dataforge/agents_and_robots
1
0
Fork 0
Code Issues Pull Requests 2 Actions Packages Projects Releases Wiki Activity
410 Commits 3 Branches 0 Tags
d1fd78324b37f5799282a0582294107e90c1e59a
Commit Graph

4 Commits

Author SHA1 Message Date
egutierrez d1fd78324b test(0145): unit + integration + launcher + claudecode coverage 
cmd/devicemesh-mcp/main_test.go (10 tests):
- TestInitialize: JSON-RPC initialize frame → serverInfo + capabilities.
- TestToolsList: tools/list → 16 user-mode entries, cada uno con name +
  inputSchema valido.
- TestToolsCallExec: tools/call name=exec → mock device-agent (httptest)
  recibe capability=shell.exec, MCP response content contiene "hi".
- TestToolsCallInvalidTool: name desconocido → isError o error envelope.
- TestNotificationsInitializedNoResponse: notification (sin id) → cero
  responses.
- TestUserModeFiltersPkgInstall: --mode user oculta pkg.install,
  --mode sudo la expone.
- TestToolsAllowedNarrows: --tools-allowed exec,fs.read → solo 2.
- TestSplitCSV, TestParseMode, TestIsCleanShutdown: helpers.

cmd/devicemesh-mcp/integration_test.go:
- TestIntegrationBinarySubprocess: build el binario en tmp + spawn como
  child via exec.Command + pipe real + secuencia initialize ->
  notifications/initialized -> tools/list -> tools/call. Valida el path
  identico al que usara claude.

devagents/mcp_bridge_test.go (9 tests):
- Disabled paths (nil DM, ExposeViaMCP=false, provider!=claude-code).
- Applied path: /tmp/<agent>-mcp-config.json JSON valido, mode 0600,
  mcpServers.devicemesh con command apuntando al binario fake.
- AllowedTools formato mcp__<server>__<tool>.
- DisableTools=true overrideado a false.
- URLEnv override gana sobre YAML.
- Binary missing → ok=false sin panico.
- BuildClaudeAllowedToolNames default server name.
- ResolveBridgedToolNames respeta mode + ToolsAllowed.
- ShouldExposeViaMCP cubre nil/disabled/default/explicit-true/false.

shell/llm/claudecode_test.go:
- TestBuildClaudeArgs_DisableTools actualizado: solo emite --tools "" cuando
  AllowedTools ESTA vacio. La regla nueva (issue 0145) da precedencia a
  AllowedTools.
- Anadido TestBuildClaudeArgs_DisableToolsButAllowedToolsWins.
- Anadido TestBuildClaudeArgs_MCPConfigPath.

bridge.go fix: cambio NewTool + WithRawInputSchema a NewToolWithRawSchema
porque NewTool inicializa ToolInputSchema.Type="object" por default, lo
cual entra en conflicto con RawInputSchema en MarshalJSON del SDK.

Suite completa pasa con -tags goolm -count=1.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
 2026-05-24 18:33:24 +02:00
egutierrez 1bdf9344a2 feat: streaming del subproceso claude-code con --output-format stream-json 
Implementa la Fase 1 del issue 0036: soporte de streaming en tiempo real
para el provider claude-code.

- Tipos puros de streaming en pkg/llm/types.go: StreamEventKind,
  StreamEvent, StreamFunc (pure core, sin side effects)
- Refactor de shell/llm/claudecode.go: nuevo code path executeStreaming
  que usa cmd.StdoutPipe + bufio.Scanner para leer linea a linea
- Parser parseStreamLine que mapea eventos JSON del CLI (system, assistant,
  result) a StreamEvent del dominio
- buildClaudeArgs ahora selecciona --output-format stream-json cuando
  streaming esta habilitado y StreamFunc presente
- Campos Streaming y ShowToolProgress en ClaudeCodeCfg (config schema)
- Backward compatible: streaming=false (default) no cambia comportamiento
- 40 tests (20 existentes + 20 nuevos) pasan sin errores

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-04-09 22:53:41 +00:00
egutierrez 6a5cad5700 test: extraer resolveWorkDir y tests unitarios de aislamiento 
Extraer la logica de resolucion de working_dir a una funcion
resolveWorkDir() separada para hacerla testeable. Tres tests cubren:
- WorkingDir vacio → crea tmpdir con prefijo claude-agent-*
- WorkingDir configurado → crea el directorio y lo usa
- WorkingDir ya existente → lo usa sin error

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
 2026-03-08 11:47:07 +00:00
egutierrez 61f4fee5d0 test: añadir tests para claude-code provider y router 
27 tests nuevos cubriendo las funciones del provider claude-code:

- buildClaudeArgs: minimal, all options, disable_tools, disallowed_tools
- flattenMessages: empty, multi-role, skips system messages
- parseClaudeOutput: success, error response, process failed (con/sin stderr),
  fallback a plain text, content blocks, exec error con stdout parcial
- filterEnv: single key, multiple keys, no match, prefix safety
- Route: claude-code, claude-code/custom, claude-*, gpt-*, ollama/*, default
- ModelName: ollama prefix strip, passthrough

Todos pasan con 'go test -tags goolm ./shell/llm/ ./pkg/llm/'.
 2026-03-06 22:14:43 +00:00