Goal-driven agent execution for OpenCode ACP

[muhammad-fiaz]

Body

What are you proposing?

Add a Goals feature to OpenCode ACP agents in Zed that allows users to define a high-level objective and have the agent continuously work toward that objective until it is completed, verified, or requires user intervention.

Instead of the agent treating each prompt as an isolated task, the goal would remain active across multiple iterations of planning, implementation, testing, debugging, and verification.

For example:

Goal: Implement the requested feature completely and make sure all relevant tests pass.

The agent should be able to keep working toward that goal without requiring the user to manually tell it what to do after every step.

Why does this matter?

Complex development tasks often require many iterations:

Plan → Implement → Test → Inspect → Fix → Test again → Verify

With a normal agent workflow, the user may need to repeatedly prompt the agent to continue or address the next issue.

A Goals feature would make it possible to delegate the overall objective while allowing the agent to determine the necessary intermediate steps.

This would make longer-running tasks easier and reduce the amount of manual supervision required.

It would also make OpenCode ACP more suitable for tasks that cannot realistically be completed in a single agent turn.

Are there any examples or context?

This is similar to the goal-oriented execution experience in OpenAI Codex, where users can give an agent a broader objective and allow it to continue working through multiple steps toward completion.

A possible workflow in Zed could look like:

Goal
────────────────────────────────────────────
Complete the requested implementation and
verify that all relevant tests pass.
────────────────────────────────────────────

[ Start Goal ]

While the goal is running:

Goal: Complete implementation

✓ Analyze requirements
✓ Plan implementation
✓ Implement changes
✓ Run tests
→ Investigating failing tests
→ Applying fixes
○ Run final verification
○ Complete goal

The user should be able to pause, resume, cancel, or interact with the agent while the goal is running.

Possible approach

A Goal could be represented as a persistent objective associated with the current agent session.

The agent could repeatedly execute an internal loop such as:

Understand goal → Plan next action → Execute → Observe result → Evaluate progress → Continue

The agent should only consider the goal complete after evaluating whether the requested objective has actually been achieved.

It would also be useful to support optional completion criteria, for example:

Goal:
Complete the requested implementation.

Completion criteria:
- Implementation is complete
- Relevant tests pass
- No known errors remain
- Changes have been verified

The feature should continue respecting Zed/OpenCode ACP's existing permissions and approval mechanisms.

Additional safeguards could include:

• Maximum execution time
• Maximum iterations
• Token/resource limits
• Pause and cancel controls
• User approval when required
• Clear progress/status indicators
• A final summary explaining what was completed
• Verification of the goal before marking it complete

The key idea is to allow the user to specify what outcome they want, while the agent handles the iterative work required to reach that outcome.

[muhammad-fiaz]

Related thread: anomalyco/opencode#27167

[muhammad-fiaz]

I found one solution while looking into this: https://github.com/prevalentWare/opencode-goal-plugin

I’ve tested it with OpenCode and it seems to work pretty well with Zed through ACP. After installing the plugin, /goal is available in OpenCode, including when using it through ACP.

It supports long-running goals, automatic continuation, pause/resume (/pause_goal and /resume_goal), persistent goal state, progress tracking, and completion verification.

So this seems like a pretty decent workaround/proof of concept for the Goals idea in the meantime. It would be interesting to see whether some of these capabilities could eventually be integrated more natively into Zed/ACP.