Skip to content

Commit 43a0faa

Browse files
docs: journal that OpenCode custom subagents work (gh-493)
Verified live on opencode 1.18.31: all three definition forms are read, and task-tool dispatch produces a genuine child session with parent_id set and the custom agent name in the agent column. Upstream anomalyco/opencode#29616 is stale, so the caveat I added to #493 was wrong and is corrected there. Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
1 parent 0095401 commit 43a0faa

1 file changed

Lines changed: 5 additions & 0 deletions

File tree

‎docs/superpowers/journals/specs/2026-09-19-presentation-v2.md‎

Lines changed: 5 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -74,3 +74,8 @@ Harness confirmed as opencode --auto with github-copilot/gpt-5.6-terra, default
7474
### v2-opencode-parity-costs-resolved · finding [refuted] · resolves v2-opencode-parity-costs: The confirmed harness removes the Extensibility beat and makes the contract beat advisory
7575

7676
Half refuted by the operator, who asked why OpenCode subagents 'were not implemented'. The challenge holds and my account was wrong twice. (1) I attributed arm A's 14 sessions / 13 subagents to Claude. Both runs were OpenCode -- the #429 experiment's runbook puts every arm on opencode --auto, and arm A's 13 child sessions carry parent_id and agent=general. OpenCode HAS a dispatch primitive: Task tool plus @mention over built-in general/explore/scout. fr-goal SKILL.md:91's 'OpenCode has no dispatch primitive of its own' is false as written. (2) The parity scope_note blames the lack of an isolation-argument dispatch primitive, which cannot be the blocker: fr-isolation-required.md:149 names Hermes' delegate_task and OpenCode TOGETHER as lacking one, yet Hermes is declared enforced; and the fr-phase-executor carve-out FORBIDS an isolation argument for phase executors, so the surface requires its absence. Filed as gh-493. The narrower real limitation is that custom NAMED subagents may not be invocable on OpenCode (upstream anomalyco/opencode#29616) and super-fr ships no OpenCode agent definition to test it. Measured cost of the capability: arm A's 13 subagents cost 7.59 USD against roughly 1 USD inline, about 7x -- though arm A was also FASTER in wall clock, 56.2 min against 77.6 and 105.2 for the single-session arms, so it is a cost-versus-speed-and-isolation tradeoff rather than a plain loss. NOT refuted, and unchanged for the recording: operator-gate/opencode is advisory, both question gates can silently not fire, and the take-acceptance criteria stand. The practical plan is also unchanged, because 4.5.x's SHIPPED behaviour is inline regardless of how the declaration is worded.
77+
78+
<!-- fr:journal kind=discovery scope=spec id=v2-opencode-custom-subagents-work created=2026-09-19T23:26:35 -->
79+
### v2-opencode-custom-subagents-work · discovery · Custom subagents work on OpenCode 1.18.31 — definition AND dispatch, verified live
80+
81+
Operator proposed adding agent keys for OpenCode and invoking with @, believing subagents should reach more harnesses than Claude. Tested rather than assumed, and it works. Definitions are read in all three documented forms -- opencode.json 'agent' key, .opencode/agents/*.md, and .opencode/agent/*.md -- all three appeared in 'opencode agent list' as (subagent). Dispatch works too: 'opencode run' with a task-tool delegation returned the subagent's reply, and the session table shows a genuine CHILD session with parent_id set, its own cost (0.0239) and tokens (9510/8), and the CUSTOM agent name in the agent column. Structurally identical to arm A's 13 @general rows but with a named custom agent. Upstream anomalyco/opencode#29616, which I cited in gh-493 as a narrower real limitation, is STALE -- filed 2026-05-27 and fixed by the installed 1.18.31. So my caveat in that issue was wrong too, and I have corrected it there with the transcript. Consequence: subagent-dispatch/opencode: absent is simply wrong; both halves of the surface work. OpenCode agents also accept a model: field, so per-phase model tiering -- what fr models exists for -- is reachable, not only context isolation. Implementation is small and rides sync-opencode.py's existing mirroring, but needs a frontmatter translation (name -> filename, tools -> permission, add mode: subagent) AND a correction to fr-goal SKILL.md:91, without which the skill still instructs inline execution however many agents are defined.

0 commit comments

Comments
 (0)