feat: add copilot runner adapter (#48)

## What is this?

This PR adds GitHub Copilot CLI as a supported Skillgym runner. It lets
suites run against `agent.type: "copilot"`, keeps Copilot session
artifacts, and documents the new support across the README, docs, and
website.

## How does it work?

The new adapter launches `copilot` in non-interactive JSONL mode with
the configured model, isolates Copilot state under each artifact
directory, and normalizes Copilot messages, tool calls, file reads,
commands, explicit skill loads, and OpenTelemetry token usage into
Skillgym session reports. The runner registry, config parser, public
types, max-step monitor, and examples now recognize the `copilot` agent
type.

## Why is this useful?

Skill authors can now compare GitHub Copilot CLI against the existing
OpenCode, Codex, Claude Code, and Cursor Agent runners using the same
Skillgym cases and assertions. Explicit Copilot skill-load events are
captured as detected skills, so skill-selection suites can verify that
Copilot actually loaded the expected skill instead of relying only on
file-read inference.

Author: V3RON

.changeset/brave-copilot-runners.md

@@ -0,0 +1,5 @@
+---
+"skillgym": minor
+---
+
+Add GitHub Copilot CLI as a supported runner with JSONL session normalization, isolated Copilot runtime state, token usage capture, explicit skill detection, docs, and website coverage.

DICTIONARY.md

@@ -15,6 +15,7 @@ Before adding or changing a term, check this file and ask the user for approval.
 - `prompt`: Exact instruction sent to the agent.
 - `runner`: Configured execution target: agent type plus model.
 - `agent`: External coding agent doing the work.
+- `copilot`: GitHub Copilot CLI agent type.
 - `suite run`: One top-level `skillgym run ...` over one suite.
 - `execution`: One case x runner unit of work, including all its repetitions and retries.
 - `repetition`: One planned sample within an execution.

README.md

@@ -16,6 +16,7 @@ When you evaluate agent skills manually, it is hard to tell whether the agent ac
 - Codex CLI
 - Claude Code
 - Cursor Agent (Cursor CLI `agent`)
+- GitHub Copilot CLI (`copilot`)
 
 ## Quick start
 
@@ -66,6 +67,12 @@ const config = {
         model: "composer-2-fast",
       },
     },
+    "copilot-main": {
+      agent: {
+        type: "copilot",
+        model: "gpt-5.4-mini",
+      },
+    },
   },
 } satisfies SkillgymConfig;
 
@@ -165,7 +172,7 @@ Most important config properties:
 - `run.maxSteps`: best-effort limit on streamed agent steps before skillgym terminates the execution
 - `run.workspace`: default workspace mode for the suite
 - `defaults.timeoutMs`: default per-case timeout
-- `runners.<id>.agent.type`: which agent integration to use, currently `opencode`, `codex`, `claude-code`, or `cursor-agent`
+- `runners.<id>.agent.type`: which agent integration to use, currently `opencode`, `codex`, `claude-code`, `cursor-agent`, or `copilot`
 - `runners.<id>.agent.model`: model passed to that runner
 - `snapshots`: token regression baseline configuration
 
@@ -181,7 +188,7 @@ The execution unit is one case x runner pair. `skillgym` expands the suite into
 
 For concurrent schedules, `run.maxParallel` defaults to `os.availableParallelism()`. This limits how many Skillgym executions are active at once; it does not pin or limit CPU cores used by an individual agent process.
 
-Concurrent schedules do not copy or isolate the workspace by themselves. Overlapping executions may still interact through the same filesystem state and live runner output unless you use isolated workspaces. OpenCode, Codex, and Claude Code runtime state are isolated per execution under each artifact directory.
+Concurrent schedules do not copy or isolate the workspace by themselves. Overlapping executions may still interact through the same filesystem state and live runner output unless you use isolated workspaces. OpenCode, Codex, Claude Code, and Copilot runtime state are isolated per execution under each artifact directory.
 
 `run.repeat` is useful when you want stability sampling instead of a single lucky pass. Each case x runner execution keeps running until it records the requested number of successful classified results, or stops early when one repetition still fails after exhausting `run.repeatFailure` retries.
 

docs/explain.md

@@ -91,6 +91,7 @@ Deferred explain is implemented for:
 - Codex
 - Claude Code
 - Cursor Agent
+- GitHub Copilot CLI
 
 Each runner must have a resumable session id captured during the original run.
 

docs/readme.md

@@ -18,6 +18,7 @@ Supported in V1:
 - Codex CLI
 - Claude Code CLI
 - Cursor Agent CLI
+- GitHub Copilot CLI
 
 ## Core principles
 

docs/session-report.md

@@ -50,11 +50,12 @@ export interface SessionReport {
     stderrPath?: string;
     sessionPath?: string;
     exportPath?: string;
+    telemetryPath?: string;
   };
 }
 ```
 
-`runner.id` is the configured runner id, and `runner.agent.type` identifies the backing integration such as `opencode`, `codex`, or `claude-code`.
+`runner.id` is the configured runner id, and `runner.agent.type` identifies the backing integration such as `opencode`, `codex`, `claude-code`, `cursor-agent`, or `copilot`.
 
 ## Skill detection
 
@@ -110,7 +111,7 @@ Preferred source order:
 
 1. `inputTokens + outputTokens + reasoningTokens - cacheTokens`
 
-This keeps OpenCode, Codex, and Claude Code comparable even though the raw provider fields use different cache semantics.
+This keeps OpenCode, Codex, Claude Code, Cursor Agent, and Copilot comparable even though the raw provider fields use different cache semantics.
 
 Snapshot enforcement uses the selected token metric directly and does not fall back to character counts when token usage is unavailable.
 

skillgym.config.ts

@@ -35,6 +35,12 @@ const config = {
         model: "auto",
       },
     },
+    "copilot-main": {
+      agent: {
+        type: "copilot",
+        model: "gpt-5.4-mini",
+      },
+    },
   },
 } satisfies SkillgymConfig;