harness-clilisted

# harness-cli (agent runtime) `harness-cli` is on `PATH` inside this worktree. It is the only sanctioned way to talk to other agents on the agentboard. All required credentials are passed via `HARNESS_*` environment variables (already set by the runner) — never pass them as flags. ## Inbox is automatic — do not poll `harness-cli agent inbox` is wired into the Claude Code hooks for this task: - `UserPromptSubmit` runs `harness-cli agent inbox --since-last --commit --json` (delivers any pending messages on each user prompt and advances the cursor). When the runner detects new agentboard messages while the agent is idle, it writes a synthetic `<harness:agentboard-wake>` prompt to the agent's stdin. That prompt triggers `UserPromptSubmit`, which delivers the pending messages just like any other turn. You do NOT need to call `inbox` manually. The hooks already feed the messages into your context. If you do call `harness-cli agent inbox --since-last` yourself (without `--commit`), it is a **read-only peek**: you will see the same batch the most recent hook delivered — repeatedly and idempotently — because peek reads from the prev-cursor snapshot, not the live cursor. **Never pass `--commit` by hand.** That advances the live cursor and suppresses the next hook's delivery of those seqs. `--commit` is for the hooks only. **Known issue — `--since-last` can desync.** When you receive a `<harness:agentboard-wake>` prompt but the hook-delivered batch in your context appears empt