# HANDOFF.md

## Purpose
Immediate baton-pass for the next fresh implementation session.

## Current objective
Investigate and improve subagent / ACP delegation reliability with evidence-first debugging. The subagent persistence/announcement fix and the raw `agent.wait` semantics fix are now both live-verified on branch `fix/subagent-wait-error-outcome`; the next work should stay tightly scoped to ACP-specific Claude/Codex follow-up. This pass already narrowed that thread to a real bundled-acpx parser bug for Claude-style JSON-RPC auth failures and landed a focused fix/tests. The remaining work is end-to-end OpenClaw ACP-path validation (or a fresh repro of the older exit-code crash notes) plus normal commit/push/PR cleanup when desired.

## Use these state files first
1. `WIP.subagent-reliability.md` — canonical state for this pass
2. `memory/tasks.json` — task tracking for reliability items
3. `memory/2026-03-04-subagent-delegation.md` — earlier delegation context
4. `memory/2026-03-13.md` if present, otherwise append today’s evidence there
5. `external/openclaw-upstream/` — for any core-runtime fix work

## Related tasks
- `task-20260304-2215-subagent-reliability` — in progress
- `task-20260304-211216-acp-claude-codex` — open

## Known truths
- TUI noise suppression was already patched locally and upstreamed earlier.
- User still wants actual subagent reliability improved, not just UI noise hidden.
- Historical ACP notes included `Claude: acpx exited with code 1` and `Codex: acpx exited with code 5`, but those exact crashes were **not** reproduced in the latest pass.
- Fresh-session implementation discipline is now the expected approach for non-trivial work.
- One explicit failure mode is already understood: requesting `glm-5` can route into an unavailable GLM-5 provider/entitlement path in this setup.
- A deeper bug was also identified and fixed earlier: a subagent run could finish with terminal assistant errors yet still be recorded as successful with no frozen result.
- Current host state for ACP follow-up:
  - bundled plugin-local `acpx` exists and runs
  - `~/.openclaw/openclaw.json` currently has no explicit `acp` block / enabled `acpx` plugin entry, so this pass used the smallest direct acpx repro path instead of a full OpenClaw ACP session
- New confirmed acpx/runtime bug from this pass:
  - Codex direct acpx path works
  - Claude direct acpx path returns top-level JSON-RPC auth errors (`Authentication required`) and exits `0`
  - `extensions/acpx/src/runtime-internals/events.ts` previously dropped that JSON-RPC error shape during prompt streaming, so OpenClaw could falsely treat the turn as successful
- A focused upstream fix for that runtime bug now exists on `fix/subagent-wait-error-outcome` with targeted tests passing.

## Highest-priority next actions
1. Treat the generic reliability fixes as live-verified on this branch:
   - subagent persistence/announcement proof:
     - run id `b50cb91f-6219-44f7-9d2f-a1264ac7ceaf`
     - child transcript `~/.openclaw/agents/main/sessions/f114b831-000b-4070-a539-85c68d2b7057.jsonl`
     - `runs.json` stores `outcome.status: "error"`, `endedReason: "subagent-error"`, and a non-null `frozenResultText`
   - raw `agent.wait` live-fix proof:
     - gateway launch: `OPENCLAW_SKIP_CHANNELS=1 CLAWDBOT_SKIP_CHANNELS=1 pnpm exec tsx src/index.ts gateway run --port 18903 --bind loopback --auth none --allow-unconfigured`
     - run id: `gwc-live-agent-wait-gpt53-source-fixed2-1773429512008`
     - final `agent` response: `finalStatus:"error"`
     - `agent.wait`: `{"runId":"gwc-live-agent-wait-gpt53-source-fixed2-1773429512008","status":"error","endedAt":1773429514106,"error":"LLM request rejected: Your input exceeds the context window of this model. Please adjust your input and try again."}`
2. Treat the ACP follow-up as partially closed, not fully done:
   - live direct bundled-acpx Codex repro now works and returns `OK`
   - live direct bundled-acpx Claude repro returns JSON-RPC auth errors with process exit `0`
   - focused upstream fix now maps top-level JSON-RPC prompt errors into ACP runtime `type:"error"` events instead of silently dropping them
   - targeted validation passed:
     - `cd external/openclaw-upstream && pnpm exec vitest run extensions/acpx/src/runtime-internals/events.test.ts extensions/acpx/src/runtime.test.ts extensions/acpx/src/runtime-internals/control-errors.test.ts`
     - result: `22` tests passed across `3` files
3. Next, do end-to-end OpenClaw ACP validation if/when ACP is explicitly enabled here:
   - confirm or add the needed `acp` / `acpx` config in `~/.openclaw/openclaw.json` (or equivalent current config path)
   - run the smallest real OpenClaw ACP turn/session and confirm Claude auth failures now surface as terminal errors instead of false success
   - only reopen the old `acpx exited with code 1/5` thread if a fresh repro appears
4. Commit/push/PR the focused upstream reliability branch when ready.
5. Leave the dirty `/subagents log` UX diff out of this branch unless you intentionally spin a separate focused pass; it regression-passed `src/auto-reply/reply/commands.test.ts` but still lacks dedicated feature coverage.

## Success criteria
- Real-run verification of the new error/outcome fix. ✅ done for subagent persistence/announcement handling.
- Clear separation between resolved reporting bug(s) and any still-open ACP/runtime failures.
- Explicit decision on whether raw `agent.wait` behavior is acceptable or requires a follow-up fix.
- State files updated with paths, commands, and outcomes.