[BambinoSK]

Fixes #1187.

Problem

On OpenCode v2 the connector advertises an Explore subagent that cannot work, and the Agent menu lists agents that do not exist. See #1187 for the full analysis: the IDE-integrated explore prompt never reaches a v2-readable location (it was written for v1's deploy-to-directory mechanism), and v2's built-in explore is toolless in this plugin's sessions because the plugin's own config denies the native tools while the built-in has no execute permission (the only reachability path for MCP tools in v2 Code Mode). The menu's built-in list is a v1-era snapshot (general and explore are Task-tool subagents on v2, scout does not exist), and user-level custom agents from the global agent directories never appear in the dropdown even though the agent advertises them as selectable modes.

What this PR does

• Injects the explore subagent inline for v2+. When the probed OpenCode version is v2 or later, the launch config (OPENCODE_CONFIG_CONTENT) gains an agents.explore entry: mode: subagent, no model key (inherits the session model), the bundled prompt as a plain system string with its v1 frontmatter stripped, and permissions as a v2 Ruleset array: execute allowed first, the 18 read-only agentbridge_* MCP tools allowed, the 7 mutating ones denied, edit/write/patch/bash/shell denied. Config-agent rules merge after the built-in policy, last-match-wins, so the injected rules win for every action they cover and the built-in catch-all deny blocks uncovered MCP actions. The version probe result is cached and reused when building the config, so v1 and unknown versions keep the existing behavior.
• Stops advertising phantom agents. On v2+ the built-in list is only build and plan (the primary session modes); general, explore, and scout are dropped from the menu. The explore subagent stays available to the model through the injected config. v1 and unknown versions keep the full list.
• Lists user-level custom agents. The dropdown now also scans both global agent directories v2 reads, ~/.opencode/agents and ~/.config/opencode/agents, with OpenCode's own precedence: ~/.opencode wins over the XDG dir on slug collisions, a project agent wins over a same-slug global one, and built-ins are never shadowed (enforced in the merge itself). This also fixes the near-duplicate Mode and session-option sections that leaked into the menu whenever a global custom agent made the agent list and the advertised mode values diverge.

Notes for reviewers

• The agent entry uses exactly the fields the v2 config schema defines for agents entries (description, mode, permissions, system; additionalProperties: false), and the ruleset is the documented ordered array of {action, resource, effect}.
• The write/patch/bash denies duplicate edit/shell coverage under v2 action naming; they are kept as defense in depth and are normalized by the runtime.
• write/patch/bash denies and the top-level v1-style permission map are accepted and normalized by v2.0.27 even though current docs prefer edit/shell and a top-level permissions array; the top-level map is pre-existing behavior left unchanged here.
• Frontmatter stripping locates the closing fence by line scan (a --- inside the prompt body survives); an unclosed fence is treated as malformed and skips the injection rather than leaking YAML into the prompt.
• 6 files, +653/-38.

Verification

• :plugin-core:test batches run green on the squashed commit: OpenCodeInlineAgentConfigTest (8/8, including an exact 31-rule pin and a loop over all 7 mutating denies), OpenCodeClientStaticMethodsTest (30/30), ProjectAgentScannerDirectoryScanTest (4/4), OpenCodeClientBehaviorTest (11/11), plus OpenCodeVersionProbeTest and ProjectAgentScannerTest.
• The injected config was fed to a real OpenCode v2.0.27 binary (opencode debug agents) and the merged permission chain inspected: agent rules appended last, execute allowed, mutating tools denied, uncovered actions blocked by the built-in catch-all deny.
• The explore subagent was spawned in a live IDE session on OpenCode v2.0.27: the task completed and every tool call it made was read-only.

How to verify the interactive behavior

1. In an OpenCode v2 session, ask the agent to research part of the project with the explore agent. Expected: the subagent task completes and reports findings.
2. Ask it (via the explore subagent) to modify a file. Expected: refusal, the mutating tools are denied.
3. Open the Agent menu. Expected on v2: Build, Plan, plus project and user-level custom agents only; no General, Explore, or Scout entries, and no duplicate Mode/session-mode sections.
4. Add a <name>.md agent to ~/.config/opencode/agents and reopen the menu. Expected: it appears in the Agent section.

Closes #1187

[github-actions]

⚠️ ClusterFuzzLite — Maintainer Security Review Required

A maintainer needs to approve the ClusterFuzzLite PR fuzzing workflow run for this fork PR.

Before approving, please review the code changes from BambinoSK/agentbridge@b6ba210d to confirm there is no code that could:

• Exfiltrate the CFLITE_STORAGE_REPO_TOKEN CI secret
• Abuse CI infrastructure or the fuzz corpus storage repository

The fuzz targets from the fork branch will be compiled and executed with access to repository secrets. Rubber-stamping this approval without reviewing the code is a security risk.

This comment is posted automatically whenever ClusterFuzzLite is triggered on a fork PR.

[codecov]

Codecov Report

❌ Patch coverage is 87.27273% with 14 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...afishen/agentbridge/client/acp/OpenCodeClient.java	87.91%	8 Missing and 3 partials ⚠️
...en/agentbridge/client/acp/ProjectAgentScanner.java	84.21%	2 Missing and 1 partial ⚠️

📢 Thoughts on this report? Let us know!