Problem
A "when to use the cloud" scenario doesn't exist on any surface the agent can see, and the cloud read surface is functionally poor. Result (verified): 23 of 24 organic cloud sessions contain an explicit user command — the agent almost never routes to the cloud on its own.
Evidence (audit of 13,105 sessions)
- .with_instructions contains no occurrence of cloud/share/pull/team at all (kaeru-mcp/src/server.rs:700-766). The "Local vs cloud" section in SKILL.md is well written — but the skill isn't installed in any organic project.
- cloud_recall is only a full listing of ONE initiative + get by id; there is no cloud SEARCH — the agent's search reflex (219 recall, 151 search locally) physically can't reach the cloud (cloud_recall=37).
- There's no "list cloud initiatives" tool, though the REST endpoint exists (kaeru-cloud/src/api/router/initiatives.rs:28 — GET /initiatives/). In a real session the user asked for the cloud initiative list → the agent resorted to raw curl against the REST API → the user had to ask why the agent wasn't using kaeru itself.
- Fail-silent: awake/search/overview silently omit that the cloud has more nodes — the agent gets a plausible local answer and never suspects a second tier. In one team initiative, 18 of 25 reading sessions answered "what do we know" from the local graph alone.
- link_cloud=0 and cloud_links=1 in the entire history (dev included): pull strictly dominates — same id, content, and automatic recreate_local_edges; soft-link has no scenario.
Proposal
- A
cloud_initiatives tool (method in cloud_client.rs + #[tool] — REST already exists).
- Cloud search: a query param on cloud_recall, or cloud_search.
- A CLOUD ROUTING section in .with_instructions (3-4 sentences from SKILL.md), respecting the truncation budget.
- Team initiatives: a tail in awake/overview — "cloud: N shared node(s), M not pulled — cloud_recall ".
- link_cloud/cloud_links: either drop from the surface or write a concrete trigger into the description.
Problem
A "when to use the cloud" scenario doesn't exist on any surface the agent can see, and the cloud read surface is functionally poor. Result (verified): 23 of 24 organic cloud sessions contain an explicit user command — the agent almost never routes to the cloud on its own.
Evidence (audit of 13,105 sessions)
Proposal
cloud_initiativestool (method in cloud_client.rs + #[tool] — REST already exists).