agent-duelist 0.2.0 → 0.3.0

package/README.md

@@ -1,31 +1,46 @@
 # agent-duelist
 
+[![npm version](https://img.shields.io/npm/v/agent-duelist?color=f59e0b)](https://www.npmjs.com/package/agent-duelist)
 [![CI](https://github.com/DataGobes/agent-duelist/actions/workflows/ci.yml/badge.svg)](https://github.com/DataGobes/agent-duelist/actions/workflows/ci.yml)
 [![Docs](https://img.shields.io/badge/docs-landing%20page-f59e0b)](https://datagobes.github.io/agent-duelist/)
+[![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](./LICENSE)
 
-> Pit LLM providers against each other on agent tasks — Duel your models.
+> Pit LLM providers against each other on agent tasks — **Duel your models.**
 >
 > **[View the landing page →](https://datagobes.github.io/agent-duelist/)**
 
-`agent-duelist` is a TypeScript-first framework to pit multiple LLM providers against each other on the same tasks and get structured, reproducible results: correctness, latency, tokens, and cost.
+`agent-duelist` is a TypeScript-first benchmarking framework that runs the same tasks against multiple LLM providers and gives you structured, reproducible results: correctness, latency, tokens, cost, and more.
+
+```bash
+npx duelist init   # scaffold a config
+npx duelist run    # see who wins
+```
 
 ## What you get
-> ![Agent Duelist console output](docs/assets/screenshot.png)
 
+**Console output** — box-drawing tables with medals, color-ranked metrics, sparklines, and per-task winners:
+
+![Agent Duelist console output](docs/assets/screenshot.png)
+
+**HTML report** — a self-contained, shareable single-file report with sortable tables, progress bars, tab navigation, and summary cards:
 
-- Compare OpenAI, Azure OpenAI, Anthropic, Google Gemini, and any OpenAI-compatible gateway.
-- Define tasks once, run them against many providers.
-- Get CLI tables and JSON results you can feed into dashboards, CI, or docs.
+![Agent Duelist HTML report](docs/assets/screenshot-html.png)
 
 ---
 
 ## Why agent-duelist?
 
-- **Provider-agnostic**: One config, many providers. Swap models and gateways without rewriting your tasks.
-- **Agent-focused**: Designed for agent workflows and tool use, not just single-turn prompts.
-- **Realistic metrics**: Latency, token counts, and cost estimates based on a pricing catalog.
-- **TypeScript-native DX**: Strongly typed APIs, Zod schemas for structured outputs, and a simple `defineArena()` entrypoint.
-- **CLI-first**: `npx duelist init` → `npx duelist run` gets you from zero to useful table in minutes.
+| | |
+|---|---|
+| **Provider-agnostic** | One config, many providers. Swap models and gateways without rewriting tasks. |
+| **Agent-focused** | Built for agent workflows and tool use, not just single-turn prompts. |
+| **Task packs** | Built-in benchmark suites — run with `--pack structured-output`, zero config needed. |
+| **Quality-first ranking** | Medals decided by output quality. Speed and cost only break ties — a fast model that gets nothing right won't medal. |
+| **7 built-in scorers** | Correctness, latency, cost, schema validation, fuzzy similarity, LLM-as-judge, and tool usage. |
+| **Fair benchmarking** | Tasks run sequentially while providers race in parallel — fair latency comparison with no queue-induced penalties. |
+| **TypeScript-native** | Strongly typed APIs, Zod schemas for structured outputs, and a simple `defineArena()` entrypoint. |
+| **CI-ready** | Regression detection with confidence intervals, cost budgets, PR comments, and a prebuilt GitHub Action. |
+| **CLI-first** | `npx duelist init` → `npx duelist run` gets you from zero to results in minutes. |
 
 ---
 
@@ -39,7 +54,7 @@ pnpm add agent-duelist
 yarn add agent-duelist
 ```
 
-You'll also need API keys for the providers you want to benchmark, for example:
+Set API keys for the providers you want to benchmark:
 
 ```bash
 export OPENAI_API_KEY=sk-...
@@ -50,7 +65,7 @@ export GOOGLE_API_KEY=...
 
 ---
 
-## One-minute quickstart
+## Quickstart
 
 Initialize a config:
 
@@ -58,7 +73,7 @@ Initialize a config:
 npx duelist init
 ```
 
-This creates `arena.config.ts` in your project. A minimal example:
+This creates `arena.config.ts` in your project:
 
 ```ts
 // arena.config.ts
@@ -98,33 +113,78 @@ Run the benchmark:
 npx duelist run
 ```
 
-You'll see a matrix like:
+You'll see a results matrix:
 
-- Rows: tasks (`simple-qa`, `structured-extraction`)
-- Columns: providers (`openai/gpt-5-mini`, `azure/gpt-5-mini`)
-- Cells: correctness score, latency, tokens, and estimated cost.
+- **Rows**: tasks (`simple-qa`, `structured-extraction`)
+- **Columns**: providers (`openai/gpt-5-mini`, `azure/gpt-5-mini`)
+- **Cells**: correctness score, latency, tokens, and estimated cost
 
-For CI or further processing:
+Export the results in different formats:
 
 ```bash
+# JSON for CI pipelines and dashboards
 npx duelist run --reporter json > results.json
+
+# Self-contained HTML report you can share or host
+npx duelist run --reporter html --output report.html
 ```
 
 ---
 
-## Core concepts
+## Task packs
 
-### Providers
+Task packs are **built-in benchmark suites** — curated sets of tasks with recommended scorers. Use them to benchmark providers without writing any tasks yourself.
 
-Providers are **factory functions** that return plain objects implementing a shared `ArenaProvider` interface.
+```bash
+# List available packs
+npx duelist run --pack list
 
-This lets you:
+# Run a pack with your providers
+npx duelist run --pack structured-output --config arena.config.ts
+```
 
-- Swap providers without changing tasks.
-- Wrap or extend providers in your own code.
-- Mock providers in tests.
+Your config only needs to define providers — the pack supplies tasks and scorers:
 
-Examples:
+```ts
+// arena.config.ts
+import { defineArena, openai, anthropic } from 'agent-duelist'
+
+export default defineArena({
+  providers: [
+    openai('gpt-5-mini'),
+    anthropic('claude-sonnet-4.6'),
+  ],
+  tasks: [],     // ignored when --pack is used
+  scorers: [],   // pack supplies its own scorers
+})
+```
+
+### Available packs
+
+| Pack | Tasks | Description |
+|------|-------|-------------|
+| `structured-output` | 6 | Zod schema stress test — flat objects, nesting, arrays, enums, empty arrays, and adversarial input |
+
+Packs work with both `run` and `ci` commands:
+
+```bash
+# CI with a task pack
+npx duelist ci --pack structured-output --threshold correctness=0.1 --budget 1.00
+```
+
+You can also combine multiple packs:
+
+```bash
+npx duelist run --pack structured-output,another-pack
+```
+
+---
+
+## Core concepts
+
+### Providers
+
+Providers are **factory functions** that return plain objects implementing a shared `ArenaProvider` interface. This lets you swap providers without changing tasks, wrap or extend providers in your own code, and mock providers in tests.
 
 ```ts
 import {
@@ -133,25 +193,40 @@ import {
   anthropic,
   gemini,
   openaiCompatible,
-  type ArenaProvider,
 } from 'agent-duelist'
 
+// OpenAI
 const oai = openai('gpt-5-mini')
 
+// Azure OpenAI
 const azure = azureOpenai('gpt-5-mini', {
   deployment: 'my-deployment',
 })
 
+// Anthropic
 const claude = anthropic('claude-sonnet-4.6')
 
-const gem = gemini('gemini-3-flash-preview') // uses GOOGLE_API_KEY
+// Google Gemini
+const gem = gemini('gemini-3-flash-preview')
 
-const local: ArenaProvider = openaiCompatible({
+// Any OpenAI-compatible gateway (Ollama, LiteLLM, vLLM, etc.)
+const local = openaiCompatible({
   id: 'local/llama',
-  name: 'Local Gateway',
+  name: 'Local Ollama',
   baseURL: 'http://localhost:11434/v1',
   model: 'llama3.3',
   apiKeyEnv: 'LOCAL_LLM_API_KEY',
+  free: true,           // registers zero-cost pricing
+})
+
+// Reasoning models that emit <think> blocks (DeepSeek-R1, MiniMax M2.5, etc.)
+const deepseek = openaiCompatible({
+  id: 'deepseek/r1',
+  name: 'DeepSeek R1',
+  baseURL: 'https://api.deepseek.com/v1',
+  model: 'deepseek-reasoner',
+  apiKeyEnv: 'DEEPSEEK_API_KEY',
+  stripThinking: true,  // strips <think>...</think> from output
 })
 ```
 
@@ -178,6 +253,7 @@ interface ArenaTask {
   prompt: string
   expected?: unknown         // used by correctness scorers
   schema?: ZodSchema<any>    // used by schema-based scorers
+  tools?: ToolDefinition[]   // used by tool-calling scorers
 }
 ```
 
@@ -206,7 +282,7 @@ const tasks: ArenaTask[] = [
 
 ### Scorers
 
-Scorers take raw model outputs and turn them into **numeric scores** (0–1) with optional details. Built-in scorers:
+Scorers turn raw model outputs into **numeric scores** (0–1) with optional details. Seven built-in scorers ship out of the box:
 
 | Scorer | What it measures |
 |--------|-----------------|
@@ -215,7 +291,8 @@ Scorers take raw model outputs and turn them into **numeric scores** (0–1) wit
 | `correctness` | Exact match against `expected` (deep-equal, key-order independent for objects) |
 | `schema-correctness` | Validates output against the task's Zod `schema` via `safeParse()` |
 | `fuzzy-similarity` | Jaccard token-overlap similarity between output and `expected` |
-| `llm-judge-correctness` | Async LLM-as-judge — calls a judge model to score correctness 0–1 |
+| `tool-usage` | Whether the model invoked the expected tool(s) during a tool-calling task |
+| `llm-judge-correctness` | LLM-as-judge — calls a judge model to score accuracy, completeness, and conciseness |
 
 Configure them in your arena:
 
@@ -235,8 +312,6 @@ defineArena({
 
 The judge model defaults to `gpt-5-mini`. It can also be set via the `DUELIST_JUDGE_MODEL` env var. The judge backend is auto-detected from the model name — `gemini-*` models use Google's API, otherwise it falls back to OpenAI or Azure OpenAI.
 
-You can also add custom scorers for domain-specific metrics (e.g. tool-call correctness, safety, style).
-
 ---
 
 ### Arena options
@@ -245,7 +320,7 @@ You can also add custom scorers for domain-specific metrics (e.g. tool-call corr
 
 | Option | Type | Default | Description |
 |--------|------|---------|-------------|
-| `runs` | `number` | `1` | Number of runs per provider × task combination. Higher values improve statistical confidence for CI regression detection. |
+| `runs` | `number` | `1` | Number of runs per provider x task combination. Higher values improve statistical confidence for CI regression detection. |
 | `judgeModel` | `string` | `'gpt-5-mini'` | Model used by the `llm-judge-correctness` scorer. Also settable via `DUELIST_JUDGE_MODEL` env var. Gemini models auto-route to Google's API. |
 | `timeout` | `number` | `60000` | Per-request timeout in milliseconds. Requests exceeding this are marked as failures. Prevents hanging on unresponsive APIs. |
 | `sparklines` | `boolean` | `true` | Show sparkline bars next to percentage scores in the console reporter. Disable with `false` if your terminal doesn't render Unicode block characters well. |
@@ -266,43 +341,62 @@ export default defineArena({
 
 ---
 
+## Reporters
+
+agent-duelist includes four output formats, each suited to a different workflow:
+
+| Reporter | Flag | Use case |
+|----------|------|----------|
+| **Console** | `--reporter console` (default) | Interactive development — box-drawing tables with medals, sparklines, color-ranked metrics, and per-task winners |
+| **JSON** | `--reporter json` | CI pipelines, dashboards, and downstream tooling |
+| **HTML** | `--reporter html` | Shareable single-file reports with sortable tables, animated backgrounds, tab navigation, CSS progress bars, medal rankings, and summary cards |
+| **Markdown** | `--comment` (CI mode) | Auto-posted PR comment with comparison table, cost summary, and pass/fail verdict |
+
+Generate an HTML report:
+
+```bash
+npx duelist run --reporter html --output report.html
+```
+
+The HTML report is a single self-contained file — no external dependencies, no build step. Open it in any browser or host it as a static page.
+
+---
+
 ## Cost & pricing
 
 Cost estimation is intentionally transparent and conservative:
 
-1. **Token counts**  
-   Providers return token usage (prompt and completion tokens) in each `TaskResult`. These are treated as the source of truth.
+1. **Token counts** — Providers return token usage (prompt and completion tokens) in each `TaskResult`. These are the source of truth.
 
-2. **Pricing catalog**  
-   `agent-duelist` ships with a **locally bundled catalog** of per-token prices for many models, derived from OpenRouter's public pricing pages.
+2. **Pricing catalog** — `agent-duelist` ships with a **locally bundled catalog** of per-token prices for many models, derived from OpenRouter's public pricing pages.
+   - The catalog maps `(provider, model)` to `{ inputPerM, outputPerM }` in USD per 1M tokens.
+   - Azure OpenAI models resolve back to their base OpenAI models (e.g. `azure/gpt-5-mini` → `openai/gpt-5-mini`).
+   - Cross-provider fallback: models hosted on Groq, Together, Fireworks, etc. resolve to the original provider's pricing.
 
-   - The catalog maps `(provider, model)` → `{ inputPerM, outputPerM }` in USD per 1M tokens.
-   - Azure OpenAI models are resolved back to their base OpenAI models where possible (e.g. `azure/gpt-5-mini` → `openai/gpt-5-mini`) so you don't need to configure Azure pricing manually.
-
-3. **Estimated USD**  
-   The `cost` scorer computes:
+3. **Estimated USD** — The `cost` scorer computes:
 
    ```text
    estimatedUsd = (promptTokens * inputPerM + completionTokens * outputPerM) / 1_000_000
    ```
 
-   In the console reporter, you'll see:
+4. **Unknown models** — If a model is not in the catalog, tokens are still reported and cost is marked as unknown (no fake numbers).
 
-   - token counts: `prompt: X, completion: Y`
-   - cost: `~$0.0XXm` (millicents; fractions of a cent)
-   - a short disclaimer that this is an **estimate** based on a pricing snapshot.
+5. **Custom pricing** — Register pricing for models not in the catalog:
 
-4. **Unknown models**  
-   If a model is not in the catalog:
+   ```ts
+   import { registerPricing } from 'agent-duelist'
 
-   - Tokens are still reported.
-   - Cost is marked as unknown (no fake numbers).
+   registerPricing('custom/my-model', {
+     inputPerToken: 0.000003,
+     outputPerToken: 0.000015,
+   })
+   ```
 
-You can update the catalog with a script that re-scrapes OpenRouter's public pricing page when prices change.
+You can update the bundled catalog with `npm run update:pricing`, which re-scrapes OpenRouter's public pricing page.
 
 ---
 
-## CLI usage
+## CLI reference
 
 ### `duelist init`
 
@@ -310,30 +404,46 @@ Scaffold a new `arena.config.ts` in the current directory.
 
 ```bash
 npx duelist init
+npx duelist init --force  # overwrite an existing config
 ```
 
+| Option | Description |
+|--------|-------------|
+| `--force` | Overwrite existing config file |
+
 ### `duelist run`
 
 Run benchmarks defined in your arena config.
 
 ```bash
-# Run with the default config (arena.config.ts)
+# Default config, console output
 npx duelist run
 
-# Use a custom config
+# Custom config
 npx duelist run --config path/to/arena.config.ts
 
-# Get JSON instead of a table
-npx duelist run --reporter json
+# Run a built-in task pack
+npx duelist run --pack structured-output
+
+# List available packs
+npx duelist run --pack list
+
+# JSON for piping
+npx duelist run --reporter json > results.json
+
+# HTML report
+npx duelist run --reporter html --output report.html
 
-# Suppress per-result progress lines
+# Quiet mode
 npx duelist run --quiet
 ```
 
 | Option | Description |
 |--------|-------------|
 | `-c, --config <path>` | Path to config file (default: `arena.config.ts`) |
-| `--reporter <type>` | Output format: `console` (default) or `json` |
+| `--pack <names>` | Run built-in task pack(s) instead of config tasks. Comma-separated for multiple packs. Use `list` to show available packs. |
+| `--reporter <type>` | Output format: `console` (default), `json`, or `html` |
+| `--output <path>` | Output file path for HTML reporter (default: `duelist-report.html`) |
 | `-q, --quiet` | Suppress per-result progress |
 
 ### `duelist ci`
@@ -347,6 +457,9 @@ npx duelist ci --update-baseline
 # Subsequent runs — compare against baseline
 npx duelist ci --threshold correctness=0.1 --budget 1.00
 
+# Run CI with a task pack
+npx duelist ci --pack structured-output --threshold correctness=0.1
+
 # Post comparison table as a PR comment (GitHub Actions)
 npx duelist ci --threshold correctness=0.1 --comment
 ```
@@ -354,6 +467,7 @@ npx duelist ci --threshold correctness=0.1 --comment
 | Option | Description |
 |--------|-------------|
 | `-c, --config <path>` | Path to config file (default: `arena.config.ts`) |
+| `--pack <names>` | Run built-in task pack(s) instead of config tasks |
 | `--baseline <path>` | Baseline JSON file (default: `.duelist/baseline.json`) |
 | `--budget <dollars>` | Max total cost in USD — fails if exceeded |
 | `--threshold <scorer=delta>` | Regression threshold (repeatable, e.g. `--threshold correctness=0.1 --threshold cost=0.002`) |
@@ -368,24 +482,61 @@ npx duelist ci --threshold correctness=0.1 --comment
 - Without `--threshold` flags, regression detection is skipped entirely — only `--budget` is enforced.
 - Results with high variance (CV > 0.3) are flagged as **flaky** with a warning.
 
-The CLI loads TypeScript configs directly using a lightweight runtime loader so users don't need to precompile their config.
+The CLI loads TypeScript configs directly using a lightweight runtime loader so you don't need to precompile your config.
+
+---
+
+## Tool-calling agent example
+
+agent-duelist supports tool-calling tasks — define tools with Zod-typed parameters and handlers, and the provider will execute them during the benchmark:
+
+```ts
+import { defineArena, openai } from 'agent-duelist'
+import { z } from 'zod'
+
+const weatherTool = {
+  name: 'getCurrentWeather',
+  description: 'Get the current weather in a given city',
+  parameters: z.object({ city: z.string() }),
+  handler: async ({ city }: { city: string }) => ({
+    city,
+    tempC: 20,
+  }),
+}
+
+export default defineArena({
+  providers: [openai('gpt-5-mini')],
+  tasks: [
+    {
+      name: 'weather-tool-call',
+      prompt: 'What is the current temperature in Amsterdam? Use the tool.',
+      expected: { city: 'Amsterdam' },
+      tools: [weatherTool],
+    },
+  ],
+  scorers: ['latency', 'cost', 'tool-usage'],
+  runs: 1,
+})
+```
+
+The model calls `getCurrentWeather`, the handler returns a stub result, and the `tool-usage` scorer reports whether the expected tool was invoked. Tool calls and their results are included in the JSON output for inspection.
 
 ---
 
 ## Example: multi-provider benchmark
 
-Here's a richer example comparing multiple providers across tasks:
+A richer example comparing multiple providers across tasks:
 
 ```ts
 // arena.config.ts
-import { defineArena, azureOpenai } from 'agent-duelist'
+import { defineArena, azureOpenai, openai, gemini } from 'agent-duelist'
 import { z } from 'zod'
 
 export default defineArena({
   providers: [
-    azureOpenai('gpt-5-mini'),
+    openai('gpt-5-mini'),
     azureOpenai('gpt-5-nano'),
-    azureOpenai('gpt-5.2-chat'),
+    gemini('gemini-3-flash-preview'),
   ],
   tasks: [
     {
@@ -418,44 +569,13 @@ export default defineArena({
 npx duelist run
 ```
 
-Output includes box-drawing tables with medals, color-ranked metrics, sparkline bars, and a winner row per task — see the [screenshot above](#what-you-get) for a real example.
-
----
-
-## Tool-calling agent example
-
-agent-duelist supports tool-calling tasks — define tools with handlers, and the provider will execute them during the benchmark:
-
-```ts
-import { defineArena, openai } from 'agent-duelist'
-import { z } from 'zod'
-
-const weatherTool = {
-  name: 'getCurrentWeather',
-  description: 'Get the current weather in a given city',
-  parameters: z.object({ city: z.string() }),
-  handler: async ({ city }: { city: string }) => ({
-    city,
-    tempC: 20,
-  }),
-}
+**How scoring works:**
 
-export default defineArena({
-  providers: [openai('gpt-5-mini')],
-  tasks: [
-    {
-      name: 'weather-tool-call',
-      prompt: 'What is the current temperature in Amsterdam? Use the tool.',
-      expected: { city: 'Amsterdam' },
-      tools: [weatherTool],
-    },
-  ],
-  scorers: ['latency', 'cost', 'tool-usage'],
-  runs: 1,
-})
-```
-
-The model calls `getCurrentWeather`, the handler returns a stub result, and the `tool-usage` scorer reports whether the expected tool was invoked. Tool calls and their results are included in the JSON output for inspection.
+- Providers are compared **head-to-head within each task** — all providers receive the same prompt at the same time.
+- Medals are awarded only when a provider is the **sole leader** in a metric column. Ties don't award medals, keeping rankings meaningful.
+- **Quality-first ranking**: medals are decided by quality scorer wins (correctness, schema-correctness, etc.). Efficiency metrics (latency, cost) only break ties among quality-equal providers.
+- **Quality gate**: providers that score 0% on all quality scorers are ineligible for medals entirely — being fast or cheap doesn't compensate for getting nothing right.
+- The overall winner is the provider with the highest average correctness score.
 
 ---
 
@@ -503,8 +623,8 @@ When `--comment` is enabled, the CI posts (or updates) a markdown table on the P
 
 | Provider | Task | Scorer | Baseline | Current | Delta | Status |
 |----------|------|--------|----------|---------|-------|--------|
-| openai/gpt-5-mini | extract | correctness | 0.900 | 0.850 | -0.050 | ⚪ unchanged |
-| openai/gpt-5-mini | extract | latency | 0.920 ± 0.030 | 0.890 ± 0.025 | -0.030 | ⚪ unchanged |
+| openai/gpt-5-mini | extract | correctness | 0.900 | 0.850 | -0.050 | unchanged |
+| openai/gpt-5-mini | extract | latency | 0.920 ± 0.030 | 0.890 ± 0.025 | -0.030 | unchanged |
 
 With cost summary, flakiness warnings, and pass/fail verdict.
 
@@ -512,33 +632,31 @@ With cost summary, flakiness warnings, and pass/fail verdict.
 
 ## Roadmap
 
-Shipped so far:
+**Shipped:**
 
-- OpenAI, Azure OpenAI, Anthropic, Google Gemini, and OpenAI-compatible providers
+- 5 provider types: OpenAI, Azure OpenAI, Anthropic, Google Gemini, and any OpenAI-compatible gateway
 - 7 built-in scorers including LLM-as-judge, tool-usage, schema validation, and fuzzy similarity
 - Tool-calling support with local handlers for agent task benchmarking
-- Console reporter with box-drawing tables, medal rankings, sparkline bars (toggleable), and per-task winner rows
-- Configurable per-request timeout to prevent hanging on unresponsive APIs
-- JSON reporter for CI/pipeline integration
-- Markdown reporter for PR comments
-- `duelist ci` command with regression detection, cost budgets, and flakiness warnings
+- **Task packs**: built-in benchmark suites (`structured-output`) — run with `--pack`, no config writing needed
+- Quality-first medal ranking: output quality decides medals, efficiency only breaks ties
+- Fair head-to-head benchmarking with parallel provider execution
+- 4 reporters: console (tables + medals + sparklines), JSON, HTML (sortable, self-contained), and Markdown (PR comments)
+- `duelist ci` with regression detection (confidence intervals), cost budgets, and flakiness warnings
 - GitHub Action for CI/CD integration
-- Pricing catalog from OpenRouter with refresh script
+- Pricing catalog with cross-provider fallback and `registerPricing()` for custom models
+- `openaiCompatible` with `stripThinking` for reasoning models and `free` flag for local models
+- Configurable per-request timeout
 
-Planned directions (subject to community feedback):
+**Planned** (subject to community feedback):
 
-- **More providers**
-  - OpenRouter-native and more OpenAI-compatible gateways.
-- **Better reporting**
-  - HTML and CSV export options.
-- **Agent workflows**
-  - Multi-step tool chains, multi-hop reasoning, and agent traces.
-- **Plugin system**
-  - First-class support for user-defined providers and scorers.
-- **Embedding-based scoring**
-  - Semantic similarity via embedding distance.
+- **More task packs** — reasoning, summarization, tool-calling, and multi-turn conversation packs
+- **Agent workflows** — multi-step tool chains, multi-hop reasoning, and agent traces
+- **More export formats** — CSV
+- **Plugin system** — first-class support for user-defined providers and scorers
+- **Embedding-based scoring** — semantic similarity via embedding distance
+- **More providers** — OpenRouter-native and additional OpenAI-compatible gateways
 
-If you have a specific use case (framework comparisons, multi-agent competitions, tool-calling benchmarks), please open an issue — those will shape what gets built first.
+Have a use case in mind? [Open an issue](https://github.com/DataGobes/agent-duelist/issues) — community feedback shapes what gets built first.
 
 ---
 
@@ -546,18 +664,18 @@ If you have a specific use case (framework comparisons, multi-agent competitions
 
 Contributions, issues, and feature requests are welcome.
 
-- **Bug reports / ideas**: open a GitHub issue.
+- **Bug reports / ideas**: [open a GitHub issue](https://github.com/DataGobes/agent-duelist/issues).
 - **Code changes**:
-  - Fork the repo.
-  - Create a branch.
-  - Run tests: `npm test`.
-  - Run build: `npm run build`.
-  - Open a PR with a clear description and, if possible, a small repro.
+  1. Fork the repo.
+  2. Create a branch.
+  3. Run tests: `npm test`.
+  4. Run build: `npm run build`.
+  5. Open a PR with a clear description.
 
-Please try to keep PRs narrowly focused (single provider, one new scorer, etc.) so they're easy to review.
+Please keep PRs narrowly focused (single provider, one new scorer, etc.) so they're easy to review.
 
 ---
 
 ## License
 
-MIT.
+MIT