kondi-chat

kondi-chat

The terminal coding agent that picks a different model for each phase.

One coding task. Three models. About four cents. GPT-5.4 plans. Gemini 2.5 Pro codes (free). Sonnet reviews. kondi-chat reads what each phase actually needs and routes accordingly — every turn, under a cost cap you set.

Why route per phase?

Every model has a sweet spot:

• Frontier reasoning (Opus, GPT-5.4) — great at planning and architecture, painful on grunt work.
• Coding-tuned (Gemini 2.5 Pro, DeepSeek V4, GLM-4.6) — produce good code at 1/20th the cost.
• Fast and cheap (GLM-4.5-flash, Haiku) — compress context, run classifiers, summarise, often free.

Pinning to one model means you're either burning money on boilerplate or undercutting hard problems. kondi-chat looks at the current pipeline phase — plan, execute, reflect, compress — and picks from the models you've enabled. It's not a lookup table; it's an LLM intent classifier seeded with each model's description and capabilities, with a learned tier that takes over once you've accumulated enough usage data.

One turn, three models

> refactor src/auth into a separate module with its own tests

  router  plan      → gpt-5.4              ($0.011)
  router  execute   → gemini-2.5-pro       (free)
  router  reflect   → claude-sonnet-4-5    ($0.029)

  ✓ extracted src/auth/{index.ts, session.ts, tokens.ts}
  ✓ moved 14 tests to src/auth/__tests__/
  ✓ all 47 tests green; typecheck clean

  total: 3 models · 8 tool calls · $0.040 · 23s

The same task on Claude Code or Cursor runs Opus or GPT-5 end-to-end. Same outcome; ~30× the bill on the heavy bits.

How it compares

	kondi-chat	Claude Code	Cursor CLI	Aider
Models per turn	many, per-phase	one (Claude)	one (configurable)	one (configurable)
Cross-provider routing	yes	no	no	no
Cost cap enforced in-loop	yes	no	no	partial
Free-tier coding (Gemini / DeepSeek)	yes	no	no	manual
In-terminal scrollback (no alt-screen)	yes	no	no	yes
Local model support	yes	no	no	yes
IDE integration	no	no	yes	no

If you want one polished model and don't care about cost, use Claude Code. If you want the cheapest capable model for each step of every task, kondi-chat is the one that does that.

Install

# Requires Node 18+. The postinstall downloads the prebuilt TUI binary
# for your platform — no Rust toolchain needed.
npm install -g @thispointon/kondi-chat

Platform support

Platform	Status
Linux x64 / arm64	Supported — npm install and binary launch verified in a clean Debian container. Static musl binaries, so they run on any distro.
macOS x64 / arm64	Supported — same install path, native binaries.
Windows	Via WSL only. Native Windows is not supported.

Windows users run kondi-chat inside WSL (Windows Subsystem for Linux). WSL is a Linux environment, so it uses the same Linux binary — there is no separate Windows build to break. One-time setup, in PowerShell:

wsl --install

Then open the WSL / Ubuntu shell and run the npm install -g command above inside it.

From source (for hacking on it):

git clone https://github.com/thisPointOn/kondi-chat.git
cd kondi-chat
npm install --ignore-scripts            # skip postinstall when building locally
cd tui && cargo build --release && cd ..
npm run chat:tui                        # run the TUI

Set up your API keys

kondi-chat talks to whatever providers you have a key for, and skips the rest — you do not need every key. One is enough to start.

The friendliest free path: a Google AI Studio key (free Gemini tier) plus a Z.AI Coding Plan key (free GLM-4.5-flash). Each provider issues keys from its own developer console — Anthropic, OpenAI, Google AI Studio, DeepSeek, xAI, and Z.AI all have one.

You can supply keys two ways:

Option A — a .env file (recommended, persists across runs). Create a file named .env with one KEY=value per line. kondi-chat reads it from three places, checked in order:

1. the project directory you launched kondi-chat from,
2. ~/.kondi-chat/.env — a global file, set your keys once and they work in every project,
3. the kondi-chat install directory.

# ~/.kondi-chat/.env  — set once, used everywhere
ANTHROPIC_API_KEY=sk-ant-...
OPENAI_API_KEY=sk-proj-...
GOOGLE_API_KEY=...
ZAI_API_KEY=...
# DEEPSEEK_API_KEY, XAI_API_KEY, BRAVE_SEARCH_API_KEY are all optional

Option B — environment variables. export them in your shell before launching (handy for CI or one-off runs):

export GOOGLE_API_KEY=...
kondi-chat

See the full variable list for every supported key.

Quick start

Pick the cheapest path that matches what you have:

# Free path — runs entirely on free / near-free tiers.
# Gemini 2.5 Pro (free) for coding, GLM-4.5-flash (free on Z.AI Coding Plan) for compression.
export GOOGLE_API_KEY=...
export ZAI_API_KEY=...
kondi-chat                         # then inside the session:
# /mode zai                        # switch to the Z.AI-only profile

# Cheap path — DeepSeek V4 Flash for everything (~$0.14/M in, $0.28/M out).
export DEEPSEEK_API_KEY=...
kondi-chat                         # then inside the session:
# /use deepseek                    # pin all turns to DeepSeek

# Best-value path — multi-provider routing across what you have.
export ANTHROPIC_API_KEY=...
export OPENAI_API_KEY=...
export GOOGLE_API_KEY=...
kondi-chat                         # then inside the session:
# /mode best-value                 # router picks per phase across providers

(Profile switching happens inside the session via /mode <name>. Once switched, the choice persists across restarts in .kondi-chat/config.json.)

Then just talk:

> Explain this codebase
> Refactor the auth module to use JWTs instead of sessions
> @opus Architect a new ingest pipeline   # pin one turn
> /use gemini-2.5-pro                     # pin until you say otherwise
> /cost                                    # see who did what, for how much
> /routing                                  # see the router's tier-by-tier decisions

Features

Multi-model routing

Every message is classified (coding task vs discussion) and routed to the best available model based on your active budget profile. The router has three tiers, evaluated in order:

1. NN Router — fast, trained on your accumulated usage data. Runs automatically once you have ≥100 samples across ≥2 models. Falls through if not yet trained or low-confidence.
2. Intent Router — the primary tier. An LLM reads every enabled model's description + capabilities and classifies which one best fits the current task. Scoped to the active profile's allowedProviders (so zai mode never escapes to Claude). Uses a cheap classifier model chosen from the in-profile registry (e.g. glm-4.5-flash — free — in zai mode).
3. Rule Router — minimal phase/task-kind fallback. Only runs if the intent tier fails or produces no candidate.

You see the routing decision on every turn:

router: phase=execute (coding intent detected)
→ glm-4.6 (intent: coding)

Run /routing at any time to see the tier distribution (intent/nn/rules), per-model success rates and cost, model×tier matrix, NN training readiness, and per-phase breakdown. This is the tool for verifying that routing is actually hitting the intent tier as expected.

Budget profiles

A mode is a budget profile — a named bundle of cost caps, iteration caps, and model preferences. You switch modes with /mode; the words "mode" and "profile" mean the same thing. kondi-chat ships these:

Mode	Use case	Iteration cap	Cost cap
quality	Complex architecture, frontier reasoning	30	$10.00
balanced	Everyday coding and chat (default)	20	$3.00
cheap	Quick lookups, high-volume exploration	8	$0.75
zai	Z.AI (GLM) Coding Plan — glm-5.1 plans, glm-4.6 codes, glm-4.5-flash compresses (free)	20	$3.00
best-value	Multi-provider routing — Sonnet/GPT-5.4 plan, Gemini codes (free), Sonnet reviews	24	$5.00
orchestra	Deterministic pipeline — GPT-5.4 plans, Gemini codes, GLM-5.1 reviews	24	$5.00

Run /mode with no argument to see the list and which one is active. Switch at any time: /mode quality. The active profile is persisted to .kondi-chat/config.json so it survives restarts.

Provider scoping. A profile can restrict routing to a subset of providers by setting allowedProviders. When set, the intent router, rule router, cross-turn compactor, and intent classifier LLM all stay inside that allow-list — nothing leaks out. See the zai profile for an example.

Create custom profiles by adding JSON files to .kondi-chat/profiles/:

{
  "name": "my-profile",
  "description": "Custom workflow",
  "executionPreference": ["coding", "fast-coding"],
  "planningPreference": ["reasoning", "planning"],
  "loopIterationCap": 15,
  "loopCostCap": 5.00,
  "contextBudget": 40000,
  "maxOutputTokens": 8192,
  "allowedProviders": ["anthropic", "openai"]
}

contextBudget is also the ceiling the compactor enforces. Inside an agent loop, old tool results are progressively stubbed to stay under it — no LLM calls, just local string rewriting. Between turns, cross-turn compaction fires at contextBudget × 1.2 and summarizes older messages using the profile-scoped compression model (glm-4.5-flash in zai mode, claude-haiku in unrestricted profiles). See /help compression and /help intent-router.

Multi-provider pipelines and model preferences

A profile declares which models are available via rolePinning and the router intelligently selects among them per phase. Pins are soft preferences with fallback semantics, not hard overrides — the intent router gets first shot at picking the best model for each step (informed by phase context, model descriptions, and cost), and the pin only fires if the router produces no result.

The best-value profile demonstrates the design:

{
  "name": "best-value",
  "allowedProviders": ["anthropic", "openai", "google", "zai"],
  "rolePinning": {
    "discuss":      "claude-sonnet-4-5-20250929",
    "dispatch":     "gpt-5.4",
    "execute":      "models/gemini-2.5-pro",
    "reflect":      "claude-sonnet-4-5-20250929",
    "compress":     "glm-4.5-flash",
    "state_update": "glm-4.5-flash"
  }
}

The classifier sees exactly these 5 models (Sonnet, GPT-5.4, Gemini Pro, GLM-flash — plus Opus which is also enabled in the registry). For the dispatch phase, the profile suggests GPT-5.4 — but the classifier also sees Opus and can choose it when the task is genuinely complex enough to justify the 6× price premium. For simpler planning calls, GPT-5.4 wins on cost. The router makes that call per turn, not per session.

The pipeline passes context between phases so the classifier makes informed decisions: "Gemini just wrote the code, tests passed, now pick a reviewer — and don't pick the same model that wrote the code." The phase descriptions are baked into the prompt so the classifier understands what reflect means (code review, catch bugs) vs. dispatch (architecture, planning, task decomposition).

Two bundled profiles use this:

• best-value — Sonnet + GPT-5.4 for chat/planning, Gemini Pro for coding (free), Sonnet for review, GLM-flash for compression (free). The router chooses between comparable models based on task complexity.
• orchestra — deterministic pipeline: GPT-5.4 plans, Gemini codes, GLM-5.1 reviews. More rigid, for workflows where you want explicit role binding.

Activate with /mode best-value or /mode orchestra.

Agent tools

The agent has access to:

Tool	Description
read_file	Read files from the project
write_file	Create or overwrite files
edit_file	Search/replace edits with diff output
list_files	List directory contents
search_code	Grep for patterns across the codebase
run_command	Execute shell commands
create_task	Dispatch multi-phase coding tasks (routes each phase to a profile-appropriate model)
consult	Ask a domain-expert consultant for an opinion — see the Consultants section
update_plan	Update the session goal and plan
update_memory	Write to KONDI.md memory files
git_status	View git repository state
git_commit	Create git commits
git_diff	View diffs
web_search	Search the web (requires Brave API key)
web_fetch	Fetch and extract web page content
spawn_agent	Spawn sub-agents for parallel work

Council deliberation

For decisions that matter, run a multi-model council explicitly:

/council list                                  # see available council profiles
/council run analysis "Should we use microservices or a monolith here?"

Multiple models debate the question across several rounds, with a manager model synthesizing the final recommendation. The deliberation engine is bundled — no extra install. Council profiles live in .kondi-chat/councils/*.json (curated presets — coding, analysis, debate, code-planning — are seeded on first run); each one defines the personas that participate, their models and stances, how many rounds run, and the debate format. Edit those files or drop in your own.

Councils are explicit-only. The agent cannot auto-invoke a council — COUNCIL_TOOL is deliberately not registered in the agent toolset. Councils are expensive (fan out across frontier models for multiple rounds) and blocking (synchronous subprocess) so they only run when the user types /council themselves.

Domain-expert consultants

The agent can call on domain experts via the consult tool when it decides a problem benefits from a specialized perspective. Defaults ship with:

• aerospace-engineer — flight safety, fault tolerance, margins, certification
• security-auditor — OWASP top-10, authn/authz, input validation, crypto misuse
• database-architect — indexes, query plans, migration safety, isolation levels

Consultants are defined in .kondi-chat/consultants.json (auto-created on first run with the defaults above). Each entry:

{
  "role": "ml-researcher",
  "name": "ML Research Scientist",
  "description": "Review experimental designs, loss functions, evaluation protocols, distribution shift, and reproducibility.",
  "provider": "anthropic",
  "model": "claude-sonnet-4-5-20250929",
  "system": "You are an ML research scientist. When reviewing an experimental design, think about: sample size and power, evaluation leakage, distribution shift between train and deploy, ablation coverage, baseline fairness, reproducibility (seeds, data provenance, code), and what conclusion the reported results actually support vs. what is being claimed. Be blunt about overclaiming.",
  "contextText": "Project is a recommender system for a mid-size e-commerce site. Eval is offline NDCG@10 against a 30-day holdout. Production serves 1M users/day.",
  "contextFiles": ["docs/eval-protocol.md", "docs/data-splits.md"],
  "maxOutputTokens": 2048
}

Field reference:

Field	Purpose
role	Machine id — what the agent passes in consult({role: "..."}).
name	Human-readable display name.
description	Shown to the agent so it can decide when to reach for this consultant. Keep it concrete — "review for flight safety and fault tolerance" beats "do engineering review."
provider + model	Which LLM runs the persona. Can be any enabled model, regardless of the active profile's allowedProviders.
system	The persona definition — this is where the actual expertise lives.
contextText (optional)	Static baseline context baked into every call: mission specs, target platform, stable constraints, vocabulary.
contextFiles (optional)	Relative paths read from disk lazily on each call (not at startup), so edits to spec files show up in the next consultation without restarting. Capped per-file at 50KB and 200KB total by default — override with contextFileMaxBytes / contextTotalMaxBytes if you need more. Paths are sandboxed to the working directory; ../ escapes are rejected.
maxOutputTokens (optional)	Default 2048.

The agent decides when to consult. Consultants are pure text-in / text-out — they see only the question (plus any caller-supplied context arg, plus the consultant's own contextText + contextFiles), not the session history, and they cannot call any tools themselves. If you need an expert that can actually read arbitrary files or run commands, use spawn_agent instead.

Consultations log to the ledger as phase: consult with the role in the reason field, so /routing and /cost attribute the spend to the consultant that did the work. Run /consultants in the TUI to see the roster, including a preview of each consultant's baseline context and attached files.

Autonomous loop mode

Run the agent against a goal until it explicitly reports completion or hits the profile's iteration/cost caps:

/loop fix all the failing tests and commit when green
/loop find every TODO in src/ and resolve them

Unlike a regular turn — which stops as soon as the model returns a final answer without calling tools — /loop synthesizes a "continue" follow-up whenever the model appears to stop early, and keeps iterating. The model signals termination itself by emitting DONE or STUCK: <reason> on a line by itself, at which point the loop ends and the final summary is written to scrollback.

Safety rails:

• LoopGuard enforces the active profile's loopIterationCap and loopCostCap. The loop can't outrun your budget.
• Checkpoints are still created before the first mutating tool call, so /undo works the same way as for a normal turn.
• Permission prompts still fire for every confirm-tier tool call. Use t in the permission dialog to yolo-approve everything for the duration of the current iteration if you trust the loop.
• Ctrl+C aborts the TUI (and therefore the backend), stopping the loop immediately.
• All tool-call, activity, and message events stream in real time — you can watch the loop work and Ctrl+O into the tool-call detail view at any moment.

@mention routing

Direct a message to a specific model by prefixing your prompt with @<alias>:

> @opus Analyze the security implications of this auth flow
> @deep Write the implementation based on the analysis above
> @gemini Review the code for edge cases

Autocomplete. Typing @ as the first character of the input pops an autocomplete list of every enabled model alias (same source as /models). Keep typing to narrow it — @ge filters to @gemini and @gemini-pro.

Prefix matching. Aliases resolve on an unambiguous prefix, so you don't have to type the whole thing. @gemi lands on @gemini because it's the only enabled alias starting with those letters. If your prefix is ambiguous (e.g. @gem when both @gemini and @gemini-pro are enabled), the backend reports the ambiguity and lists the candidates so you can disambiguate.

/use <alias> is the persistent equivalent: it pins all subsequent turns to the given model until you run /use auto to return to router-based selection. The bottom-of-viewport model indicator updates immediately when /use runs — no need to send a turn first.

Session management

• Session resume — pick up where you left off with /resume
• Undo / checkpoints — revert file changes with /undo
• Auto-save — sessions are saved periodically and on exit

MCP support

Connect to any MCP-compatible tool server:

/mcp add filesystem npx -y @modelcontextprotocol/server-filesystem /home
/mcp add github npx -y @modelcontextprotocol/server-github
/mcp add my-api http https://api.example.com/mcp

MCP tools appear alongside built-in tools and are available to the agent automatically.

Git integration

The TUI shows your current branch and dirty-file count in the status bar. Git tools (git_status, git_commit, git_diff) let the agent interact with your repository. Checkpoints are created before mutating operations so /undo can roll back.

Permission system

Tool calls that write files, run commands, or access the network require approval:

┌─ permission ──────────────────────────────────────────┐
│ Permission required [confirm]                          │
│                                                        │
│ Tool: run_command                                      │
│ npm test                                               │
│                                                        │
│ [y/⏎] approve   [n] deny   [a] same cmd (session)      │
│ [t] yolo — approve everything for the rest of this turn│
└────────────────────────────────────────────────────────┘

• y / Enter — approve this one call
• n / Esc — deny
• a — approve this exact command (fingerprint-matched) for the rest of the session
• t — yolo: approve every confirm-tier tool call until the assistant turn ends. Cleared automatically when the turn finishes. Does not bypass always-confirm tier (rm -rf, sudo, force-push to main, etc. — still prompt every time)

Configure defaults in .kondi-chat/permissions.json.

Analytics and cost tracking

/analytics          # usage by model/provider (last 30 days)
/analytics 7        # last 7 days
/analytics export   # export all data as JSON
/cost               # cost breakdown for current session

Non-interactive mode

Run kondi-chat in CI, scripts, or pipelines:

# Pipe a prompt
echo "Explain this error" | kondi-chat --pipe

# Direct prompt
kondi-chat --prompt "Add error handling to auth.ts" --json

# Auto-approve specific tools
kondi-chat --prompt "Fix the tests" --auto-approve run_command,write_file

Commands

Command	Description
/mode [profile]	Show or switch budget profile. Persisted across restarts via config.json.
/use <alias>	Force a specific model (/use auto for router). Supports unambiguous prefix matching — /use gemi → gemini. Updates the model indicator immediately.
/models	List available models and aliases
/health	Check model availability
/routing	Routing stats dashboard — tier distribution (intent/nn/rules), per-model cost, model×tier matrix, NN training readiness, per-phase breakdown
/status	Session stats and context utilization
/cost	Cost breakdown by model
/analytics [days]	Usage analytics
/consultants	List domain-expert consultants the agent can call via the consult tool
/council [list|run]	Council deliberation — explicit-only, never auto-invoked by the agent
/loop <goal>	Autonomous agent loop with guards — cycles until the model emits DONE / STUCK or LoopGuard caps hit
/undo [n]	Undo last n file changes
/resume	Resume a previous session
/sessions	List saved sessions
/mcp	List MCP servers and tools
/tools	List agent tools
/help [topic]	Show all commands or a specific help topic (zai, compression, intent-router, type-ahead, mentions, consultants, etc.)
/quit	Exit

Keyboard shortcuts

Key	Action
Enter	Send message — or queue it if a turn is already running
Ctrl+N	Insert newline in input
Ctrl+O	Toggle tool-call detail view (current turn)
Ctrl+T	Toggle token-stats detail view (current turn)
Ctrl+R	Toggle reasoning detail view — hidden chain-of-thought from reasoning models (GLM-5.x, OpenAI o-series, DeepSeek-R1, Anthropic extended thinking)
Ctrl+Y	Copy last assistant response to system clipboard (raw markdown)
Ctrl+A	Toggle activity log
Ctrl+P	Toggle copy mode — pauses all viewport redraws + scrollback inserts so terminal-native text selection isn't killed by streaming updates. Press again to resume; queued content drains.
← / →	Move cursor within input
Home / End	Jump to start / end of input
Backspace / Delete	Delete before / at cursor
↑ / ↓	Recall input history (bash-style)
Esc	Close detail view → clear input → clear queued submits (in that order)
Ctrl+C	Exit

Type-ahead queue. If you hit Enter while a turn is still running, the new message is queued instead of fired concurrently. The TUI renders a dim ⧗ queued: … line in scrollback as confirmation, and the status bar shows ⧗ queued: N (Esc to clear). When the current turn finishes, the oldest queued entry fires automatically and the spinner picks back up. This guarantees at most one handleSubmit is ever in flight on the backend — concurrent turns can't race over shared session state, tool call attribution, or the permission dialog. Esc on an empty input clears the queue if you change your mind mid-stack.

Mouse wheel scrolls the terminal scrollback. Text selection and copy work natively — no special mode needed.

Markdown tables in assistant responses are rendered with box-drawing characters. Code fences, headers, and lists render as-is. When a response was produced by a reasoning model, a dim magenta [^R reasoning] tag appears in the header so you know Ctrl+R will show something.

Configuration

Environment variables

Set these in a .env file or export them — see Set up your API keys for where .env is read from. The router auto-excludes any provider whose key is missing, so an unset variable is never an error.

Variable	Provider
ANTHROPIC_API_KEY	Anthropic (Claude)
OPENAI_API_KEY	OpenAI (GPT)
DEEPSEEK_API_KEY	DeepSeek
GOOGLE_API_KEY	Google (Gemini)
XAI_API_KEY	xAI (Grok)
ZAI_API_KEY	Z.AI (GLM) — Coding Plan endpoint
BRAVE_SEARCH_API_KEY	Brave Search (web tools)
OLLAMA_BASE_URL	Ollama (local models, default: http://localhost:11434)

Project-level config

All configuration lives in .kondi-chat/ in the project root:

.kondi-chat/
  config.json           # General settings
  permissions.json      # Tool permission tiers
  profiles/             # Budget profiles (quality.json, balanced.json, cheap.json, + custom)
  councils/             # Council profiles (coding.json, analysis.json, debate.json, + custom)
  models.yml            # Model registry
  sessions/             # Saved sessions
  analytics.json        # Usage data
  backend.log           # Backend diagnostic log

Providers

Provider	Models	Key required
Anthropic	Claude Opus, Sonnet, Haiku (with prompt caching + extended thinking)	Yes
OpenAI	GPT-5.4, GPT-4o, o3	Yes
DeepSeek	DeepSeek Chat, Coder	Yes
Google	Gemini 2.5 Pro, Flash	Yes
xAI	Grok	Yes
Z.AI	GLM 5.1, 5, 4.7, 4.6, 4.5, 4.5-air, 4.5-flash (free) — via OpenAI-compatible Coding Plan endpoint	Yes
Ollama	Any local model	No (local)

kondi-chat works with any combination of providers. The router automatically excludes providers without keys and routes to what's available.

Z.AI (GLM Coding Plan)

Z.AI's OpenAI-compatible API is used through the Coding Plan endpoint (https://api.z.ai/api/coding/paas/v4) rather than the general-purpose /api/paas/v4. If you subscribed to the GLM Coding Plan on z.ai, your key is authorized on the coding endpoint only — hitting the general PaaS endpoint returns HTTP 429 with error code 1113 ("insufficient balance"). kondi-chat handles this automatically; just set ZAI_API_KEY in your .env.

Use /mode zai to activate the bundled zai profile, which restricts routing to Z.AI models exclusively via allowedProviders: ["zai"]:

Phase	Capability	Routed to	In/Out per 1M
planning / reasoning / analysis / code-review	planning, reasoning, analysis	glm-5.1	$1.40 / $4.40
execution / coding / fast-coding / general	coding, fast-coding, general	glm-4.6	$0.60 / $2.20
compression / state_update / summarization	summarization	glm-4.5-flash	free

Reasoning tax caveat. glm-5.1 is a reasoning model — it emits hidden chain-of-thought that is billed as output tokens at the full $4.40/M rate but not shown inline. A single 20-char reply can cost 500+ output tokens of invisible thinking. Press Ctrl+R in the TUI to see what the model was actually reasoning about. For high-volume agent-loop work, consider pinning execution to @glm (glm-4.6, non-reasoning) with /use glm so you only pay the reasoning premium on planning phases.

Prompt caching. z.ai's Coding Plan endpoint serves prompt_tokens_details.cached_tokens automatically for repeated prefixes ≥1k tokens. kondi-chat tracks cache hits per call and discounts them 50% in the cost estimator. Cache hit totals appear in /routing and /cost.

Running the backend directly

npm run chat:tui (after building, see Install) is the interactive entry point. For non-interactive use — CI, scripts, piping a prompt — bypass the TUI and call the Node backend directly:

npm start                              # tsx src/cli/backend.ts (JSON-RPC over stdio)
npx tsx src/cli/backend.ts --prompt "Explain this codebase"

The Rust TUI is the only frontend; the Node backend is the engine it talks to over JSON-RPC on stdio. There is no pure-Node "chat" frontend.

Rust toolchain install (if you don't have one): curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh

Architecture

┌─────────────────────────────────────────┐
│  Rust TUI (tui/)                         │
│  Ratatui + Crossterm, inline viewport    │
│  Renders to terminal, handles input      │
├──────────── JSON-RPC over stdio ────────┤
│  Node.js Backend (src/)                  │
│  LLM routing, tools, MCP, context mgmt  │
│  Providers: Anthropic, OpenAI, etc.      │
└─────────────────────────────────────────┘

The Rust TUI spawns the Node.js backend as a child process. They communicate via JSON-RPC over stdin/stdout. All LLM calls, tool execution, and state management happen in the backend. The TUI is purely display and input.

License

MIT -- see LICENSE.