Subagent Task Worker — Pitfalls and Examples

You're seeing this skill because the dispatcher spawned you for a subagent task — your Session.task_id is set and the harness has exposed the worker_complete, worker_block, and worker_context self-tools to you. This skill is the deeper detail beyond the system-prompt basics: good handoff shapes, retry diagnostics, edge cases, and what NOT to do.

See Tasks (Subagent Task Layer) for the conceptual chapter and Tools for the parameter tables.

What you have access to

You're running for a Task. Three self-tools are available only to you, gated by Session.task_id:

Tool	Use it when
worker_context	At the start of your run, especially on retry. Returns goal, accumulated context (from prior unblocks), parent task results, and prior attempt summaries.
worker_complete	When you've actually finished. Writes the explicit summary + structured metadata to the task row; the parent agent sees this in its worker.complete event.
worker_block	When you need new context that isn't available without a human or peer providing it. Does NOT consume a retry attempt.

You also have the same tool surface a spawn_worker child would have: read/write files, web, terminal, etc. -- subject to the AgentDef filter pinned to this attempt.

Step 0 — Orient yourself

The first thing you should do on a non-trivial task is worker_context. It tells you:

• The original goal.
• Any context accumulated from prior unblocks (timestamped).
• attempt_count -- if > 1, you are a retry. Read the prior_attempts list before doing anything.
• parents -- if non-empty, each entry's result and result_metadata is the upstream work you're building on. Don't re-derive it.

Your initial USER_MESSAGE already includes a short summary of prior attempts (when attempt_count > 1), bounded to the last 5. worker_context is how you read the full detail.

Good summary + metadata shapes

worker_complete(summary, metadata) is how downstream readers (the parent agent, future retries, humans) understand what you did. Aim for summary to be 1-3 sentences a human can scan; metadata to be machine-readable facts.

Coding task:

worker_complete(
    summary="shipped rate limiter — token bucket keyed on user_id with IP fallback; 14 tests pass",
    metadata={
        "changed_files": ["rate_limiter.py", "tests/test_rate_limiter.py"],
        "tests_run": 14,
        "tests_passed": 14,
        "decisions": ["user_id primary; IP fallback for unauthenticated requests"],
    },
)

Coding task that needs human review:

For most code-changing work, "done" should mean "human reviewed and approved." Use worker_block instead of worker_complete, with a reason that starts review-required:. Leave the structured info (diff path, test counts, what to look at) for the parent agent or human to discover via worker_context.

# (NOT worker_complete — block instead so a reviewer steps in)
worker_block(
    reason=(
        "review-required: rate limiter shipped, 14/14 tests pass — "
        "needs eyes on the user_id-vs-IP fallback choice before merge"
    ),
)

A reviewer (human, or another task spawned by the orchestrator) then calls unblock_task to resume you, OR cancel_task if a fresh attempt is wanted.

Research task:

worker_complete(
    summary="3 inference servers reviewed; vLLM wins on throughput, SGLang on latency, TRT-LLM on memory",
    metadata={
        "sources_read": 12,
        "recommendation": "vLLM",
        "benchmarks": {"vllm": 1.0, "sglang": 0.87, "trtllm": 0.72},
    },
)

Review task:

worker_complete(
    summary="reviewed PR #123; 2 blocking issues: SQL injection in /search, missing CSRF on /settings",
    metadata={
        "pr_number": 123,
        "findings": [
            {"severity": "critical", "file": "api/search.py", "line": 42, "issue": "raw SQL concat"},
            {"severity": "high", "file": "api/settings.py", "issue": "missing CSRF middleware"},
        ],
        "approved": False,
    },
)

Shape metadata so downstream parsers (the parent orchestrator, an aggregator task, a reviewer skill) can use it without re-reading your prose.

The coordination board — share as you work

Because you were spawned into a fan-out, you also have share_note, read_board, and expand_note: a shared, verified board that all sibling workers, retries, and the coordinator read. This is how your discoveries help peers while they work, instead of only at handoff time.

When to write (batch related notes into one share_note call):

• FAIL — highest value, post immediately. You tried something and it dead-ended: name what you ran and the observed error. A sibling is one note away from repeating it.
• FACT for reusable knowledge anchored to specifics (path, symbol, endpoint, error class). Not narration — anchors a peer can act on.
• CLAIM before starting a substantial unit of work that a sibling might also pick ("claiming the slack adapter"). Expires automatically; re-post to renew.
• RESULT when you have a candidate outcome, as outcome=…|evidence=…|risk=… where the evidence names a check you ACTUALLY ran and what it printed. Your latest RESULT replaces your previous one.

When to read: [Board update] messages arrive in your history automatically as peers post. Additionally call read_board at decision points — before committing to an approach, and before retrying anything that smells like a known dead end — because inline updates may be stale (superseded results, expired claims).

Notes are admission-verified: vague or unevidenced notes are rejected with a reason. Write them anchored the first time. Sharing on the board does NOT replace worker_complete — the board is live coordination; the completion summary is your durable handoff.

Block reasons that get answered fast

Bad: "stuck". The human or parent has no context.

Good: one sentence naming the specific decision. If you need more context to justify the question, do it in the work you've done so far -- don't stuff a paragraph into the reason. The parent/human can call worker_context to read your full state.

worker_block(
    reason=(
        "Rate-limit key choice: should I key on IP (simple, NAT-unsafe) "
        "or user_id (requires auth — skips anonymous endpoints)?"
    ),
)

Retry scenarios

When worker_context returns task.attempt_count > 1, you are a retry. The prior_attempts array tells you what earlier sessions did:

• outcome: "completed" -- they emitted a structured summary but the task was reopened (rare). Their summary is in the entry. Don't redo their work.
• outcome: "blocked" -- a previous attempt blocked; an unblock_task re-launched you. Read the accumulated task.context for what was added on unblock.
• outcome: "crashed" -- the prior session ended without emitting either a complete or block event (crash, OOM, timeout, hard-kill). No structured summary is available; check the parent's events or worker_context for any clue.

Don't repeat what failed. If three prior attempts crashed at the same step, change your approach, narrow scope, or block for guidance.

Tenant isolation

If your session has org_id set, you are scoped to that tenant. Any persistent memory you write should be prefixed by the tenant so context doesn't leak across orgs. The Surogates memory tool generally namespaces by tenant automatically; if you write directly to shared scratch files, prefix manually.

Do NOT

• Call delegate_task as a substitute for spawn_task. delegate_task is a synchronous fork-join for short reasoning subtasks inside YOUR run; spawn_task is for durable cross-agent handoffs that outlive one API loop.
• Call spawn_task if you are a leaf worker. Children spawned by either spawn_worker or spawn_task have WORKER_EXCLUDED_TOOLS applied; if spawn_task is in your toolset, you're an orchestrator-shaped session (and you should see the orchestrator skill instead).
• Complete a task you didn't actually finish. Use worker_block to ask for help; the retry budget is not consumed by blocking.
• Modify files outside your sandbox workspace unless the task body says to. The parent's workspace is shared via inheritance; don't surprise it.
• Hand-write task ids into your prose. When you spawn child tasks (only if you're an orchestrator-shaped worker), keep the returned task_id from each spawn_task call and reference them in your worker_complete summary by quoting the actual return value, not making one up.

Pitfalls

Task state can change between dispatch and your startup. Between when the dispatcher claimed the task and your process boot, the task may have been cancelled or reblocked. Always worker_context first. If the task is no longer running (e.g. cancelled or blocked), stop — you shouldn't be doing the work.

Your attempt may have been reclaimed. If the session lease expired while you were inside a long-running tool call, the dispatcher's stale-claim recovery may have started a new attempt. The worker_complete / worker_block tools refuse with "this attempt is no longer the current task attempt" — that's the signal. Exit cleanly; the new attempt has the work.

Don't rely on a CLI. The task_* tools work uniformly across all execution backends (sandbox, Modal, remote SSH). There is no surogates kanban CLI to fall back on — use the tool surface.

Read your parents' results. When the task has parents, each parent's result and result_metadata is the upstream work. The orchestrator placed you here because their output is your input. Read it via worker_context; don't re-derive.