionsDir": "", // path | undefined, defaults to .local/state or the appropriate Unix convention,
25- "customProviders": [], // see right below.
26-}
27-```
28-
29-Custom providers example:
30-
31-```json
32-{
33- "customProviders": [
34- {
35- "id": "ollama",
36- "name": "Ollama",
37- "baseUrl": "http://localhost:11434/v1",
38- "api": "openai-completions",
39- "models": ["qwen3-coder:30b", "gpt-oss:20b"]
40- },
41- // ...
42- ]
43-}
44-```
45-
46-### Extra details
47-
48-**Theme**: It can be a choice of `tokyo-night` or `auto`, defaults to tokyo. Auto is a theme that uses the host terminal theme. Plenty of CLIs do this, it shouldn't be a complicated thing. It should be just as detailed and accurate as tokyo-night.
49-**provider**: The provider id you want to use from one of the supported providers. Initially this is opencode-go and opencode-zen via an OPENCODE_API_KEY from the env. Or the id of one of your configured customProviders.
50-**model**: the model id for the specific model you want to use from the configured provider.
51-**effort**: pre-clamp, user configured value for effort. (models vary a lot on this, so the effective effort post clamp is shown in the TUI). This configures the desire of the user, not the effective value.
52-**custom providers**: api is openai-completions, openai-responses, anthropic-messages, or google-generative-ai. Then point provider and model at it.
53-
54-## Agent core
55-
56-Stays very similar in concept to mini-coder's with the exception of the tools.
57-
58-The agent core needs to be interruptible, a polite request to pause at the next boundary for steering.
59-It also needs to be cancelable, and this needs to be responsive.
60-
61-### Tools
62-
63-A single, codemode style tool. https://blog.cloudflare.com/code-mode/ similar in concept to this. Instead of providing several tools, we provide a single code runtime, where we can expose configurable mcps, along with a runtime and standard library. QuickJS comes to mind as a nice embeddable option. Sandboxing is not a feature for now, but an embedded runtime would help us go this way in the future.
64-We currently don't plan to support MCP, but this is the foundation we need to do so eventually.
65-We render the call, and an elided body just like mini-coder does it, and we show the diff of the changes just like mini-coder does with bash calls.
66-
67-- Capabilities: bash exec, read/write to file system, read images. (same as mini-coder but in code mode).
68-
69-## Providers
70-
71-We aim to support Opencode Zen, Opencode Go and local models via customProviders.
72-
73-Instead of supporting all specific providers, we implement 4 modules, one for each major API format. This allows us to support the whole opencode catalog, without a specific implementation per model.
74-Model catalog: we should generate this from a reliable source so we have details about models, like context size, supported effort levels, etc.
75-
76-We can look at the open-source pi-ai for reference.
77-
78-## Headless mode
79-
80-Prints the ongoing statuses to stderr, and tool activity (no tool bodies), and prints the final assistant response to stdout.
81-Can be used in script scenarios, pipes and input redirect, etc.
82-Accepts a `-p/--print` CLI argument as well.
83-
84-## TUI
85-
86-We are going to use the mini-coder UX/Appearance as a specification for how we render reasoning, tool bodies, user and assistant responses, errors, etc.
87-
88-At startup as part of the banner, we should show the loaded agents file if any. Same for skills. So the user knows whats up right away.
89-
90-Same commands as mini-coder, but with the addition of `/new` to start a new session.
91-
92-### Themes
93-
94-Our themes appear both in the chrome of the app, as well as syntax highlighting and diff rendering (see how
95-mini-coder does it). We use tree-sitter and their respective grammars. Then we follow `nvim`'s usage of it.
96-We intend to have this small set of themes:
97-
98-- Tokyo Night (night variant).
99-- Auto (tries to use the host terminal's theme).
100-
101-## Tech
102-
103-Modern C11
104-tree-sitter and its grammars (nvim is a good reference I think, open source and also uses tree sitter).