[Dhravya]

By default the plugin now makes zero Supermemory calls from its before_prompt_build (recall) and agent_end (capture) hooks for OpenClaw subagent sessions, host helper runs, host-generated inputs, and headless runs. Interactive TUI / Control UI / channel turns behave exactly as before. Two opt-in flags restore today's behaviour.

What is detected and how (OpenClaw's own signals)

Checked against openclaw/openclaw main @ 7ab0143 and the published openclaw@2026.11.1 CLI.

Class	Host signal	Source
Subagents	ctx.sessionKey contains a subagent: segment (agent:<id>:subagent:<uuid>, incl. incognito- and nested)	docs/tools/subagents.md ("Each one runs in its own session (agent:<agentId>:subagent:<uuid>)"), src/sessions/session-key-utils.ts isSubagentSessionKey / getSubagentDepth
Host helper runs	session-key segments the host mints for its own runs: active-memory: (active-memory recall run, extensions/active-memory/recall-run.ts), helper: (session-memory slug generator, src/hooks/bundled/session-memory/handler.ts), update-repair: (src/infra/update-repair-agent.runtime.ts), internal-session-effects: (session forks, side-chat companion, memory flush, src/config/sessions/internal-session-key.ts), dreaming-narrative- (isDreamingNarrativeSessionStoreKey)	as listed
Host-generated input (e.g. a subagent announcing its result into the parent session)	inputProvenance.kind is inter_session or internal_system: ctx.inputProvenance on before_prompt_build, the last user message's provenance on agent_end	src/sessions/input-provenance.ts isHostGeneratedInputProvenance; docs/plugins/hooks/prompt-and-session.md ("ctx.trigger === "user" is a run trigger, not an origin classification … Use typed provenance to distinguish those inputs")
Headless	OpenClaw has no explicit "headless" flag in hook context. openclaw agent exec ("Run one isolated headless embedded agent turn", src/cli/program/register.agent-turn.ts; "exec matches no channel", src/commands/agent-exec-input.ts) and scripted openclaw agent --message turns run with no messageProvider/channel, while every interactive surface stamps one (TUI/Control UI webchat = INTERNAL_MESSAGE_CHANNEL, chat channels, node, system agent openclaw). A run is treated as headless when ctx.trigger is set and both fields are empty; hosts that predate ctx.trigger keep the legacy behaviour.	src/utils/message-channel-constants.ts, src/agents/command/run-embedded-attempt.ts, observed ctx below

Existing cron / heartbeat / memory-flush skips are unchanged. Subagents inherit messageProvider: "webchat" and trigger: "user" from their parent (observed below), so the session key is the only usable subagent signal.

Opt-in flags (plugin config in openclaw.json, or env for the OpenClaw process; each re-enables both recall and capture for its class):

• captureSubagents: true / SUPERMEMORY_CAPTURE_SUBAGENTS=true: subagents, helper runs, host-generated inputs
• captureHeadless: true / SUPERMEMORY_CAPTURE_HEADLESS=true: headless runs (a subagent spawned from a headless run needs both)

Stored data, container tags, customIds, and metadata shapes are untouched; explicit memory tool calls are unaffected. README and CHANGELOG are updated.

Verification on the real host

openclaw@2026.11.1 from npm, gateway + openclaw tui driven in a tmux pty (interactive), openclaw agent exec and openclaw agent --message (headless). The model was a local OpenAI-compatible stub (models.providers.stub, api: "openai-completions") that returns a sessions_spawn tool call when the prompt says SPAWNCHILD, so the host spawned a real agent:main:subagent:<uuid> session. The plugin was pointed at a recording Supermemory stub (baseUrl: http://127.0.0.1:18081, dummy sm_ key) that logs every method+path. The same 4-step script ran per scenario (stub servers and scripts are not committed):

Step	main (before)	this PR default	captureSubagents config	SUPERMEMORY_CAPTURE_SUBAGENTS=true	captureHeadless config	SUPERMEMORY_CAPTURE_HEADLESS=true
1 interactive TUI turn	2	2	2	2	2	2
2 TUI turn that spawns a subagent	6	2 (parent turn only)	6	6	2	2
3 openclaw agent exec	2	0	0	0	2	2
4 openclaw agent --message (gateway, no channel)	2	0	0	0	2	2
total	12	4	8	8	8	8

Default run, Supermemory stub log (all calls):

STEP 1 interactive TUI turn
  POST /v4/profile q="I prefer dark roast coffee every morning"
  POST /v3/documents content="[role: user]\nI prefer dark roast coffee every morning ..."
STEP 2 TUI turn that spawns a subagent (sessions_spawn)
  POST /v4/profile q="Please SPAWNCHILD a worker to check the build"
  POST /v3/documents content="[role: user]\nPlease SPAWNCHILD a worker to check the build ..."
STEP 3 headless: openclaw agent exec
STEP 4 headless: scripted openclaw agent --message via gateway
total supermemory calls: 4

Hook contexts the host passed in that run (from a second, logging-only plugin), showing the subagent and the announce turn really ran:

before_prompt_build agent:main:main trigger=user provider=webchat
subagent_spawned child=agent:main:subagent:d19f29a6-…
agent_end agent:main:main trigger=user provider=webchat lastUser.provenance={"kind":"external_user"}
before_prompt_build agent:main:subagent:d19f29a6-… trigger=user provider=webchat
agent_end agent:main:subagent:d19f29a6-… trigger=user provider=webchat
before_prompt_build agent:main:main trigger=user provider=webchat ctx.inputProvenance={"kind":"inter_session",…,"sourceTool":"subagent_announce"}
agent_end agent:main:main trigger=user provider=webchat lastUser.provenance={"kind":"inter_session",…,"sourceTool":"subagent_announce"}
before_prompt_build agent:main:explicit:c3873ca1-… trigger=user provider=undefined   (agent exec)
before_prompt_build agent:main:main trigger=user provider=undefined                    (agent --message)

On main, the 4 extra calls in step 2 were the subagent's own profile + capture and the parent's capture of [Inter-session message] sourceSession=agent:main:subagent:….

The helper-run session keys (active-memory, helper, update-repair, internal-session-effects, dreaming) were verified by replaying the host's key shapes and triggers through the hook handlers, not by driving those host features live: 0 calls by default, 2 with captureSubagents. Interactive webchat, a Telegram-shaped context, the system agent (manual / openclaw), and a legacy context without trigger keep their 2 calls in every configuration.

bun run check-types and biome ci . (CI's 2.3.8) pass. The repo has no test suite.

Not detectable / out of scope

• OpenClaw exposes no explicit headless flag to plugins. A scripted gateway turn that names a channel (openclaw agent --channel … --deliver) carries that channel and is treated as interactive.
• ACP sessions (agent:<id>:acp:…, external Codex/Claude harnesses via sessions_spawn runtime: "acp") run outside OpenClaw's embedded hooks, so this plugin never saw them and nothing changes.
• Session titles and conversation labels (src/gateway/dashboard-session-title.ts, src/auto-reply/reply/conversation-label-generator.ts) go through runUtilityCompletion (src/agents/utility-completion.ts), a plain completion that never fires before_prompt_build/agent_end, so the plugin never saw them. I did not find a separate branch-name generator.