feat(forge): M16 skills + M18 subagents in-process + fix read_file cap - #29
Open
turbillon50 wants to merge 3 commits into
Open
feat(forge): M16 skills + M18 subagents in-process + fix read_file cap#29turbillon50 wants to merge 3 commits into
turbillon50 wants to merge 3 commits into
Conversation
V ahora elige modelo según task kind y costo, cae automáticamente al
siguiente del cascade si el provider falla (402/429/5xx), y expone su
consumo en tiempo real. Cubre lo que Luis pidió: "mezclar gratuitas
con paga, que vea consumo en vForge en tiempo real, que se auto-analize
si OpenRouter va".
NUEVO:
lib/forge/models.ts (registry):
Catálogo TS de 8 modelos OpenRouter con tier, kind, costo per 1M,
context window, soporta tools, y fallback_chain ordenado por
conveniencia. Source of truth para routing + costo.
- Premium: anthropic/claude-opus-4.7
- Balanced: anthropic/claude-sonnet-4.6, google/gemini-2.5-pro, openai/gpt-5
- Cheap (paid): anthropic/claude-haiku-4.5, google/gemini-2.5-flash
- Free: minimax/minimax-m2.5:free, google/gemma-4-31b-it:free
normalizeSlug() resuelve los slugs versionados que devuelve OR
(ej. 'anthropic/claude-4.6-sonnet-20260217' → 'anthropic/claude-sonnet-4.6').
lib/forge/routing.ts (policy):
routeFor(task, { costPreference?, excludeSlugs?, forceSlug? }) →
{ primary, cascade, reason }. TASK_PREFERENCES table mapea kind →
modelos preferidos en orden. costPreference: cheapest | balanced |
premium | free-only. suggestTierForBudget() sugiere tier según
saldo restante (M9.5 lo usará para self-throttle automático).
app/api/forge/cost/route.ts (NUEVO):
GET /api/forge/cost?period=today|24h|this_month|last_7d|last_30d
Agrega conversations.cost_usd + tokens + by_model + by_day +
fallback_events count. Cero auth en MVP (Clerk M11).
app/api/admin/health/route.ts (NUEVO):
GET /api/admin/health → { db, vault (con self-test del pepper),
openrouter (balance, usage, latency, status) }. Responde 503 si
alguno está roto. Activity dashboard + futura cron Trigger.dev.
MODIFICADO:
app/api/forge/run/route.ts:
- Importa routeFor + estimateCostForModel + normalizeSlug.
- Reemplaza LEGACY_MODEL_MAP por routeFor("chat-main", { forceSlug })
cuando system_config.default_model está en la registry; si no,
free-form override + cascade del task default.
- Inner retry loop: try create() → si error 402/429/5xx + hay cascade
siguiente, emit "model_fallback" event + retry con next slug.
Non-recoverable bubble out.
- audit_events.payload incluye cascade_tried, fallbacks[], routing_reason.
- Cost calc unificado vía estimateCostForModel (elimina PRICING
duplicado al final del archivo).
- errorStatus() helper que extrae status code de OpenAI APIError o
del prefix "402 ..." en err.message.
lib/forge/tools.ts:
- import routeFor + MODELS.
- Tool nueva: model_recommend(task, cost_preference?) → devuelve
{ primary, cascade, reason, primary_info }. Para que V tome
decisiones de modelo informadas antes de openrouter_query.
- Tool nueva: forge_cost_report(period?) → query agregado en
conversations + audit_events; devuelve totales, by_model top 8,
fallback_events count. Para que V responda "cuánto vas gastando"
con datos reales, no inventados.
Smoke local en /api/forge/cost?period=this_month: $25.20 / 175 turns
across 3 models (Sonnet via OR $18.85, legacy claude-sonnet-4-6 $6.25,
MiniMax free $0.10). Health: db 57ms, vault round-trip OK,
OpenRouter balance $20 usage $19.61 latency 235ms.
NO cambia el contract SSE hacia el frontend. NO cambia tools existentes.
NO agrega deps.
https://claude.ai/code/session_01Parr3CGTU4ucH6xReyNJ2G
V ya puede descubrir y "instalar" habilidades reusables al vuelo. Cada
skill es un paquete autodescriptivo: system prompt fragment + lista de
tools requeridas + ring level + tags. Cuando Luis describe una tarea,
V invoca skill_search → skill_install y opera con instrucciones
específicas durante el resto del turno.
NUEVO:
migrations/005_skills.sql:
- Tabla skills con id, name, description, system_prompt, required_tools,
examples, ring_max, source, tags, invocation_count.
- Tabla skill_invocations para audit (skill_id, session_id, outcome).
- pg_trgm + índices GIN sobre name/description/tags para ILIKE rápido.
- Sin embeddings (coherente con "solo OpenRouter" — OR no vende
embeddings). Si el catálogo crece, agregamos pgvector + OpenAI
text-embedding-3-small en M16.5.
migrations/006_skills_builtin.sql:
Seeds: 7 skills cubriendo los flujos canónicos de vForge.
- repo-rescue (ring 1) protocolo RESCATE
- new-project-bootstrap (ring 1) protocolo NUEVO
- dns-vercel-namecom (ring 1) apuntar dominio
- github-pr-author (ring 1) cambio en repo via PR
- vercel-deploy-debug (ring 1) diagnóstico build fallido
- code-review (ring 0) review sin escribir
- bulk-categorizer (ring 1) clasificar repos en masa
lib/forge/skills.ts:
- tokenize(query): split en palabras + strip de acentos + stopwords
ES+EN para que "rescatar" matche "rescate" y "vercel" matche
"Vercel" sin importar acentos.
- searchSkills(query, topK): WITH tokens AS (...) + score híbrido
(tag exacto > name substring > tag substring > desc substring).
- getSkillBody(id): full skill + bump invocation_count + insert en
skill_invocations. Fire-and-forget para no bloquear.
- listSkills() para futuros /modules dashboards.
lib/forge/tools.ts:
- skill_search (Ring 0): { query, top_k } → top matches resumidos.
- skill_install (Ring 0): { skill_id } → full body con
system_prompt + required_tools + instrucciones explícitas para V.
Smoke test contra Neon production: los 6 queries naturales devuelven
el skill correcto en top 1 con scores 75-565.
NO cambia tools existentes. NO requiere deploy de Vercel para que las
migrations apliquen (ya las apliqué directo en Neon).
Pendiente próxima tanda: M18 subagentes paralelos in-process.
https://claude.ai/code/session_01Parr3CGTU4ucH6xReyNJ2G
…ile cap
V puede ahora despachar tareas a sub-agentes especializados que corren
en paralelo dentro del mismo turno. Sin Trigger.dev (M9 lo upgrade a
durable), pero suficiente para descargar 5-50 tareas cortas a la vez:
clasificar repos, sumarizar dumps, reviewer múltiples archivos.
NUEVO:
lib/forge/subagents.ts:
8 roles built-in con system prompt + modelo default optimizado:
- researcher → google/gemini-2.5-flash
- reviewer → anthropic/claude-sonnet-4.6
- coder → anthropic/claude-sonnet-4.6
- tester → anthropic/claude-haiku-4.5
- categorizer → google/gemini-2.5-flash
- summarizer → anthropic/claude-haiku-4.5
- extractor → anthropic/claude-haiku-4.5
- translator → google/gemini-2.5-flash
runSubagent({ role, task, model?, maxTokens?, signal? }) → llama
openRouterAdapter.execute() con system prompt del rol. Devuelve
{ role, model, content, tokens_in, tokens_out, cost_usd,
duration_ms, finish_reason }.
recordSubagentRun() persiste a conversations + audit_events para
que el cost dashboard agregue. Fire-and-forget. Cada run aparece
con session_id = parent + ':subagent:' + role.
lib/forge/tools.ts:
Tools nuevas (M18):
- spawn_subagent(role, task, model?, max_tokens?) — Ring 0.
V puede llamar N en paralelo en una sola tool round; el
dispatcher las ejecuta con Promise.all (ya existente).
- list_subagent_roles() — Ring 0. V consulta qué roles existen
antes de spawn si no recuerda.
Fix bug crítico (V reportó hoy):
github_read_file YA NO trunca a 5KB. Default 50KB. max_bytes
hasta 500KB. offset para paginar archivos enormes. Devuelve
total_bytes + next_offset para que V sepa cuándo seguir leyendo
el siguiente chunk. Resuelve "V ciega en archivos grandes" —
tools.ts pesa 58KB, era invisible al modelo.
Smoke test PASS contra producción:
3 subagents en paralelo (summarizer + categorizer + translator),
wall-clock 3673ms vs serial sum 9100ms → speedup 2.48x.
Total cost: \$0.000442 (Haiku + 2x Flash).
V puede ahora orquestar: "audita los 52 repos del catálogo" →
spawn_subagent('categorizer', ...) x52 en una sola round → resultados
en ~5-10s wall-clock vs ~4 minutos serial.
https://claude.ai/code/session_01Parr3CGTU4ucH6xReyNJ2G
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
turbillon50
marked this pull request as ready for review
May 13, 2026 07:07
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 02f044196d
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Comment on lines
+80
to
+82
| const cascade = isKnownSlug | ||
| ? routing.cascade | ||
| : [configuredModel, ...routing.cascade]; |
There was a problem hiding this comment.
This path now prepends unknown system_config.default_model values directly into the model cascade, but seeded installs still use legacy values like claude-sonnet-4-6 (see existing migrations), which are not valid OpenRouter slugs. Because fallback only retries for 402/429/5xx, the first call fails with a non-recoverable 4xx and the chat turn aborts instead of reaching valid fallback models, effectively breaking default deployments until config is manually edited.
Useful? React with 👍 / 👎.
Comment on lines
+43
to
+45
| case "this_month": | ||
| since = new Date(now.getUTCFullYear(), now.getUTCMonth(), 1); | ||
| break; |
There was a problem hiding this comment.
this_month mixes UTC getters with the local-time Date(year, month, day) constructor, so in non-UTC runtimes the window starts at local midnight (shifted in UTC) instead of 00:00:00Z on day 1. That skews month-to-date spend/tokens by several hours and makes this period inconsistent with the explicit UTC handling already used for today.
Useful? React with 👍 / 👎.
turbillon50
added a commit
that referenced
this pull request
May 13, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Resumen
Tres avances en un PR. V ahora descubre skills relevantes, los aplica, y delega trabajo paralelo a sub-agentes especializados. Plus un bug fix crítico: V dejaba de ver archivos > 5KB.
M16 — Skills registry (commit
5925c1c)V tiene un catálogo de 7 habilidades reusables que puede buscar por texto natural y "instalar" al vuelo:
repo-rescuenew-project-bootstrapdns-vercel-namecomgithub-pr-authorvercel-deploy-debugcode-reviewbulk-categorizerTools:
skill_search(query, top_k)+skill_install(skill_id). Search por trigram ILIKE + tokenize ES+EN + stopwords. Cada query natural devuelve el skill correcto en top 1 (smoke contra Neon production con 6 queries PASS).M18 light — Subagentes paralelos in-process (commit
02f0441)V puede invocar N sub-agentes en una sola tool round → el dispatcher los corre con
Promise.all. Smoke test 3 subagents en paralelo → speedup 2.48x (3.67s vs 9.1s serial), costo total $0.000442.8 roles built-in con modelo optimizado:
researcherreviewercodertestercategorizersummarizerextractortranslatorTools:
spawn_subagent(role, task, model?, max_tokens?)+list_subagent_roles(). Persistencia automática aconversations+audit_eventscon session_id = parent + ':subagent:' + role para que el cost dashboard agregue.Sin Trigger.dev todavía (M9 lo upgrade a durable). Si el SSE cierra, results se pierden. Suficiente para tareas < 60s totales.
Fix bug crítico:
github_read_filecapV reportó hoy: "Mi tool github_read_file trunca a 5KB y el repo tiene archivos de 58KB+. No puedo ver el código completo — estoy ciega en archivos grandes."
max_bytes(hasta 500KB)offsetpara paginar archivos enormestotal_bytes+next_offsetpara que V sepa cuándo seguirArchivos
migrations/005_skills.sqlmigrations/006_skills_builtin.sqllib/forge/skills.tslib/forge/subagents.tslib/forge/tools.tsTest plan
tsc --noEmitpasaLo que NO incluye
Pendiente próxima tanda
https://claude.ai/code/session_01Parr3CGTU4ucH6xReyNJ2G
Generated by Claude Code