fix: respect OpenCode model resolution priority in bootstrap injection #294

1orZero · 2026-01-19T02:19:47Z

Summary

Add model resolution to bootstrap injection following OpenCode's documented priority
Ensures bootstrap uses the same model the session is configured to use

Problem

The Superpowers plugin injects a bootstrap prompt via session.prompt() without specifying a model or agent. This means OpenCode falls back to its internal default, which may differ from the model the user actually selected or configured.

Example of the issue

User selects GPT-5.2 Codex in the TUI:

But the bootstrap is processed by a different model (gemini-3-pro-preview):

The bootstrap injection used OpenCode's internal default instead of the user's selected model, causing the skill context to be processed by a completely different model than expected.

Solution

Added getBootstrapTarget() that resolves the expected model following OpenCode's documented priority:

config.agent[default_agent].model — Agent-specific model (highest priority)
Markdown-defined agent model — Via client.app.agents() for agents defined in .md files
config.model — Global model setting
Last used model — Fallback from ~/.local/state/opencode/model.json

Now passes both agent and model parameters to session.prompt() for proper alignment.

References

OpenCode Models Documentation — Loading models priority
OpenCode Agents Documentation — Agent model override behavior

Previously getPreferredModel() only read from ~/.local/state/opencode/model.json (last used model), ignoring user-configured agent and global models. Now getBootstrapTarget() follows OpenCode's documented priority: 1. default_agent's configured model (config.agent[name].model) 2. Agent model from markdown-defined agents (via client.app.agents()) 3. Global model (config.model) 4. Last used model (fallback) Also passes agent parameter to session.prompt() for proper context.

coderabbitai · 2026-01-19T02:19:56Z

📝 Walkthrough

Walkthrough

The pull request adds model and session state resolution capabilities to the bootstrap flow. It introduces helper functions to parse model references, read persisted model state, and resolve bootstrap targets from configuration or fallback sources, enabling the system to remember and reuse selected models across sessions.

Changes

Cohort / File(s)	Summary
State Resolution & Model Persistence `.opencode/plugin/superpowers.js`	Adds `stateModelFile` path configuration, introduces helper functions (`parseModelRef`, `getLastUsedModel`, `getBootstrapTarget`) to resolve bootstrap target from config hierarchy (default_agent → agent..model → agents list → global model → last used), persists model selection across sessions, and integrates resolved agent/model into `injectBootstrap` via await pattern with payload augmentation. Retains existing injection points at session.created and session.compacted.

Estimated code review effort

🎯 3 (Moderate) | ⏱️ ~20 minutes

Possibly related PRs

fix: properly inherit agent model from main #144: Modifies default inherited agent model to "opus", directly affecting the same model-selection and resolution behavior as this PR's bootstrap target logic.

Poem

🐰 A model remembers its favorite hop,
Across sessions, the choice won't drop!
State files whisper what came before,
Bootstrap roots deeper than ever before,
Persistence hops on, forevermore! ✨

🚥 Pre-merge checks | ✅ 3

✅ Passed checks (3 passed)

Check name	Status	Explanation
Title check	✅ Passed	The pull request title clearly and specifically describes the main change: implementing proper model resolution priority in bootstrap injection by respecting OpenCode's documented hierarchy.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing touches

📝 Generate docstrings

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

coderabbitai

Actionable comments posted: 1

🤖 Fix all issues with AI agents

In @.opencode/plugin/superpowers.js:
- Around line 60-97: The top comment above getBootstrapTarget no longer matches
the actual resolution order—update it to reflect the current logic: first check
default_agent.model from config (config.agent[agentName].model parsed via
parseModelRef), then fallback to agents returned by client.app.agents
(markdown-defined agents) and parseModelRef on the matched agent, then fallback
to global config.model (parseModelRef), and finally use getLastUsedModel(); also
reflect the outer try/catch behavior that returns { agentName: 'build', model:
lastUsed } on errors.

coderabbitai · 2026-01-19T02:24:43Z

.opencode/plugin/superpowers.js

+  // Resolve the model the session is expected to use:
+  // default_agent.model (if set) -> global model -> last used.
+  const getBootstrapTarget = async () => {
+    const lastUsed = getLastUsedModel();
+
+    try {
+      const configRes = await client.config.get();
+      const config = configRes?.data ?? configRes;
+
+      const agentName =
+        typeof config?.default_agent === 'string' && config.default_agent.trim()
+          ? config.default_agent.trim()
+          : 'build';
+
+      const agentModelFromConfig = parseModelRef(config?.agent?.[agentName]?.model);
+      if (agentModelFromConfig) return { agentName, model: agentModelFromConfig };
+
+      // Agents defined via markdown files might not appear under config.agent.
+      try {
+        const agentsRes = await client.app.agents();
+        const agents = agentsRes?.data ?? agentsRes;
+        if (Array.isArray(agents)) {
+          const agent = agents.find((a) => a?.id === agentName || a?.name === agentName);
+          const agentModelFromList = parseModelRef(agent?.model);
+          if (agentModelFromList) return { agentName, model: agentModelFromList };
+        }
+      } catch {
+        // ignore; fall back to other sources
+      }
+
+      const globalModel = parseModelRef(config?.model);
+      if (globalModel) return { agentName, model: globalModel };
+
+      return { agentName, model: lastUsed };
+    } catch {
+      return { agentName: 'build', model: lastUsed };
+    }
+  };


⚠️ Potential issue | 🟡 Minor

Update the resolution-order comment to match the code.

The inline comment omits the markdown-defined agent model check now present in the logic.

📝 Suggested comment fix

- // Resolve the model the session is expected to use: - // default_agent.model (if set) -> global model -> last used. + // Resolve the model the session is expected to use: + // default_agent.model -> markdown-defined agent model -> global model -> last used.

🤖 Prompt for AI Agents

In @.opencode/plugin/superpowers.js around lines 60 - 97, The top comment above getBootstrapTarget no longer matches the actual resolution order—update it to reflect the current logic: first check default_agent.model from config (config.agent[agentName].model parsed via parseModelRef), then fallback to agents returned by client.app.agents (markdown-defined agents) and parseModelRef on the matched agent, then fallback to global config.model (parseModelRef), and finally use getLastUsedModel(); also reflect the outer try/catch behavior that returns { agentName: 'build', model: lastUsed } on errors.

obra · 2026-01-22T18:52:16Z

Reproduction Attempt

I tried to reproduce this issue on OpenCode v1.1.31 (macOS) with the superpowers plugin installed via the documented symlink method.

Test Performed

cd /tmp
opencode run -m opencode/gpt-5-nano "What skills do you have? Use find_skills tool."

Result

The model specified (opencode/gpt-5-nano) was used correctly throughout the session. The bootstrap injection via session.prompt({ noReply: true }) did not cause any model or agent switching behavior.

Logs showed:

service=llm providerID=opencode modelID=gpt-5-nano sessionID=... agent=build mode=primary stream

Questions

To help reproduce this:

What OpenCode version are you using?
Are you using the TUI or CLI?
What model/agent configuration do you have in opencode.json?
Can you provide exact steps to reproduce the model switching?
Does this happen on every session, or intermittently?

The 81 lines of model resolution code is substantial - I'd like to confirm the issue exists before merging a fix that might not be necessary.

obra · 2026-01-22T21:37:48Z

Thank you for this contribution! However, this PR has been superseded by PR #330.

Context: PR #330 completely rewrote the bootstrap injection mechanism. Instead of using session.prompt() (which had model resolution issues as you correctly identified), it now uses the experimental.chat.system.transform hook:

'experimental.chat.system.transform': async (_input, output) => {
  const bootstrap = getBootstrapContent();
  if (bootstrap) {
    (output.system ||= []).push(bootstrap);
  }
}

This approach:

Injects into the system prompt, not the chat history
Uses whatever model the user has selected (no model resolution needed)
Avoids the agent reset bug (OpenCode: bootstrap injection can reset custom agent to build (possible noReply interaction) #226)

Since session.prompt() is no longer used for bootstrap injection, the model resolution logic in this PR is no longer necessary.

Closing as superseded by #330.

obra · 2026-01-22T21:37:52Z

Superseded by PR #330

coderabbitai bot reviewed Jan 19, 2026

View reviewed changes

marcusrbrown mentioned this pull request Jan 19, 2026

[opencode]Conversation Automatically Switches Agent Mode and Model After Starting Chat #301

Closed

unsync added a commit to garocode/superpowers that referenced this pull request Jan 20, 2026

feat: integrate obra#294

78064dc

obra added the opencode label Jan 22, 2026

obra closed this Jan 22, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

fix: respect OpenCode model resolution priority in bootstrap injection #294

fix: respect OpenCode model resolution priority in bootstrap injection #294

Uh oh!

1orZero commented Jan 19, 2026 •

edited

Loading

Uh oh!

coderabbitai bot commented Jan 19, 2026 •

edited

Loading

Walkthrough

Changes

Estimated code review effort

Possibly related PRs

Poem

Uh oh!

coderabbitai bot left a comment

Uh oh!

coderabbitai bot Jan 19, 2026

Uh oh!

obra commented Jan 22, 2026

Uh oh!

obra commented Jan 22, 2026

Uh oh!

obra commented Jan 22, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

fix: respect OpenCode model resolution priority in bootstrap injection #294

fix: respect OpenCode model resolution priority in bootstrap injection #294

Uh oh!

Conversation

1orZero commented Jan 19, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Problem

Example of the issue

Solution

References

Uh oh!

coderabbitai bot commented Jan 19, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Walkthrough

Changes

Estimated code review effort

Possibly related PRs

Poem

Uh oh!

coderabbitai bot left a comment

Choose a reason for hiding this comment

Uh oh!

coderabbitai bot Jan 19, 2026

Choose a reason for hiding this comment

Uh oh!

obra commented Jan 22, 2026

Reproduction Attempt

Test Performed

Result

Questions

Uh oh!

obra commented Jan 22, 2026

Uh oh!

obra commented Jan 22, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

1orZero commented Jan 19, 2026 •

edited

Loading

coderabbitai bot commented Jan 19, 2026 •

edited

Loading