Subagent relay turns pay full cold-cache bootstrap re-injection tax

OpenClaw version: 2026.4.8 (9ece252) Severity: Cost / performance (per-delegation tax, ~$0.10–0.15 on Sonnet) Impact: Every sessions_spawn → subagent → relay cycle rewrites the entire cached prefix on the parent agent, regardless of whether the underlying workspace bootstrap changed.

Symptom

When a parent agent (e.g. main/Reeves) delegates to a subagent via sessions_spawn, yields its turn, and later receives the subagent's task-completion event, the follow-up relay turn shows:

cacheRead:  0
cacheWrite: ~25,000–40,000 tokens

The cache prefix has been fully rewritten, even when:

• The relay follows the delegation by only ~30–60 seconds
• Anthropic's ephemeral cache (5-min default, 1-hour with cacheRetention: "long") has not expired
• The workspace bootstrap files on disk have not changed

A working example of the cost profile:

The second turn's cost is dominated by the cold cache rewrite of ~37K tokens of bootstrap + history that did not change between the two turns.

Root cause

File: dist/pi-embedded-CNTNdlGw.js Lines: 31273–31274

const isContinuationTurn = resolveContextInjectionMode(params.config) === "continuation-skip"
    && params.bootstrapContextRunKind !== "heartbeat"
    && await hasCompletedBootstrapTurn(params.sessionFile);

const shouldRecordCompletedBootstrapTurn = !isContinuationTurn
    && params.bootstrapContextMode !== "lightweight"
    && params.bootstrapContextRunKind !== "heartbeat";

The relay turn that processes a subagent's task_completion event is invoked with bootstrapContextRunKind: "default". There is no "relay" case, and no code path that detects "this incoming event is a subagent completion that should inherit the parent session's existing context rather than re-inject bootstrap."

As a result:

1. isContinuationTurn evaluates to false even though the relay follows an already-completed bootstrap turn (because the continuation-skip workaround only fires if the config opts in, AND the subagent event is treated as a new "user message" that requires full-bootstrap)
2. shouldRecordCompletedBootstrapTurn evaluates to true
3. A openclaw:bootstrap-context:full custom event fires
4. The context engine rebuilds the full bootstrap prefix for the relay turn, invalidating the prompt cache

Confirmation in session trace (/path/to/agent/main/sessions/<id>.jsonl):

{"type":"custom","customType":"openclaw.cache-ttl",...}
{"type":"custom","customType":"openclaw:bootstrap-context:full",...}

These two events fire in sequence right before every relay turn, not on any other turn type.

Partial workaround (what I've tried)

Setting agents.defaults.contextInjection: "continuation-skip" does reduce the emission frequency of openclaw:bootstrap-context:full — it no longer fires on every CLI call within the same session — but the relay turn itself still pays the cache write because the subagent completion message is treated as new content requiring the full bootstrap prefix, not as a continuation.

Tested result with continuation-skip enabled:

• Inter-call bootstrap re-emission: eliminated ✅
• Relay turn cache rewrite: still ~23K tokens cacheWrite, cacheRead=0 ❌

Proposed fix

Introduce a dedicated bootstrapContextRunKind: "relay" (or "subagent_return") value and exempt it from full bootstrap re-injection the same way "heartbeat" is exempted:

// dist/pi-embedded-CNTNdlGw.js:31273
const isContinuationTurn = 
    (resolveContextInjectionMode(params.config) === "continuation-skip" 
     || params.bootstrapContextRunKind === "relay")
    && params.bootstrapContextRunKind !== "heartbeat"
    && await hasCompletedBootstrapTurn(params.sessionFile);

// dist/pi-embedded-CNTNdlGw.js:31274
const shouldRecordCompletedBootstrapTurn = !isContinuationTurn
    && params.bootstrapContextMode !== "lightweight"
    && params.bootstrapContextRunKind !== "heartbeat"
    && params.bootstrapContextRunKind !== "relay";

And the server-side subagent completion handler (likely in dist/server.impl-WjqjRArz.js) should set bootstrapContextRunKind: "relay" when invoking the embedded Pi runner to process a task_completion event targeted at a parent session.

Estimated saving

Per delegation, on Sonnet 4.6:

• Current: ~$0.14 structural tax per relay turn (37K token cache rewrite)
• After fix: ~$0.01–0.02 (cache hit on unchanged prefix + small cache extension for the relay content)
• Savings: ~$0.12 per delegation, or ~87% of the relay turn cost

For any fleet doing meaningful parent→subagent orchestration (Chief of Staff → CTO → technician patterns, model comparison protocols, etc.), this compounds fast.

Related

• Subagent session keys are hardcoded to crypto.randomUUID() (dist/pi-embedded-CNTNdlGw.js:13901), which means every delegation also pays a cold-start bootstrap on the subagent side — see separate issue 02-subagent-session-reuse.md.
• Together these two issues make parent → subagent delegation ~10× more expensive than it structurally needs to be.