Agent providers

Tasuku supports CODEX, CLAUDE, and OPENCODE. Select the provider, optional model, and default reasoning effort at instance scope or as a repository override.

Provider authentication

The chosen CLI must be installed in the selected agent runtime. In Settings → Coding agent, paste the provider's subscription credential while saving the agent configuration:

• For Codex, first enable Device code authorization for Codex in ChatGPT's security settings. Then select Connect ChatGPT, open the OpenAI authorization page, and enter the one-time device code. Tasuku encrypts the resulting credential without returning it to the browser. If device authorization is unavailable, expand Use auth.json instead and paste the complete ~/.codex/auth.json.
• Claude expects an OAuth token.
• OpenCode expects the OpenCode CLI's auth.json.

Tasuku stores these as the encrypted CODEX_AUTH_JSON, CLAUDE_OAUTH_TOKEN, or OPENCODE_AUTH_JSON secret and never returns plaintext. The bundled runtime images install all three tools and inject only the selected credential while an agent runs. On a repository page, choose Reuse instance credential or Use repository credential; the latter creates a repository-scoped override.

During a Codex run, Codex refreshes managed ChatGPT tokens when needed. Tasuku captures the updated auth.json from the private runtime channel and replaces the encrypted secret after the run, including when the agent later fails. Concurrent refreshes use version checks, so a newer reconnect or refresh is never overwritten by an older run.

Leaving the credential field blank preserves the existing secret. Credentials are not part of the configuration JSON returned by the API.

OpenCode

OpenCode routes requests through the providers configured in its own credential. The pasted auth.json is stored as the encrypted OPENCODE_AUTH_JSON secret; an optional OPENCODE_PROVIDER_CONFIG_JSON secret supplies additional provider configuration. OpenCode models use provider/model format, and the model suggestions come from the connection's provider catalog.

Subscription usage

After saving the provider and credential, the Review agent card shows the usage windows returned by that subscription. These commonly include a five-hour and weekly limit. Monthly spend, extra-usage, or model-specific limits appear only when the provider returns them.

Usage snapshots are cached for one minute and refreshed by the dashboard every five minutes. Use Refresh to request a current snapshot. If Codex usage returns 401 Unauthorized, Tasuku asks Codex to refresh the managed ChatGPT token, saves the updated encrypted credential, and retries the usage request once. Other failures do not trigger a token refresh; reconnect ChatGPT if the panel still reports an authentication failure.

Repository usage follows the same precedence as review runs: a repository credential overrides the instance credential, otherwise the instance credential is reused.

Codex usage follows the provider's current usage contract. Claude usage relies on the private machine-readable endpoint used by Claude Code, so a Claude CLI or service change may temporarily make the panel unavailable without affecting review runs.

Invocation context

Tasuku supplies the provider with:

• the primary repository at the workflow's exact commit;
• a catalog of approved supporting repositories that can be loaded read-only on demand;
• workflow instructions and repository policy;
• custom instructions and enabled learnings;
• provider-specific skill packaging;
• configured HTTPS MCP servers.

The provider does not receive GitHub App credentials. It produces structured output or a bounded workspace artifact for the worker to validate.

Model selection

An empty model value leaves model selection to the installed provider's default. A non-empty value is passed as provider configuration. Confirm it is supported by the CLI version in your runtime image.

The bundled model selector suggests Codex model identifiers (gpt-6-astra, gpt-6-sol, gpt-6-luna, gpt-5.6-sol, gpt-5.6, gpt-5.6-terra, gpt-5.6-luna, gpt-5.5, gpt-5.4, gpt-5.3-codex-spark) and Claude models (claude-fable-5-1, claude-opus-5-5, claude-fable-5, claude-opus-5, claude-opus-4-8, claude-sonnet-5, claude-sonnet-4-6, claude-haiku-4-5). These are Tasuku's configured suggestions, not a live guarantee of availability on a provider or subscription. The reasoning selector exposes these levels:

• low, medium, high, xhigh, and max for the bundled GPT-6 and Claude 5 presets;
• low, medium, high, and xhigh for Codex models that do not support max.

medium is the stored connection default, but every bundled workload route sets its own effort, so the connection default applies only when a route inherits it. ultra is not accepted by Tasuku's reasoning-effort configuration.

Workload routing

Every run maps to one workload: research, plan, implement, review, or deep scan. Each workload routes to its own provider, model, reasoning effort, and speed mode, editable in Coding agent → Workload routing. A route set to inherit uses the connection default. The bundled route defaults depend on the connection provider:

These are starting points selected for capability on analysis and scans, with Luna at maximum effort for implementation. Verify model access in the configured subscription and compare representative runs before relying on them in production. Existing saved routes keep their selected models and efforts.

OpenCode routes inherit the connection default for every workload.

Speed mode

The connection and each workload route select a speed mode: STANDARD or FAST, with INHERIT on a route following the connection. Tasuku maps Codex fast mode to service_tier="fast" and allows Claude fast mode on Opus 5.5, Opus 5, and Opus 4.8. Tasuku rejects fast mode for OpenCode. Provider availability, model support, and billing are separate from these local configuration rules.

Behavior and instructions

The Instructions tab controls the agent's communication and working style. DEFAULT is neutral and concise; FRIENDLY, PRAGMATIC, and NERDY apply Tasuku's built-in guidance. CUSTOM requires up to 4,000 characters of custom behavior guidance. Additional instructions are appended after the selected behavior.

Behavior and additional instructions inherit together as the repository's instructions category. They influence presentation and approach, but do not weaken diff validation, path exclusions, permissions, publication limits, or approval gates.

Overrides

Use a repository override when one codebase needs a different provider, model, or reasoning default. Keep the override disabled when it should track instance changes. See Inheritance and overrides.