Repository navigation
Follow-up (#2229): ACP-runtime agent entry for OpenClaw→OpenCode dispatch (opencode-acp) #2256
Description
Follow-up to #2229 (method 1 of 3). Method 3 — the opencode-route provider front — shipped as the priority build (PR #2250, live-proven). This method was descoped by Drew on 2026-10-29 to a follow-up: sometime soon / when an OpenClaw-origin workload needs ACP dispatch into OpenCode.
What it adds
An ACP-runtime agent entry in home/openclaw/default.nix (agents.entries): opencode-acp with runtime.type="acp", runtime.acp.agent="opencode", backend="acpx", mode="oneshot", heartbeat disabled. Per the shipped OpenClaw runbook, an entry configured with runtime.type="acp" makes it a valid sessions_spawn({runtime:"acp"}) target and binding-default holder, with runtime.acp.agent naming the underlying harness (opencode). That gives OpenClaw a standing capability to dispatch a job into OpenCode over ACP, instead of only reaching OpenCode models through the provider front.
Starting point (already done pre-descope)
- PR feat(openclaw): opencode-acp ACP-runtime agent (#2229 mechanism 1) #2237 (branch
feat/openclaw-acp-agent-2229) already contains the entry. The shape was verified against the pinned OpenClaw 2026.10.7 zod schema —modemust bepersistent|oneshot(an earlierrundraft was rejected byopenclaw config getand corrected). - Established semantics: a direct
openclaw agent --agent opencode-acpturn still runs embedded on the default model — the entry's value is as a spawn target, so its live proof depends on the sessions_spawn follow-up (headless/CLI sessions currently never receive that tool). Once that lands, prove this with a spawn toagentId="opencode-acp"and check theacp_sessionstable plus a new row in OpenCode'ssession_v2.
Activity
developing-today commentedon Oct 29, 2026
CODE-1119
fix(openclaw): remove opencode-acp entry that cannot evaluate
dezren39 commentedon Oct 30, 2026
Closed by Drew's decision (2026-10-29, main chat): the embedded runtime is what he wants for OpenClaw's agents. The standing ACP-runtime agent (opencode-acp) is therefore not wanted, not merely blocked — this restore is cancelled, not deferred. OpenClaw spawning OpenCode over ACP on demand from sessions (#1728) remains enabled and is unaffected. The nix-openclaw schema bug that made the entry unevaluable is still tracked at #2509 / upstream openclaw/nix-openclaw#176 for what it's worth to the generator, but nothing here waits on it anymore.
dezren39 commentedon Oct 30, 2026
Reopened on Drew's directive (2026-10-29 21:44 CDT, main chat): "I want the other options built and functional." This supersedes the close from ~25 minutes earlier — embedded stays the default runtime for OpenClaw's agents, AND the ACP-runtime agent option is to be built and proven functional as a standing capability. Build note: the nix-openclaw schema cannot express the entry (#2509, upstream openclaw/nix-openclaw#176); the working route is to carry a patch against the nix-openclaw module in this repo (single-submodule runtime type) or an equivalent local encoding, proven by a full toplevel evaluation WITH the entry present. Acceptance: entry evaluates, a switch activates it, and 'openclaw agent --agent opencode-acp' completes a trivial dispatched turn. Scope clarification (Drew, 21:45 CDT): ACP dispatch is task handoff — OpenCode doing the work there is the point of the option. The routing-only principle in the new follow-up issue applies to the opencode-route MODEL door, not to dispatch.
dezren39 commentedon Nov 3, 2026
Claimed by the OpenChamber scheduled queue lane (task 7860, triage, 20 min box). Revalidating against current
main: entry still absent, blocker #2509 / upstream openclaw/nix-openclaw#176 still open at the pinned rev. Scope + ordering follow in this same run.dezren39 commentedon Nov 3, 2026
Triage at current
main(f04a439, 2026-11-03)Verdict: live, unbuilt, and correctly ordered behind #2509. Not closed.
Verified state
home/openclaw/default.nix(~line 377) has only the removal comment: theopencode-acpentry is absent, pointing at OpenClaw ACP-runtime agent entry cannot evaluate: nix-openclaw generated schema makes the acp runtime branch unreachable #2509 and this issue.2256/opencode-acp).priority/high) and carries a queued queue task. Upstream openclaw/nix-openclaw#176 is still open and its review was run against rev4b261b08c173— exactly ourflake.lockpin — so there is no upstream fix to bump to.Scoped fix path
The only viable route is a local workaround, because the pinned input cannot express
runtime.type=acp:type = enum [ "embedded" "acp" ]plus optionalacpattrset, keeping the rendered JSON shape identical. This repo already accepts exactly this pattern for a nix-openclaw piece:home/openclaw/plugins/summarize/flake.nixis vendored for the same class of upstream-pin defect.opencodeandopenchamber; do not addnix-openclawas a fork.Ordering / blockers recorded
opencode-acpdispatch actually works, plus the README fix.openclaw agent --agent opencode-acpturn still runs embedded on the default model — the entry's value is as asessions_spawntarget. So the acceptance in your 21:44 CDT comment ("openclaw agent --agent opencode-acpcompletes a trivial dispatched turn") may not be satisfiable by the entry alone. The implementing worker must record which path actually proved dispatch (acp_sessionsrow + newsession_v2row), and if it is only reachable viasessions_spawn, note what remains for the headless-sessions follow-up rather than claiming a false proof.home/openclaw/README.mdstill says OpenCode is the only allowed ACP agent whiledefault.nixallowsopencode/claude/codex/agy(openclaw: enable claude, codex, agy as ACP agents alongside opencode #1728), and still says entries-based ACP agents are not used.Follow-up
Enqueued implementation task 8040 (
repo-health:v1:fix:developing-today/code:issue:2256),not_before2026-11-03T20:00Z, with the #2509 ordering gate written into its acceptance criteria. No new labels needed on the issue beyond the queue lifecycle ones already applied.116 remaining items
dezren39 commentedon Nov 5, 2026
🤖 Agent note · picked up
dezren39 commentedon Nov 5, 2026
🤖 Agent note · parked
dezren39 commentedon Nov 5, 2026
Drew's adjudication (2026-11-05 01:44 CST, main chat): embedded is what I want — the Oct 30 ruling stands; the ACP-runtime agent entry is NOT restored as a standing agent. Two refinements recorded: (1) the architecture focus is OpenClaw calls OpenCode, and OpenClaw handles the loop — Shed drives, OpenCode is the called backend. (2) Drew wants to know ACP could work: a bounded capability proof (OpenClaw driving its own loop, dispatching real work to OpenCode over ACP, end-to-end) is spun out as its own issue — a proof, not a restore, and it changes no standing config. Closing this issue as not planned per the ruling; the reopen on 2026-10-30T02:45Z has no recorded basis and is superseded by this comment.
dezren39 commentedon Nov 5, 2026
agent-note:v1
state: done (task 8040)