subagent: spawn always fails with "free tier can only be used from within OpenCode" on stock install #55471

Activity

1. UnixUserDev commented on Oct 14, 2026

   

   UnixUserDev

   on Oct 14, 2026

   Author

   More actions

   OS clarification: this is Windows 11 Insider Preview (build 10.0.26340, Insider channel). The 10.0.26340 version string in the environment section is the Win11 Insider build number, not Windows 10 — flagging it in case the free-tier gate has any OS/build-dependent path (e.g. launcher detection under an unsigned preview build).
2. opencode-agent commented on Oct 14, 2026

   

   opencode-agentBot

   on Oct 14, 2026 – with OpenCode Agent

   Contributor

   More actions

   Thanks for the detailed report! I think this is a duplicate of #50806. That issue reports the same thing: spawning a custom subagent on a free OpenCode Zen model fails with "OpenCode's free tier can only be used from within OpenCode", while the parent session on the same model works.

   The cause is tracked in #50627. People there found that the free-tier check on the Console side rejects requests whose tool list leaves out some of the standard tools. This most often happens because the agent's permissions deny shell (so bash is dropped), but a small tool allowlist triggers it as well. That would explain why your main agent works and the planner subagent doesn't, even though it's the same model in the same session.

   Possible duplicates and related issues, ordered from most to least likely:

   • subagent: custom agent spawn fails with "free tier can only be used from within OpenCode" #50806: custom subagent spawn fails on the free tier with this exact error (v2)
   • policy: deny shell * on custom agent breaks free tier with "can only be used from within OpenCode" #50627: the underlying cause (restricted tool list or shell deny on a custom agent makes free-tier requests fail); workarounds are discussed there
   • Bug: opencode/xxxxx-model fails when used by a custom subagent #51612: custom subagent using an opencode/*-free model rejected with the same error
   • subagents: explore agent denied free tier can only be used from within OpenCode while running inside the CLI — general works with identical models #49723: the same error for built-in explore subagents while general works

   Until it's fixed, people in #50627 report that changing the custom agent's shell rule from effect: deny to effect: ask, and not limiting it to a small tool allowlist, avoids the error. Using a built-in agent like general also works.

   The bot will close this issue as a duplicate in 1 day unless you reply explaining how your problem is different. For example, tell us if your planner agent has no permission or tool restrictions at all.
3. opencode-agent

   added

   pending close

   triaging

   on Oct 14, 2026
4. UnixUserDev commented on Oct 14, 2026

   

   UnixUserDev

   on Oct 14, 2026

   Author

   More actions

   Confirmed duplicate of #50806 / root cause #50627 — happy for this to close as duplicate.

   Evidence from this workspace (custom planner agent):

   • Before: planner frontmatter had permission: { edit: deny, bash: deny, read: allow, glob: allow, grep: allow }, while the main session config had only allow rules (no deny). That matches the "parent works, subagent fails" pattern exactly — 2/2 spawn failures with the error from this issue on opencode/mimo-v2.6-flash-free.
   • Workaround from policy: deny shell * on custom agent breaks free tier with "can only be used from within OpenCode" #50627 verified live: changing bash: deny → bash: ask made subagent(planner) succeed immediately on the same free model (spawned + completed).
   • Isolation test: kept edit: deny as-is with only bash: ask → spawn still succeeds. So bash: deny is the sole trigger here; edit: deny is harmless. If policy: deny shell * on custom agent breaks free tier with "can only be used from within OpenCode" #50627's check is "any missing standard tool", this narrows it — the shell/bash tool specifically appears to be the one whose absence trips the free-tier gate.
   • 1 of 9 custom agents in this workspace had bash: deny (planner), which also lines up with subagents: explore agent denied free tier can only be used from within OpenCode while running inside the CLI — general works with identical models #49723's some-agents-fail-some-don't behavior.

   Environment detail from the follow-up applies here too: Windows 11 Insider build 10.0.26340, opencode v2.0.27, stock install (no modified app.asar).
5. opencode-agent commented on Oct 14, 2026

   

   opencode-agentBot

   on Oct 14, 2026 – with OpenCode Agent

   Contributor

   More actions

   Thanks for testing this so carefully, and for confirming.

   Your isolation test is useful: bash: deny alone triggers it, while edit: deny doesn't. One caveat is that #54744 reported a case where a small tool allowlist with bash still fails. So bash looks like the most common trigger rather than the only one. Please add your deny→ask and edit: deny results to #50627, where this is tracked and where they'll be seen next to the other reports.

   On the Windows 11 Insider note: the OS shouldn't matter here. The same error has been reproduced on Linux and macOS (see #50627 and #49723), and in your case it went away once you changed only the agent's permissions.

   Until it's fixed, bash: ask on the planner agent is the workaround. This issue will be closed as a duplicate of #50806 as planned. Please follow that issue and #50627 for updates.
6. UnixUserDev commented on Oct 14, 2026

   

   UnixUserDev

   on Oct 14, 2026

   Author

   More actions

   FYI maintainers: the triage job that produced the verdict comment on this issue failed (the verdict itself was posted, then the run crashed):

   • Run/job: https://github.com/anomalyco/opencode/actions/runs/38599144520/job/115973533867
   • Step Triage issue: Process completed with exit code 1

   Log tail:

   16:34:23Z > triage · gpt-5.4-mini
   16:34:42Z (duplicate-verdict comment posted on this issue)
   16:34:43Z Error: {
                "name": "UnknownError",
                "data": {
                  "message": "Unexpected server error. Check server logs for details.",
                  "ref": "err_dd23143e"
                }
              }
   16:34:43Z ##[error]Process completed with exit code 1.
   

   So the triage model emitted its reply and the comment landed, but opencode run then returned an UnknownError server failure ~1s later and the job went red. Worth checking server logs for ref err_dd23143e in case any post-comment triage actions (label transitions / duplicate-close scheduling beyond the visible pending close + triaging labels) were skipped.

   Separately, the verdict itself checks out: I've since verified the #50627 workaround locally (see comment above) — bash: deny on the custom agent was the sole trigger, bash: ask fixes the spawn on opencode/mimo-v2.6-flash-free.
7. opencode-agent

   closed this as duplicateon Oct 15, 2026
8. opencode-agent commented on Oct 15, 2026

   

   opencode-agentBot

   on Oct 15, 2026 – with OpenCode Agent

   Contributor

   More actions

   Closing this as a duplicate of #50806, as planned. Your own tests confirmed it's the same problem: the bash: deny rule on your custom planner agent trips the free-tier check on opencode/mimo-v2.6-flash-free, and changing it to bash: ask makes the subagent spawn work.

   The cause is tracked in #50627, so please follow that issue and #50806 for updates. Your isolation results (bash: deny fails, edit: deny alone is fine) would help most there. Until it's fixed, bash: ask is the workaround. Thanks again for the careful testing!