Repository navigation
Releases
v0.11.2
Fixed
- "continue" after a stopped turn no longer fails. When a turn was stopped mid tool loop,
claude-agent-acp 0.31 failed the next prompt at once with the CLI's diagnostic for the
interrupted cycle ([ede_diagnostic] result_type=user last_content_type=n/a stop_reason=tool_use); OpenCode retried into the same error and the session looked dead.
The prompt is now re-sent once on a re-opened session, on the same stream.
Added
- Warning when OpenCode drops a second copy of the plugin. Configured with different
specs in two configs (e.g. a remote/organization config pinning another version), OpenCode
runs the first layer's copy and silently ignores the other. The running copy now logs
which spec runs and which was ignored.
Changed
- Dependencies updated within their current majors (
@mcpc-tech/acp-ai-provider0.3.9;
dev:@opencode/plugin2.2.5, TypeScript 5.9).aiand@ai-sdk/providerstay on 6 / 3,
the majors OpenCode V2 serves to plugins.
v0.11.0
Changed — breaking
-
One agent per Claude Code mode. The plugin injects a
claude-<mode>agent for every
session mode the agent reports:claude-default,claude-plan,claude-acceptedits,
claude-auto,claude-dontask,claude-bypasspermissions(the six known modes when
discovery has none). They replaceclaude-codeandclaude-code-plan, with no aliases:
rename those agents in youragentssection (claude-code→claude-default,
claude-code-plan→claude-plan). The mode only governs Claude's native tools; OpenCode's
tools keep the agent's OpenCode permissions. Effort stays a model variant;claude-plan
keepshigh. -
The
acpmodels are paired with the Claude Code agents (pairModels, defaulttrue).
OpenCode V2 offers every model to every agent, so the plugin enforces it:- selecting an agent or model that breaks the pair moves the session back to a valid model
(the agent's own, the session's last non-acpmodel, or OpenCode's defaultmodel); - an
acprequest from any other agent fails at once (HTTP 400,x-should-retry: false); - OpenCode subagents launched from a
claude-*agent get their own model or the default one
instead of inheritingacp.
Agents you declare with an
acp/…model count as Claude Code agents.pairModels: false
restores the old behaviour. - selecting an agent or model that breaks the pair moves the session back to a valid model
-
The usage-limit error now says to switch agent rather than model.
Fixed
- Each effort variant started its own agent process. OpenCode V2 builds a new SDK instance
for every distinct modelsettings, and the effort variants are settings, so#high,
#lowand the plain model each spawned aclaude-agent-acpthat resumed the same Claude
session apart from the others (cache partly cold; going back to a variant found a process
that had missed the turns in between). Provider instances built from the same agent config
now share one orchestrator: effort switches live withsetSessionConfigOption, and the
model without a variant restores the session's own effort.
🤖 Generated with Claude Code
v0.10.1
Fixed
- The usage-limit error was retried by OpenCode ≥ 2.1. OpenCode 2.1 no longer decides
retries by the wording of the error message: any error it cannot classify is retried,
about 10 attempts over ~84 s, before the user can switch models. The error is now an
APICallErrorwith HTTP 402 andx-should-retry: false. OpenCode classifies that as
QuotaExceededand fails at once.
Changed
- Documentation translated to English and rewritten for newcomers; OpenCode V2 only stated
explicitly. - Runtime messages (usage-limit error, session notes, mode-change notice) translated to English.
@opencode/pluginpeer range narrowed to^2.0.14(excludes a future 3.x).
Added
- MIT
LICENSEfile. - Load-time compatibility warnings: OpenCode major version, and
claude-agent-acpolder than
the tested 0.31. prepublishOnlyinstall check (scripts/check-install.mjs).
v0.8.0 — continuidad de sesión, coste de tokens y tooling
Versión centrada en el coste de tokens y en la continuidad de sesión. La causa común de casi todos los problemas era la misma: el provider recreaba la sesión ACP con demasiada facilidad, y cada recreación implica replay del historial completo y cache de prompt frío.
⚠️ Breaking
claudeCode.configDirhereda~/.claudepor defecto (antes"isolated"). Las sesiones iniciadas en OpenCode se retoman conclaude --resume <id>(el provider imprime el id). No reintroduce tu ecosistema:settingSources/tools/strictMcpConfigsiguen mandando. Vuelve al comportamiento anterior con"configDir": "isolated".- El límite de contexto deja de inflarse a 1M para el modelo
default; anunciar 1M impedía que la auto-compactación de OpenCode se disparase nunca.
Fixed
- Gasto de tokens desproporcionado (causa principal). Cada turno reenviaba como texto plano todos los
tool-call/tool-resultdel turno anterior — volcados íntegros de ficheros y salidas de shell — que el agente ya tenía, invalidando además el prompt cache. - Cambiar de agente destruía la sesión y replayaba todo.
systemKeyhasheaba el system prompt de OpenCode incluso conpreset, modo en el que ese system nunca se envía al agente. Ahoraclaude-code-plan→ build →claude-codemantiene la misma sesión ACP. - Un cambio en el set de tools ya no recrea la sesión (se aplica en caliente vía
notifications/tools/list_changed), y la llamada del modelo de títulos (tools=0) va a una sesión efímera aparte. - Texto truncado y razonamiento invisible (misma causa). Texto y thinking llegan por stdio ACP y las tool-calls por el socket del
McpHost, sin orden garantizado; al cerrar el stream se descartaba en silencio lo que seguía llegando. Ahora se bufferiza y se vuelca en la continuación. - Revertir mensajes dejaba las dos partes desincronizadas y reenviaba el historial entero. Ahora se compara una huella del historial y se distingue append / revert / divergencia por compactación.
- Dos conversaciones simultáneas se destruían la sesión mutuamente (había un único
TurnStatepor instancia de modelo). Ahora hay una sesión por conversación, con cota LRU. claude-code-planno tenía forma de salir de modo plan:ExitPlanModeno estaba expuesta. Aprobar un plan además concedíabypassPermissions; ahora se prefieredefault.- El agente no usaba la tool
questionni tenía búsqueda web; ambos corregidos. - Resultados de tools con estructura (imágenes/adjuntos), aviso cuando un
effortes rechazado, y cancelar un turno ya no tira la sesión.
Changed
claude-codeyclaude-code-plancomparten el mismo perfil de tools y permisos, difiriendo solo en el modo inicial: salir de plan ya no deja al agente a medias.- Nueva opción
claudeCode.forwardEnvContext(defaulttrue): reenvía la parte factual del system prompt de OpenCode (<env>,<available_references>) como contexto de usuario. Verificado que mantiene la facturación por suscripción.
Changelog completo: CHANGELOG.md