Add Opencode

OpenCode agent provider

Install OpenCode as an optional NanoClaw runtime. The payload is included in this skill; it needs no separate provider branch. It uses the upstream runtime, instructions, host, and setup metadata contracts. The host contract remains at version 1; the container owns its non-secret ChatGPT placeholder file.

OpenCode is offered by the standard setup provider picker. Existing installs can add or authenticate it with pnpm exec tsx setup/index.ts --step provider-auth opencode. To replace an installed payload and update its pins, append --refresh; back up local payload edits first. Ordinary re-authentication leaves installed files and the container image alone. Backend defaults are installation-wide; model and reasoning effort can be overridden per group through the existing container configuration. Per-group backend/auth selection and structured channel attachment transport are separate work.

Authentication checks the installed files, registration lines, and exact pins against this skill's declarations without launching a subprocess or container. Install and refresh run the existing provider contract verification; the build step owns image freshness. Model selection does not repeat installation checks. A working backend and account are checked separately by sending a real request.

Install

After installing this payload, run pnpm exec tsx scripts/opencode-host.ts --configure for host OpenCode setup, or use --update / --debug for the corresponding operational skill. An existing OpenCode CLI can also run directly in the checkout; it discovers .claude/skills natively. Host sign-in uses OpenCode's own settings and is independent of the container's OneCLI credentials. Installed setup failures use the existing provider failure-assist hook, including wizard authentication and installation-check failures. Host diagnostic context is model input and may remain in native OpenCode history; deleting its private temporary file does not erase those records. The helper requires stable OpenCode 1.18.25 or newer with --prompt and prefers the newest compatible installation it finds. Automatic help before payload installation is optional and is not part of the runtime contract.

Install and refresh require host contract version 1. The compatibility predicate below guards every subsequent step, so an unsupported core receives no partial payload or dependency changes. Update core first if it reports a missing prerequisite.

node -e "const fs=require('fs'); const p='src/provider-contracts/registry.ts'; if(fs.existsSync(p) && /PROVIDER_HOST_CONTRACT_SEAM_VERSION = 1/.test(fs.readFileSync(p,'utf8'))) console.log('yes'); else console.log('no')"

Copy only the files listed below from this skill's payload/ to the matching paths at the project root. Do not copy ignored dependency directories or other generated native-test files. These are skill-owned files; overwrite them together when refreshing the skill. Keep the core-owned cwd-shim.ts, registries, and contract realization files in place.

When refreshing an older installation, remove its unused opencode-memory-plugin.ts, opencode.compaction.test.ts, and dedicated opencode-managed-config tree from container/agent-runner/src/providers/. Recreate affected containers after the refresh to discard their old config symlinks. Keep other tools' settings and persisted session data.

The obsolete host Dockerfile guard must also be removed during refresh; current OpenCode installation is declared by the SDK and CLI manifests.

rm -f src/opencode-dockerfile.test.ts

payload/container/agent-runner/src/provider-contracts/opencode.ts -> container/agent-runner/src/provider-contracts/opencode.ts
payload/container/agent-runner/src/providers/mcp-to-opencode.test.ts -> container/agent-runner/src/providers/mcp-to-opencode.test.ts
payload/container/agent-runner/src/providers/mcp-to-opencode.ts -> container/agent-runner/src/providers/mcp-to-opencode.ts
payload/container/agent-runner/src/providers/opencode-config.ts -> container/agent-runner/src/providers/opencode-config.ts
payload/container/agent-runner/src/providers/opencode-memory.ts -> container/agent-runner/src/providers/opencode-memory.ts
payload/container/agent-runner/src/providers/opencode-registration.test.ts -> container/agent-runner/src/providers/opencode-registration.test.ts
payload/container/agent-runner/src/providers/opencode-turn.ts -> container/agent-runner/src/providers/opencode-turn.ts
payload/container/agent-runner/src/providers/opencode.attachments.test.ts -> container/agent-runner/src/providers/opencode.attachments.test.ts
payload/container/agent-runner/src/providers/opencode.config.test.ts -> container/agent-runner/src/providers/opencode.config.test.ts
payload/container/agent-runner/src/providers/opencode.conformance.test.ts -> container/agent-runner/src/providers/opencode.conformance.test.ts
payload/container/agent-runner/src/providers/opencode.empty-resume.test.ts -> container/agent-runner/src/providers/opencode.empty-resume.test.ts
payload/container/agent-runner/src/providers/opencode.factory.test.ts -> container/agent-runner/src/providers/opencode.factory.test.ts
payload/container/agent-runner/src/providers/opencode.memory.test.ts -> container/agent-runner/src/providers/opencode.memory.test.ts
payload/container/agent-runner/src/providers/opencode.native.test.ts -> container/agent-runner/src/providers/opencode.native.test.ts
payload/container/agent-runner/src/providers/opencode.question.test.ts -> container/agent-runner/src/providers/opencode.question.test.ts
payload/container/agent-runner/src/providers/opencode.shared-runtime.test.ts -> container/agent-runner/src/providers/opencode.shared-runtime.test.ts
payload/container/agent-runner/src/providers/opencode.sse-cleanup.test.ts -> container/agent-runner/src/providers/opencode.sse-cleanup.test.ts
payload/container/agent-runner/src/providers/opencode.ts -> container/agent-runner/src/providers/opencode.ts
payload/container/agent-runner/src/providers/opencode-auth.ts -> container/agent-runner/src/providers/opencode-auth.ts
payload/container/agent-runner/src/providers/opencode-auth.test.ts -> container/agent-runner/src/providers/opencode-auth.test.ts
payload/scripts/opencode-auth-config.test.ts -> scripts/opencode-auth-config.test.ts
payload/scripts/opencode-auth.test.ts -> scripts/opencode-auth.test.ts
payload/scripts/opencode-auth.ts -> scripts/opencode-auth.ts
payload/scripts/opencode-host.ts -> scripts/opencode-host.ts
payload/scripts/opencode-host.test.ts -> scripts/opencode-host.test.ts
payload/scripts/opencode-model-config.ts -> scripts/opencode-model-config.ts
payload/scripts/opencode-models.test.ts -> scripts/opencode-models.test.ts
payload/scripts/opencode-models.ts -> scripts/opencode-models.ts
payload/scripts/opencode-vault.test.ts -> scripts/opencode-vault.test.ts
payload/scripts/opencode-vault.ts -> scripts/opencode-vault.ts
payload/scripts/tsconfig.opencode-auth.json -> scripts/tsconfig.opencode-auth.json
payload/setup/providers/opencode.test.ts -> setup/providers/opencode.test.ts
payload/setup/providers/opencode.ts -> setup/providers/opencode.ts
payload/src/provider-contracts/opencode.ts -> src/provider-contracts/opencode.ts
payload/src/providers/opencode-auth-stub.ts -> src/providers/opencode-auth-stub.ts
payload/src/providers/opencode-registration.test.ts -> src/providers/opencode-registration.test.ts
payload/src/providers/opencode.ts -> src/providers/opencode.ts

Append import './opencode.js'; once to each of the five setup, provider, and contract barrels below. Keep all existing imports.

import './opencode.js';

import './opencode.js';

import './opencode.js';

import './opencode.js';

import './opencode.js';

Install the SDK in the runner's Bun package and add the matching CLI manifest entry with trusted postinstall enabled. Both pins must remain exactly 1.18.25. When refreshing an existing install, replace both old pin entries; presence alone does not establish compatibility. This updates the runner package and lockfile; there is no host SDK dependency.

@opencode-ai/sdk@1.18.25

{"name":"opencode-ai","version":"1.18.25","onlyBuilt":true}

Run the host build, runner typecheck, host/auth tests, and all provider tests. The tests exercise real barrel registration and the provider-owned contract conformance suite. All checks must pass before rebuilding the agent image.

pnpm run build

pnpm exec tsc -p scripts/tsconfig.opencode-auth.json

cd container/agent-runner && bun run typecheck

pnpm exec vitest run src/providers/opencode-registration.test.ts scripts/opencode-auth*.test.ts scripts/opencode-host.test.ts scripts/opencode-models.test.ts scripts/opencode-vault.test.ts setup/providers

cd container/agent-runner && bun test --isolate src/providers/opencode*.test.ts src/providers/mcp-to-opencode.test.ts

Build the local image with ./container/build.sh build. The new SDK dependency requires a full local build; a CLI-only overlay cannot supply it. This switches a published-image installation to locally built images.

./container/build.sh build

Authenticate and select a group

Run pnpm exec tsx setup/index.ts --step provider-auth opencode from the project root to install a missing payload and image, then choose authentication. If the provider is already installed, this command leaves its files and image alone; append --refresh only when intentionally replacing its payload and pins. Choose ChatGPT sign-in, a local OpenAI-compatible endpoint, OpenRouter, DeepSeek, or a supported native backend. Automatic API-key configuration supports OpenAI, OpenRouter, DeepSeek, Google, and Anthropic; other native authentication schemes require separate integration. The command stores credentials in OneCLI and backend defaults in .env. The full setup wizard also offers this flow and selects OpenCode for new groups only after configuration succeeds. The standalone command leaves the instance default unchanged.

For ChatGPT, native OpenCode sign-in runs in a temporary container directory. The OAuth credential is translated into OneCLI's supported vault format, and the temporary native file is removed. The container initializes fixed onecli-managed placeholders before every OpenCode server start at $XDG_DATA_HOME/opencode/auth.json; tokens and account metadata stay in OneCLI. API-key mode clears stale OAuth state. Refresh the payload and restart the host service and affected containers when updating from the earlier read-only-bind candidate; old containers retain their mounts until recreated.

Before using a group, grant its OneCLI agent access to the chosen secret. Read its existing secret assignments first and merge the new secret ID into that list: onecli agents set-secrets replaces assignments. Verify the result with onecli agents secrets. Do not put a key in .env, command arguments, or the container environment.

After installing on a running NanoClaw host, restart its actual host service before waking any OpenCode group. This reloads the host provider registration and backend settings. On Linux use systemctl --user restart nanoclaw-v2-<install-slug>.service (or the installation's system service command); on macOS use its normal launchd restart workflow. Confirm the service is running, then select and restart the test group:

ncl groups config update --id <group-id> --provider opencode
ncl groups restart --id <group-id>

Send a message and verify a reply, then send a second message to check session continuation. The test requires a reachable backend and the correct OneCLI secret grant. No provider is switched by the install steps alone. If memory needs to move from another provider, follow /migrate-memory before switching.

Recover a ChatGPT login

NanoClaw currently pins OneCLI 1.41.0, which cannot refresh these OAuth tokens after expiry because its refresh request omits the required client ID. Use manual reauthentication when that happens. Reliable unattended ChatGPT operation requires a separately validated gateway refresh fix; upgrading to OneCLI 1.43.1 also requires migrating its removed agent-grant API.

The container uses only the fixed sentinel; do not implement token refresh in the provider or copy live credentials into a group. A saved credential is not proof that authentication still works.

If a request fails because the login expired or was revoked, run on the host:

pnpm exec tsx scripts/opencode-auth.ts --reauth
# For a browser on the host instead of device pairing:
pnpm exec tsx scripts/opencode-auth.ts --reauth --method browser

This pairs again and updates the existing OneCLI secret ID, preserving its agent permissions and all backend/model defaults. It uses NanoClaw's ONECLI_URL and ONECLI_API_KEY management connection. If no credential exists, it creates one; grant that new secret to the group as described above. Retry the failed request.

An unavailable vault, duplicate name, or incompatible credential entry stops the operation before sign-in. Resolve the gateway/permissions or entry metadata in OneCLI and retry; do not delete a credential to force setup to run. Failed pairing leaves the old entry intact; failed saves leave defaults unchanged. Temporary native credentials are removed after either success or failure.

Change or refresh the default model

Run pnpm exec tsx scripts/opencode-models.ts to keep the current default or choose another model without signing in again. This changes only OPENCODE_MODEL; the small model, endpoint, credentials, and group overrides stay as configured. Restart the NanoClaw host and affected groups afterward.

pnpm exec tsx scripts/opencode-models.ts --list --refresh
pnpm exec tsx scripts/opencode-models.ts --model openai/<model-id>

Discovery runs the installed container's opencode models command and filters for text and tool support, including its ChatGPT-specific filter when selected. Only a disposable fixed sentinel is used for that filter; no credentials or host OpenCode files are mounted for discovery. --refresh fetches the runtime's current model catalog; it does not upgrade the CLI or SDK. Account access is checked by a real request, not by catalog membership. Standalone host OpenCode is never consulted. If discovery is unavailable, keep the existing model or enter an id manually. There is no static fallback list. A custom OpenAI-compatible endpoint is queried through its own /models endpoint; other custom endpoints use manual IDs. The configured backend must match the model prefix; changing backends still uses the authentication command. Exported defaults take precedence over .env, so conflicting exported values must be cleared before changing the saved model.

This separate command avoids rerunning authentication merely to change a model, and querying the container avoids disagreement with a separately upgraded host CLI. New models needing newer runtime support require a matched CLI/SDK update and image rebuild. Model changes do not automatically change context limits or modalities; adjust any custom overrides to match the new model.

Backend defaults

The host reads these values from exported environment variables, then .env. Put comments on separate lines. These settings affect only OpenCode containers.

• OPENCODE_PROVIDER: OpenCode backend ID, such as openai or openrouter.
• OPENCODE_MODEL: default full provider/model ID. The group's model wins.
• OPENCODE_SMALL_MODEL: optional separate model for lighter work, using the same backend prefix as OPENCODE_PROVIDER.
• OPENCODE_BASE_URL: backend URL, or native to use the native endpoint. For an openai backend with a custom URL, the runtime uses Chat Completions. An absent setting retains the historical ANTHROPIC_BASE_URL fallback for existing installs. The auth command writes this provider-owned setting and preserves Claude's endpoint.
• OPENCODE_AUTH_MODE=chatgpt: initialize the container's non-secret ChatGPT stub. Leave unset for API-key and local endpoints; the auth command handles this when switching.
• OPENCODE_MODEL_CONTEXT_LIMIT: positive token count for the main model.
• OPENCODE_MODEL_OUTPUT_LIMIT: positive output limit, requiring a context limit.
• OPENCODE_MODEL_INPUT_MODALITIES: optional comma-separated main-model input types from text,audio,image,video,pdf.
• OPENCODE_NATIVE_ATTACHMENT_MAX_COUNT / OPENCODE_NATIVE_ATTACHMENT_MAX_BYTES: optional limits for already-staged structured attachments. Upstream channel attachment transport remains text-only until that separate feature lands.

Custom model limits and modalities apply only to the main model. MCP servers come from the core's resolved runner configuration. Before each external turn, the shared hook renders memory into one instruction file under writable XDG data, alongside current core instructions and delivery wording. Native continuation after compaction and Task children reread that file. Compaction uses the turn-start memory snapshot; the next external turn refreshes it.

The container gets model, permission, and MCP configuration from NanoClaw and keeps OPENCODE_DISABLE_PROJECT_CONFIG enabled. It declares no plugin and uses normal writable native config locations. Upstream may attempt a background authoring-dependency install, but startup does not wait for it. Offline native tests verify turns without registry access. Host OpenCode keeps its native configuration and plugin support.

OpenCode keeps one server and continuously read event subscription per container. Prompts are serialized and completion comes from the native prompt response plus its stored messages. A stale idle event cannot complete a new turn. Failed or uncertain turns are never automatically replayed. Aborts stop native execution; if completion cannot be confirmed within the cleanup bound, the server is stopped and the next query resumes from persisted state. Effective configuration changes restart the shared server.

MCP calls allow 330 seconds, covering the core's five-minute human question window plus transport overhead. Cancelling a turn cancels its active tool wait; a question already posted to chat remains visible.

For reproducible native integration coverage, download the official OpenCode 1.18.25 binary and run from container/agent-runner:

OPENCODE_TEST_BINARY=/absolute/path/opencode bun test --isolate src/providers/opencode.native.test.ts

The test checks the binary version, starts a local model fixture, and exercises native tools, automatic and overflow compaction, cold resume, child memory, terminal errors, a 65-second MCP call, and cancellation. It writes its requests and server logs to the temporary evidence directory printed at completion. It takes about two minutes and requires no account credentials. The ordinary test suite skips this check unless OPENCODE_TEST_BINARY is set.

To remove the provider, follow REMOVE.md.