diff --git a/README.md b/README.md index 799c687..0f7de26 100644 --- a/README.md +++ b/README.md @@ -147,6 +147,11 @@ Start `codex`, choose **Trust and continue** for the folder, then at **Hooks nee choose **Trust all and continue**. The hooks take effect from your next Codex session, so quit (`/exit`), start `codex` again, and check with `$tokeneconomizer:status`. +**GitHub Copilot in VS Code** + +See [VS Code setup on Windows](plugin/README.md#github-copilot-in-vs-code) for project hooks +with the **Copilot** session target, without GitHub CLI or standalone Copilot CLI. + **Copilot CLI** ```sh diff --git a/plugin/README.md b/plugin/README.md index 0395d66..ebfd659 100644 --- a/plugin/README.md +++ b/plugin/README.md @@ -179,6 +179,98 @@ codex plugin add tokeneconomizer@tokeneconomizer To update it, rebuild the bundle and run `codex plugin add tokeneconomizer@tokeneconomizer` again. +## GitHub Copilot in VS Code + +On Windows, use project [hooks](https://code.visualstudio.com/docs/agent-customization/hooks) +with **Copilot** selected as the chat session target. This setup needs Python >= 3.10, +PowerShell 7, and Copilot signed in inside VS Code; neither `gh` nor `copilot` is required. +It enables model-free micro-skills only. The older **Local** session target uses a different hook format. + +**1. Clone TokenEconomizer**, or use your existing checkout: + +```powershell +git clone https://github.com/microsoft/TokenEconomizer.git C:/dev/TokenEconomizer +``` + +**2. Configure your project.** In PowerShell 7, edit the first two paths and run this block. +Keep the checkout in place: the hooks reference it directly. Use this setup once per project; +do not also enable the packaged TokenEconomizer plugin there. + +```powershell +$teRoot = (Resolve-Path 'C:/dev/TokenEconomizer').Path +$projectRoot = (Resolve-Path 'C:/work/my-project').Path +$pythonExe = (Get-Command python -CommandType Application | + Where-Object Source -NotLike '*WindowsApps*' | Select-Object -First 1).Source +if (-not $pythonExe) { throw 'Set $pythonExe to a real Python 3.10+ executable' } +& $pythonExe --version # Must be a real Python 3.10+ installation. + +$bootstrap = Join-Path $teRoot 'plugin/scripts/te_bootstrap.py' +$teState = Join-Path $projectRoot '.te-vscode-state' +$hookDir = Join-Path $projectRoot '.github/hooks' +$hookFile = Join-Path $hookDir 'tokeneconomizer-vscode.json' +if (Test-Path -LiteralPath $hookFile) { + throw "TokenEconomizer hooks already exist: $hookFile" +} + +$previousTeState = $env:TOKENECONOMIZER_HOME +try { + $env:TOKENECONOMIZER_HOME = $teState + & $pythonExe $bootstrap cli --host copilot --cwd $projectRoot config set ` + microskills=envfacts toolprune=off loopguard=off repomap=off history_import=off + if ($LASTEXITCODE -ne 0) { throw 'Could not configure TokenEconomizer' } +} finally { + $env:TOKENECONOMIZER_HOME = $previousTeState +} + +$quotedPython = "'" + $pythonExe.Replace("'", "''") + "'" +$quotedBootstrap = "'" + $bootstrap.Replace("'", "''") + "'" +$command = "& $quotedPython $quotedBootstrap hook --host copilot" +$hooks = [ordered]@{} +foreach ($event in @('SessionStart', 'UserPromptSubmit', 'PostToolUse', + 'PostToolUseFailure', 'Stop', 'SessionEnd')) { + $hooks[$event] = @(@{ + type = 'command' + powershell = $command + timeoutSec = 60 + env = @{ TOKENECONOMIZER_HOME = $teState; PYTHONUTF8 = '1' } + }) +} +New-Item -ItemType Directory -Path $hookDir -Force | Out-Null +@{ version = 1; hooks = $hooks } | ConvertTo-Json -Depth 8 | + Set-Content -LiteralPath $hookFile -Encoding utf8NoBOM +``` + +Add `/.te-vscode-state/` and `/.github/hooks/tokeneconomizer-vscode.json` to the project's +local Git excludes (normally `.git/info/exclude`). They contain local state and absolute paths. + +**3. Check it.** Open that project in VS Code, select the **Copilot** session target and +**Agent**, then open **Chat: Open Customizations** > **Hooks** to check the workspace hooks. +Start a new chat and ask Copilot to read a project file. In the same PowerShell terminal: + +```powershell +$previousTeState = $env:TOKENECONOMIZER_HOME +try { + $env:TOKENECONOMIZER_HOME = $teState + & $pythonExe $bootstrap cli --host copilot --cwd $projectRoot status --explain +} finally { + $env:TOKENECONOMIZER_HOME = $previousTeState +} +``` + +Check for a session record under `.te-vscode-state/repos//copilot/sessions/`. +After saving your work and closing that project's VS Code window, its record should eventually +show `folded: true`. For hook errors, check `.te-vscode-state/errors.log`. + +Tested with VS Code 1.146.0 and the Copilot session target. Hooks, transcript processing, and +context injection were verified; natural fact learning still needs validation. Model-based +proposals, loop guard, and tool pruning stay disabled. This hook setup does not install slash commands. + +**Update.** Update the clone and start a new chat. If its path or the Python path changes, +remove the hook file and repeat step 2. + +**Uninstall.** Remove only `.github/hooks/tokeneconomizer-vscode.json` and start a fresh chat. +Keep `.te-vscode-state/` if you want to retain the learned state. + ## Copilot CLI **1. Install.** diff --git a/tests/test_trace_copilot_detection.py b/tests/test_trace_copilot_detection.py new file mode 100644 index 0000000..f45ff04 --- /dev/null +++ b/tests/test_trace_copilot_detection.py @@ -0,0 +1,27 @@ +"""Copilot metadata must not make its event stream look like a Claude transcript.""" +import json + +from tokeneconomizer.trace import call_cwds, load_trajectory, tool_usage + + +def test_copilot_tool_use_finish_reason_preserves_calls(tmp_path): + path = tmp_path / "events.jsonl" + events = [ + {"type": "session.start", "data": {"context": {"cwd": str(tmp_path)}}}, + {"type": "assistant.turn_end", "data": {"finishReason": "tool_use"}}, + {"type": "tool.execution_start", "data": { + "toolCallId": "read-1", "toolName": "view", "arguments": {"path": "fixture.txt"}}}, + {"type": "tool.execution_complete", "data": { + "toolCallId": "read-1", "success": True, "result": {"content": "amber"}}}, + {"type": "tool.execution_start", "data": { + "toolCallId": "read-2", "toolName": "view", "arguments": {"path": "missing.txt"}}}, + {"type": "tool.execution_complete", "data": { + "toolCallId": "read-2", "success": False, "result": {"content": "Path does not exist"}}}, + ] + path.write_text("\n".join(json.dumps(event) for event in events), encoding="utf-8") + assert load_trajectory(path) == [ + ("view", {"path": "fixture.txt"}, "amber", True), + ("view", {"path": "missing.txt"}, "Path does not exist", False), + ] + assert tool_usage(path) == {"view": 2} + assert call_cwds(path) == [str(tmp_path), str(tmp_path)] diff --git a/tokeneconomizer/trace/calls.py b/tokeneconomizer/trace/calls.py index b8ce150..1d85048 100644 --- a/tokeneconomizer/trace/calls.py +++ b/tokeneconomizer/trace/calls.py @@ -41,9 +41,11 @@ def load_trajectory(path): lines = [l.strip() for l in open(path, encoding="utf-8", errors="replace") if l.strip()] if _is_codex_exec(lines): return codex_exec_calls(lines) - if any('"tool_use"' in line or '"type": "assistant"' in line or '"type":"assistant"' in line - for line in lines): - return [c[:4] for c in claude_calls(lines)] + # Copilot can also say "tool_use" in metadata (for example a model's finish reason). + # Recognize actual Claude conversation blocks, not strings anywhere in the event log. + claude = claude_calls(lines) + if claude: + return [c[:4] for c in claude] return copilot_event_calls(lines)