karajan-code 1.35.0 → 1.36.1

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/docs/README.es.md CHANGED
@@ -29,8 +29,9 @@ En lugar de ejecutar un agente de IA y revisar manualmente su output, `kj` encad
29
29
 
30
30
  **Caracteristicas principales:**
31
31
  - **Pipeline multi-agente** con 11 roles configurables
32
- - **4 agentes de IA soportados**: Claude, Codex, Gemini, Aider
33
- - **Servidor MCP** con 15 herramientas — usa `kj` desde Claude, Codex o cualquier host compatible con MCP sin salir de tu agente. [Ver configuracion MCP](#servidor-mcp)
32
+ - **5 agentes de IA soportados**: Claude, Codex, Gemini, Aider, OpenCode
33
+ - **Servidor MCP** con 20 herramientas — usa `kj` desde Claude, Codex o cualquier host compatible con MCP sin salir de tu agente. [Ver configuracion MCP](#servidor-mcp)
34
+ - **Bootstrap obligatorio** — valida prerequisitos del entorno (git, remote, config, agentes, SonarQube) antes de cada ejecucion. Si algo falta, para con instrucciones claras
34
35
  - **TDD obligatorio** — se exigen cambios en tests cuando se modifican ficheros fuente
35
36
  - **Integracion con SonarQube** — analisis estatico con quality gates (requiere [Docker](#requisitos))
36
37
  - **Perfiles de revision** — standard, strict, relaxed, paranoid
@@ -128,33 +129,135 @@ Guias completas: [`docs/multi-instance.md`](multi-instance.md) | [`docs/install-
128
129
 
129
130
  `kj init` auto-detecta los agentes instalados. Si solo hay uno disponible, se asigna a todos los roles automaticamente.
130
131
 
131
- ## Inicio rapido
132
+ ## Tres formas de usar Karajan
133
+
134
+ Karajan instala **tres comandos**: `kj`, `kj-tail` y `karajan-mcp`.
135
+
136
+ ### 1. CLI — Directamente desde terminal
132
137
 
133
138
  ```bash
134
- # Ejecutar una tarea con defaults (claude=coder, codex=reviewer, TDD)
135
139
  kj run "Implementar autenticacion de usuario con JWT"
136
-
137
- # Solo coder (sin revision)
138
140
  kj code "Anadir validacion de inputs al formulario de registro"
139
-
140
- # Solo reviewer (revisar diff actual)
141
141
  kj review "Revisar los cambios de autenticacion"
142
+ kj plan "Refactorizar la capa de base de datos"
143
+ ```
142
144
 
143
- # Generar un plan de implementacion
144
- kj plan "Refactorizar la capa de base de datos para usar connection pooling"
145
+ ### 2. MCP Dentro de tu agente de IA
145
146
 
146
- # Pipeline completo con todas las opciones
147
- kj run "Corregir inyeccion SQL critica en el endpoint de busqueda" \
148
- --coder claude \
149
- --reviewer codex \
150
- --reviewer-fallback claude \
151
- --methodology tdd \
152
- --enable-triage \
153
- --enable-tester \
154
- --enable-security \
155
- --auto-commit \
156
- --auto-push \
157
- --max-iterations 5
147
+ El caso de uso principal. Karajan corre como servidor MCP dentro de Claude Code, Codex o Gemini. El agente tiene acceso a 20 herramientas (`kj_run`, `kj_code`, `kj_review`, etc.) y delega el trabajo pesado al pipeline de Karajan.
148
+
149
+ ```
150
+ Tu → Claude Code → kj_run (via MCP) → triage → coder → sonar → reviewer tester → security
151
+ ```
152
+
153
+ **El problema**: cuando Karajan corre dentro de un agente de IA, pierdes visibilidad. El agente te muestra el resultado final, pero no las etapas del pipeline, iteraciones o decisiones de Solomon en tiempo real.
154
+
155
+ ### 3. kj-tail — Monitorizar desde otro terminal
156
+
157
+ **La herramienta companera.** Abre un segundo terminal en el **mismo directorio del proyecto** donde esta trabajando tu agente de IA:
158
+
159
+ ```bash
160
+ kj-tail
161
+ ```
162
+
163
+ Veras la salida del pipeline en vivo — etapas, resultados, iteraciones, errores — tal como ocurren.
164
+
165
+ ```bash
166
+ kj-tail # Seguir pipeline en tiempo real (por defecto)
167
+ kj-tail -v # Verbose: incluir heartbeats de agente y presupuesto
168
+ kj-tail -t # Mostrar timestamps
169
+ kj-tail -s # Snapshot: mostrar log actual y salir
170
+ kj-tail -n 50 # Mostrar ultimas 50 lineas y seguir
171
+ kj-tail --help # Todas las opciones
172
+ ```
173
+
174
+ > **Importante**: `kj-tail` debe ejecutarse desde el mismo directorio donde el agente de IA esta trabajando. Lee `<proyecto>/.kj/run.log`, que se crea cuando Karajan arranca un pipeline via MCP.
175
+
176
+ **Flujo tipico:**
177
+
178
+ ```
179
+ ┌──────────────────────────┐ ┌──────────────────────────┐
180
+ │ Terminal 1 │ │ Terminal 2 │
181
+ │ │ │ │
182
+ │ $ claude │ │ $ kj-tail │
183
+ │ > implementa la │ │ │
184
+ │ siguiente tarea │ │ ├─ 📋 Triage: medium │
185
+ │ prioritaria │ │ ├─ 🔬 Researcher ✅ │
186
+ │ │ │ ├─ 🧠 Planner ✅ │
187
+ │ (Claude llama a kj_run │ │ ├─ 🔨 Coder ✅ │
188
+ │ via MCP — solo ves │ │ ├─ 🔍 Sonar: OK │
189
+ │ el resultado final) │ │ ├─ 👁️ Reviewer ❌ │
190
+ │ │ │ ├─ ⚖️ Solomon: 2 cond. │
191
+ │ │ │ ├─ 🔨 Coder (iter 2) ✅ │
192
+ │ │ │ ├─ ✅ Review: APPROVED │
193
+ │ │ │ ├─ 🧪 Tester: passed │
194
+ │ │ │ └─ 🏁 Result: APPROVED │
195
+ └──────────────────────────┘ └──────────────────────────┘
196
+ ```
197
+
198
+ **Ejemplo con pipeline completo** — tarea compleja con todos los roles:
199
+
200
+ ```
201
+ ┌─ Terminal 1 ─────────────────────────────────────────────────────────────────┐
202
+ │ │
203
+ │ $ claude │
204
+ │ │
205
+ │ > Construye una API REST para un sistema de reservas. Requisitos: │
206
+ │ > - Express + TypeScript con validacion Zod en cada endpoint │
207
+ │ > - Endpoints: POST /bookings, GET /bookings/:id, │
208
+ │ > PATCH /bookings/:id/cancel │
209
+ │ > - Una reserva tiene: id, guestName, roomType (standard|suite|penthouse), │
210
+ │ > checkIn, checkOut, status (confirmed|cancelled) │
211
+ │ > - Validar: checkOut posterior a checkIn, sin fechas pasadas, │
212
+ │ > roomType debe ser un valor valido del enum │
213
+ │ > - Cancelar devuelve 409 si ya esta cancelada │
214
+ │ > - Usa TDD. Ejecutalo con Karajan con architect y planner activos, │
215
+ │ > modo paranoid. Coder claude, reviewer codex. │
216
+ │ │
217
+ │ Claude llama a kj_run via MCP con: │
218
+ │ --enable-architect --enable-researcher --enable-planner --mode paranoid │
219
+ │ │
220
+ └──────────────────────────────────────────────────────────────────────────────┘
221
+
222
+ ┌─ Terminal 2: kj-tail ────────────────────────────────────────────────────────┐
223
+ │ │
224
+ │ kj-tail v1.36.1 — .kj/run.log │
225
+ │ │
226
+ │ ├─ 📋 Triage: medium (sw) — activando researcher, architect, planner │
227
+ │ ├─ ⚙️ Preflight passed — all checks OK │
228
+ │ ├─ 🔬 Researcher: 8 ficheros, 3 patrones, 5 restricciones │
229
+ │ ├─ 🏗️ Architect: diseno 3 capas (routes → service → validators) │
230
+ │ ├─ 🧠 Planner: 6 pasos — tests primero, luego rutas, servicio, validadores │
231
+ │ │ │
232
+ │ ▶ Iteracion 1/5 │
233
+ │ ├─ 🔨 Coder (claude): 3 endpoints + 18 tests │
234
+ │ ├─ 📋 TDD: PASS (3 src, 2 test) │
235
+ │ ├─ 🔍 Sonar: Quality gate OK │
236
+ │ ├─ 👁️ Reviewer (codex): REJECTED (2 blocking) │
237
+ │ │ "Falta 404 para GET booking inexistente" │
238
+ │ │ "Endpoint cancel sin test de idempotencia" │
239
+ │ ├─ ⚖️ Solomon: approve_with_conditions (2 condiciones) │
240
+ │ │ "Anadir respuesta 404 y test para GET /bookings/:id con id desconocido" │
241
+ │ │ "Anadir test: cancelar reserva ya cancelada devuelve 409, no 500" │
242
+ │ │ │
243
+ │ ▶ Iteracion 2/5 │
244
+ │ ├─ 🔨 Coder (claude): corregido — 22 tests │
245
+ │ ├─ 📋 TDD: PASS │
246
+ │ ├─ 🔍 Sonar: OK │
247
+ │ ├─ 👁️ Reviewer (codex): APPROVED │
248
+ │ ├─ 🧪 Tester: passed — cobertura 94%, 22 tests │
249
+ │ ├─ 🔒 Security: passed — 0 criticos, 1 bajo (helmet recomendado) │
250
+ │ ├─ 📊 Audit: CERTIFIED (3 advertencias) │
251
+ │ │ │
252
+ │ 🏁 Resultado: APPROVED │
253
+ │ 🔬 Investigacion: 8 ficheros, 3 patrones │
254
+ │ 🗺 Plan: 6 pasos (tests primero) │
255
+ │ 🧪 Cobertura: 94%, 22 tests │
256
+ │ 🔒 Seguridad: OK │
257
+ │ 🔍 Sonar: OK │
258
+ │ 💰 Presupuesto: $0.42 (claude: $0.38, codex: $0.04) │
259
+ │ │
260
+ └──────────────────────────────────────────────────────────────────────────────┘
158
261
  ```
159
262
 
160
263
  ## Comandos CLI
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "karajan-code",
3
- "version": "1.35.0",
3
+ "version": "1.36.1",
4
4
  "description": "Local multi-agent coding orchestrator with TDD, SonarQube, and code review pipeline",
5
5
  "type": "module",
6
6
  "license": "AGPL-3.0",
@@ -38,6 +38,7 @@
38
38
  ],
39
39
  "bin": {
40
40
  "kj": "bin/kj.js",
41
+ "kj-tail": "bin/kj-tail",
41
42
  "karajan-mcp": "bin/karajan-mcp.js"
42
43
  },
43
44
  "scripts": {
@@ -5,21 +5,28 @@ import { resolveBin } from "./resolve-bin.js";
5
5
  export class AiderAgent extends BaseAgent {
6
6
  async runTask(task) {
7
7
  const role = task.role || "coder";
8
- const args = ["--yes", "--message", task.prompt];
9
8
  const model = this.getRoleModel(role);
10
- if (model) args.push("--model", model);
11
- const res = await runCommand(resolveBin("aider"), args, {
12
- onOutput: task.onOutput,
13
- silenceTimeoutMs: task.silenceTimeoutMs,
14
- timeout: task.timeoutMs
15
- });
16
- return { ok: res.exitCode === 0, output: res.stdout, error: res.stderr, exitCode: res.exitCode };
9
+ const result = await this._exec(task, model);
10
+ if (!result.ok && model && this.isModelNotSupportedError(result)) {
11
+ this.logger?.warn(`Aider model "${model}" not supported — retrying with agent default`);
12
+ return this._exec(task, null);
13
+ }
14
+ return result;
17
15
  }
18
16
 
19
17
  async reviewTask(task) {
20
18
  const role = task.role || "reviewer";
21
- const args = ["--yes", "--message", task.prompt];
22
19
  const model = this.getRoleModel(role);
20
+ const result = await this._exec(task, model);
21
+ if (!result.ok && model && this.isModelNotSupportedError(result)) {
22
+ this.logger?.warn(`Aider model "${model}" not supported — retrying with agent default`);
23
+ return this._exec(task, null);
24
+ }
25
+ return result;
26
+ }
27
+
28
+ async _exec(task, model) {
29
+ const args = ["--yes", "--message", task.prompt];
23
30
  if (model) args.push("--model", model);
24
31
  const res = await runCommand(resolveBin("aider"), args, {
25
32
  onOutput: task.onOutput,
@@ -1,3 +1,12 @@
1
+ const MODEL_NOT_SUPPORTED_PATTERNS = [
2
+ /model.{0,30}is not supported/i,
3
+ /model.{0,30}not available/i,
4
+ /model.{0,30}does not exist/i,
5
+ /unsupported model/i,
6
+ /invalid model/i,
7
+ /model_not_found/i
8
+ ];
9
+
1
10
  export class BaseAgent {
2
11
  constructor(name, config, logger) {
3
12
  this.name = name;
@@ -24,4 +33,10 @@ export class BaseAgent {
24
33
  if (role === "reviewer") return false;
25
34
  return Boolean(this.config?.coder_options?.auto_approve);
26
35
  }
36
+
37
+ isModelNotSupportedError(result) {
38
+ const text = [result?.error, result?.output, result?.stderr, result?.stdout]
39
+ .filter(Boolean).join("\n");
40
+ return MODEL_NOT_SUPPORTED_PATTERNS.some(re => re.test(text));
41
+ }
27
42
  }
@@ -35,6 +35,29 @@ function collectAssistantText(obj) {
35
35
  .map(block => block.text);
36
36
  }
37
37
 
38
+ /**
39
+ * Extract usage metrics from stream-json/json NDJSON output.
40
+ * Looks for the "result" line which contains total_cost_usd,
41
+ * usage.input_tokens/output_tokens, and modelUsage.
42
+ * Returns an object with tokens_in, tokens_out, cost_usd, model or null if not found.
43
+ */
44
+ export function extractUsageFromStreamJson(raw) {
45
+ const lines = (raw || "").split("\n").filter(Boolean);
46
+ for (let i = lines.length - 1; i >= 0; i--) {
47
+ const obj = tryParseJson(lines[i]);
48
+ if (!obj || obj.type !== "result") continue;
49
+
50
+ const tokens_in = obj.usage?.input_tokens ?? 0;
51
+ const tokens_out = obj.usage?.output_tokens ?? 0;
52
+ const cost_usd = obj.total_cost_usd ?? undefined;
53
+ const modelUsage = obj.modelUsage;
54
+ const model = modelUsage ? Object.keys(modelUsage)[0] || null : null;
55
+
56
+ return { tokens_in, tokens_out, cost_usd, model };
57
+ }
58
+ return null;
59
+ }
60
+
38
61
  /**
39
62
  * Extract the final text result from stream-json NDJSON output.
40
63
  * Each line is a JSON object. We collect assistant text content from
@@ -135,8 +158,28 @@ const ALLOWED_TOOLS = [
135
158
  export class ClaudeAgent extends BaseAgent {
136
159
  async runTask(task) {
137
160
  const role = task.role || "coder";
138
- const args = ["-p", task.prompt, "--allowedTools", ...ALLOWED_TOOLS];
139
161
  const model = this.getRoleModel(role);
162
+ const result = await this._runTaskExec(task, model, role);
163
+ if (!result.ok && model && this.isModelNotSupportedError(result)) {
164
+ this.logger?.warn(`Claude model "${model}" not supported — retrying with agent default`);
165
+ return this._runTaskExec(task, null, role);
166
+ }
167
+ return result;
168
+ }
169
+
170
+ async reviewTask(task) {
171
+ const role = task.role || "reviewer";
172
+ const model = this.getRoleModel(role);
173
+ const result = await this._reviewTaskExec(task, model);
174
+ if (!result.ok && model && this.isModelNotSupportedError(result)) {
175
+ this.logger?.warn(`Claude model "${model}" not supported — retrying with agent default`);
176
+ return this._reviewTaskExec(task, null);
177
+ }
178
+ return result;
179
+ }
180
+
181
+ async _runTaskExec(task, model, role) {
182
+ const args = ["-p", task.prompt, "--allowedTools", ...ALLOWED_TOOLS];
140
183
  if (model) args.push("--model", model);
141
184
 
142
185
  // Use stream-json when onOutput is provided to get real-time feedback
@@ -150,7 +193,8 @@ export class ClaudeAgent extends BaseAgent {
150
193
  }));
151
194
  const raw = pickOutput(res);
152
195
  const output = extractTextFromStreamJson(raw);
153
- return { ok: res.exitCode === 0, output, error: res.exitCode === 0 ? "" : raw, exitCode: res.exitCode };
196
+ const usage = extractUsageFromStreamJson(raw);
197
+ return { ok: res.exitCode === 0, output, error: res.exitCode === 0 ? "" : raw, exitCode: res.exitCode, ...usage };
154
198
  }
155
199
 
156
200
  // Without streaming, use json output to get structured response via stderr
@@ -158,12 +202,12 @@ export class ClaudeAgent extends BaseAgent {
158
202
  const res = await runCommand(resolveBin("claude"), args, cleanExecaOpts());
159
203
  const raw = pickOutput(res);
160
204
  const output = extractTextFromStreamJson(raw);
161
- return { ok: res.exitCode === 0, output, error: res.exitCode === 0 ? "" : raw, exitCode: res.exitCode };
205
+ const usage = extractUsageFromStreamJson(raw);
206
+ return { ok: res.exitCode === 0, output, error: res.exitCode === 0 ? "" : raw, exitCode: res.exitCode, ...usage };
162
207
  }
163
208
 
164
- async reviewTask(task) {
209
+ async _reviewTaskExec(task, model) {
165
210
  const args = ["-p", task.prompt, "--allowedTools", ...ALLOWED_TOOLS, "--output-format", "stream-json", "--verbose"];
166
- const model = this.getRoleModel(task.role || "reviewer");
167
211
  if (model) args.push("--model", model);
168
212
  const res = await runCommand(resolveBin("claude"), args, cleanExecaOpts({
169
213
  onOutput: task.onOutput,
@@ -171,6 +215,7 @@ export class ClaudeAgent extends BaseAgent {
171
215
  timeout: task.timeoutMs
172
216
  }));
173
217
  const raw = pickOutput(res);
174
- return { ok: res.exitCode === 0, output: raw, error: res.exitCode === 0 ? "" : raw, exitCode: res.exitCode };
218
+ const usage = extractUsageFromStreamJson(raw);
219
+ return { ok: res.exitCode === 0, output: raw, error: res.exitCode === 0 ? "" : raw, exitCode: res.exitCode, ...usage };
175
220
  }
176
221
  }
@@ -2,27 +2,48 @@ import { BaseAgent } from "./base-agent.js";
2
2
  import { runCommand } from "../utils/process.js";
3
3
  import { resolveBin } from "./resolve-bin.js";
4
4
 
5
+ /**
6
+ * Extract token usage from Codex CLI stdout.
7
+ * Codex prints "tokens used\n<number>" at the end, where number may have comma separators.
8
+ * Returns { tokens_out } with the total token count, or null if not found.
9
+ * Since Codex doesn't split input/output, we assign the total to tokens_out
10
+ * as a conservative estimate for cost calculation.
11
+ */
12
+ export function extractCodexTokens(stdout) {
13
+ const match = (stdout || "").match(/tokens?\s+used\s*\n\s*([\d,]+)/i);
14
+ if (!match) return null;
15
+ const total = Number(match[1].replace(/,/g, ""));
16
+ if (!Number.isFinite(total) || total <= 0) return null;
17
+ return { tokens_in: 0, tokens_out: total };
18
+ }
19
+
5
20
  export class CodexAgent extends BaseAgent {
6
21
  async runTask(task) {
7
22
  const role = task.role || "coder";
8
- const args = ["exec"];
9
23
  const model = this.getRoleModel(role);
10
- if (model) args.push("--model", model);
11
- if (this.isAutoApproveEnabled(role)) args.push("--full-auto");
12
- args.push("-");
13
- const res = await runCommand(resolveBin("codex"), args, {
14
- onOutput: task.onOutput,
15
- silenceTimeoutMs: task.silenceTimeoutMs,
16
- timeout: task.timeoutMs,
17
- input: task.prompt
18
- });
19
- return { ok: res.exitCode === 0, output: res.stdout, error: res.stderr, exitCode: res.exitCode };
24
+ const result = await this._exec(task, model, role);
25
+ if (!result.ok && model && this.isModelNotSupportedError(result)) {
26
+ this.logger?.warn(`Codex model "${model}" not supported — retrying with agent default`);
27
+ return this._exec(task, null, role);
28
+ }
29
+ return result;
20
30
  }
21
31
 
22
32
  async reviewTask(task) {
33
+ const role = task.role || "reviewer";
34
+ const model = this.getRoleModel(role);
35
+ const result = await this._exec(task, model, role);
36
+ if (!result.ok && model && this.isModelNotSupportedError(result)) {
37
+ this.logger?.warn(`Codex model "${model}" not supported — retrying with agent default`);
38
+ return this._exec(task, null, role);
39
+ }
40
+ return result;
41
+ }
42
+
43
+ async _exec(task, model, role) {
23
44
  const args = ["exec"];
24
- const model = this.getRoleModel(task.role || "reviewer");
25
45
  if (model) args.push("--model", model);
46
+ if (role !== "reviewer" && this.isAutoApproveEnabled(role)) args.push("--full-auto");
26
47
  args.push("-");
27
48
  const res = await runCommand(resolveBin("codex"), args, {
28
49
  onOutput: task.onOutput,
@@ -30,6 +51,7 @@ export class CodexAgent extends BaseAgent {
30
51
  timeout: task.timeoutMs,
31
52
  input: task.prompt
32
53
  });
33
- return { ok: res.exitCode === 0, output: res.stdout, error: res.stderr, exitCode: res.exitCode };
54
+ const usage = extractCodexTokens(res.stdout);
55
+ return { ok: res.exitCode === 0, output: res.stdout, error: res.stderr, exitCode: res.exitCode, ...usage };
34
56
  }
35
57
  }
@@ -5,21 +5,29 @@ import { resolveBin } from "./resolve-bin.js";
5
5
  export class GeminiAgent extends BaseAgent {
6
6
  async runTask(task) {
7
7
  const role = task.role || "coder";
8
- const args = ["-p", task.prompt];
9
8
  const model = this.getRoleModel(role);
10
- if (model) args.push("--model", model);
11
- const res = await runCommand(resolveBin("gemini"), args, {
12
- onOutput: task.onOutput,
13
- silenceTimeoutMs: task.silenceTimeoutMs,
14
- timeout: task.timeoutMs
15
- });
16
- return { ok: res.exitCode === 0, output: res.stdout, error: res.stderr, exitCode: res.exitCode };
9
+ const result = await this._exec(task, model, "run");
10
+ if (!result.ok && model && this.isModelNotSupportedError(result)) {
11
+ this.logger?.warn(`Gemini model "${model}" not supported — retrying with agent default`);
12
+ return this._exec(task, null, "run");
13
+ }
14
+ return result;
17
15
  }
18
16
 
19
17
  async reviewTask(task) {
20
18
  const role = task.role || "reviewer";
21
- const args = ["-p", task.prompt, "--output-format", "json"];
22
19
  const model = this.getRoleModel(role);
20
+ const result = await this._exec(task, model, "review");
21
+ if (!result.ok && model && this.isModelNotSupportedError(result)) {
22
+ this.logger?.warn(`Gemini model "${model}" not supported — retrying with agent default`);
23
+ return this._exec(task, null, "review");
24
+ }
25
+ return result;
26
+ }
27
+
28
+ async _exec(task, model, mode) {
29
+ const args = ["-p", task.prompt];
30
+ if (mode === "review") args.push("--output-format", "json");
23
31
  if (model) args.push("--model", model);
24
32
  const res = await runCommand(resolveBin("gemini"), args, {
25
33
  onOutput: task.onOutput,
@@ -49,15 +49,16 @@ export function buildDefaultPricingTable() {
49
49
  }
50
50
 
51
51
  // Auto-register built-in models
52
+ // Names must match what each CLI accepts as --model argument
52
53
  registerModel("claude", { provider: "anthropic", pricing: { input_per_million: 3, output_per_million: 15 } });
53
- registerModel("claude/sonnet", { provider: "anthropic", pricing: { input_per_million: 3, output_per_million: 15 } });
54
- registerModel("claude/opus", { provider: "anthropic", pricing: { input_per_million: 15, output_per_million: 75 } });
55
- registerModel("claude/haiku", { provider: "anthropic", pricing: { input_per_million: 0.25, output_per_million: 1.25 } });
54
+ registerModel("sonnet", { provider: "anthropic", pricing: { input_per_million: 3, output_per_million: 15 } });
55
+ registerModel("opus", { provider: "anthropic", pricing: { input_per_million: 15, output_per_million: 75 } });
56
+ registerModel("haiku", { provider: "anthropic", pricing: { input_per_million: 0.25, output_per_million: 1.25 } });
56
57
  registerModel("codex", { provider: "openai", pricing: { input_per_million: 1.5, output_per_million: 4 } });
57
- registerModel("codex/o4-mini", { provider: "openai", pricing: { input_per_million: 1.5, output_per_million: 4 } });
58
- registerModel("codex/o3", { provider: "openai", pricing: { input_per_million: 10, output_per_million: 40 } });
58
+ registerModel("o4-mini", { provider: "openai", pricing: { input_per_million: 1.5, output_per_million: 4 } });
59
+ registerModel("o3", { provider: "openai", pricing: { input_per_million: 10, output_per_million: 40 } });
59
60
  registerModel("gemini", { provider: "google", pricing: { input_per_million: 1.25, output_per_million: 5 } });
60
- registerModel("gemini/pro", { provider: "google", pricing: { input_per_million: 1.25, output_per_million: 5 } });
61
- registerModel("gemini/flash", { provider: "google", pricing: { input_per_million: 0.075, output_per_million: 0.3 } });
61
+ registerModel("gemini-2.5-pro", { provider: "google", pricing: { input_per_million: 1.25, output_per_million: 5 } });
62
+ registerModel("gemini-2.0-flash", { provider: "google", pricing: { input_per_million: 0.075, output_per_million: 0.3 } });
62
63
  registerModel("aider", { provider: "aider", pricing: { input_per_million: 3, output_per_million: 15 } });
63
64
  registerModel("opencode", { provider: "opencode", pricing: { input_per_million: 0, output_per_million: 0 } });
@@ -5,22 +5,29 @@ import { resolveBin } from "./resolve-bin.js";
5
5
  export class OpenCodeAgent extends BaseAgent {
6
6
  async runTask(task) {
7
7
  const role = task.role || "coder";
8
- const args = ["run"];
9
8
  const model = this.getRoleModel(role);
10
- if (model) args.push("--model", model);
11
- args.push(task.prompt);
12
- const res = await runCommand(resolveBin("opencode"), args, {
13
- onOutput: task.onOutput,
14
- silenceTimeoutMs: task.silenceTimeoutMs,
15
- timeout: task.timeoutMs
16
- });
17
- return { ok: res.exitCode === 0, output: res.stdout, error: res.stderr, exitCode: res.exitCode };
9
+ const result = await this._exec(task, model, false);
10
+ if (!result.ok && model && this.isModelNotSupportedError(result)) {
11
+ this.logger?.warn(`OpenCode model "${model}" not supported — retrying with agent default`);
12
+ return this._exec(task, null, false);
13
+ }
14
+ return result;
18
15
  }
19
16
 
20
17
  async reviewTask(task) {
21
18
  const role = task.role || "reviewer";
22
- const args = ["run", "--format", "json"];
23
19
  const model = this.getRoleModel(role);
20
+ const result = await this._exec(task, model, true);
21
+ if (!result.ok && model && this.isModelNotSupportedError(result)) {
22
+ this.logger?.warn(`OpenCode model "${model}" not supported — retrying with agent default`);
23
+ return this._exec(task, null, true);
24
+ }
25
+ return result;
26
+ }
27
+
28
+ async _exec(task, model, jsonFormat) {
29
+ const args = ["run"];
30
+ if (jsonFormat) args.push("--format", "json");
24
31
  if (model) args.push("--model", model);
25
32
  args.push(task.prompt);
26
33
  const res = await runCommand(resolveBin("opencode"), args, {
@@ -1,5 +1,7 @@
1
1
  import { readFileSync } from "node:fs";
2
+ import fs from "node:fs/promises";
2
3
  import { fileURLToPath } from "node:url";
4
+ import os from "node:os";
3
5
  import path from "node:path";
4
6
  import { runCommand } from "../utils/process.js";
5
7
  import { exists } from "../utils/fs.js";
@@ -239,6 +241,101 @@ async function checkRuleFiles(config) {
239
241
  ];
240
242
  }
241
243
 
244
+ /**
245
+ * Detect duplicate TOML table headers (e.g. [mcp_servers."karajan-mcp"] appearing twice).
246
+ * Full TOML parsing would require a dependency — this catches the most common config error.
247
+ */
248
+ function findDuplicateTomlKeys(content) {
249
+ const tableHeaders = [];
250
+ const duplicates = [];
251
+ for (const line of content.split("\n")) {
252
+ const match = line.match(/^\s*\[([^\]]+)\]\s*$/);
253
+ if (match) {
254
+ const key = match[1].trim();
255
+ if (tableHeaders.includes(key)) {
256
+ duplicates.push(key);
257
+ } else {
258
+ tableHeaders.push(key);
259
+ }
260
+ }
261
+ }
262
+ return duplicates;
263
+ }
264
+
265
+ async function checkAgentConfigs() {
266
+ const checks = [];
267
+ const home = os.homedir();
268
+
269
+ // Claude: ~/.claude.json
270
+ const claudeJsonPath = path.join(home, ".claude.json");
271
+ try {
272
+ const raw = await fs.readFile(claudeJsonPath, "utf8");
273
+ JSON.parse(raw);
274
+ checks.push({ name: "agent-config:claude", label: "Agent config: claude (~/.claude.json)", ok: true, detail: "Valid JSON", fix: null });
275
+ } catch (err) {
276
+ if (err.code === "ENOENT") {
277
+ // File doesn't exist — not an error, Claude may not be configured
278
+ } else {
279
+ checks.push({
280
+ name: "agent-config:claude",
281
+ label: "Agent config: claude (~/.claude.json)",
282
+ ok: false,
283
+ detail: `Invalid JSON: ${err.message.split("\n")[0]}`,
284
+ fix: "Fix the JSON syntax in ~/.claude.json. Common issues: trailing commas, missing quotes."
285
+ });
286
+ }
287
+ }
288
+
289
+ // Codex: ~/.codex/config.toml
290
+ const codexTomlPath = path.join(home, ".codex", "config.toml");
291
+ try {
292
+ const raw = await fs.readFile(codexTomlPath, "utf8");
293
+ const duplicates = findDuplicateTomlKeys(raw);
294
+ if (duplicates.length > 0) {
295
+ checks.push({
296
+ name: "agent-config:codex",
297
+ label: "Agent config: codex (~/.codex/config.toml)",
298
+ ok: false,
299
+ detail: `Duplicate TOML keys: ${duplicates.join(", ")}`,
300
+ fix: `Remove duplicate entries in ~/.codex/config.toml: ${duplicates.join(", ")}`
301
+ });
302
+ } else {
303
+ checks.push({ name: "agent-config:codex", label: "Agent config: codex (~/.codex/config.toml)", ok: true, detail: "Valid TOML (no duplicate keys)", fix: null });
304
+ }
305
+ } catch (err) {
306
+ if (err.code !== "ENOENT") {
307
+ checks.push({
308
+ name: "agent-config:codex",
309
+ label: "Agent config: codex (~/.codex/config.toml)",
310
+ ok: false,
311
+ detail: `Cannot read: ${err.message.split("\n")[0]}`,
312
+ fix: "Check file permissions on ~/.codex/config.toml"
313
+ });
314
+ }
315
+ }
316
+
317
+ // KJ config: ~/.karajan/kj.config.yml (validate YAML)
318
+ const kjConfigPath = getConfigPath();
319
+ try {
320
+ const raw = await fs.readFile(kjConfigPath, "utf8");
321
+ const yaml = await import("js-yaml");
322
+ yaml.default.load(raw);
323
+ checks.push({ name: "agent-config:karajan", label: "Agent config: karajan (kj.config.yml)", ok: true, detail: "Valid YAML", fix: null });
324
+ } catch (err) {
325
+ if (err.code !== "ENOENT") {
326
+ checks.push({
327
+ name: "agent-config:karajan",
328
+ label: "Agent config: karajan (kj.config.yml)",
329
+ ok: false,
330
+ detail: `Invalid YAML: ${err.message.split("\n")[0]}`,
331
+ fix: `Fix YAML syntax in ${kjConfigPath}. Run 'kj init' to regenerate if needed.`
332
+ });
333
+ }
334
+ }
335
+
336
+ return checks;
337
+ }
338
+
242
339
  export async function runChecks({ config }) {
243
340
  const checks = [];
244
341
 
@@ -260,6 +357,7 @@ export async function runChecks({ config }) {
260
357
  checks.push(...await checkBecariaInfra(config));
261
358
  }
262
359
 
360
+ checks.push(...await checkAgentConfigs());
263
361
  checks.push(...await checkRuleFiles(config));
264
362
  checks.push(await checkRtk());
265
363