Run the permanent checks from a clean checkout:
pnpm exec vitest run server/tool-results.test.ts server/drivers/agents-result.test.ts server/drivers/agents-proxy.test.ts server/agent-tool-policy.test.ts server/tool-results.e2e.test.tsThe end-to-end test launches the shared launchVerificationServer fixture with
its fake engine, temporary home and explicit loopback URL. It uses
control-omb doctor, new-bot, thread-pinned send, messages, and
interrupt. Profile/task setup uses that same isolated server's HTTP API.
It creates 70 idle profiles to produce a genuinely large roster, then calls
the real provider-mounted agents MCP process. No live models, computers,
credentials, user data or external services are used.
The assertions cover:
list_botsreturns a bounded preview and a real saved-result id.tool_result_readpages the missing text without repeating the original operation; every fixture profile remains present across the pages.- The saved result belongs to both the bot and conversation. Another bot or sibling thread cannot read it, even with a valid capability of its own.
- Spoofed conversation ids, unauthenticated requests, oversized writes and invalid offsets are refused. Stop revokes the old capability; a fresh turn on the same conversation can still read its unexpired result.
- Unit/transport checks cover Unicode boundaries, redaction, expiry, cache pressure, missing handles and storage failure. Oversized errors remain errors; success is not turned into failure when caching is unavailable.
The test prints a persistent *.tool-results.json evidence path next to the
server log, containing the control commands, preview and read outcome. Raw
provider launch dumps and their temporary credentials are never retained.
Cleanup stops the exact fixture children and removes their temporary data.
Only built-in agents-tool text results are changed. Replies above 24,000 UTF-16 code units receive a prefix of at most 16,000 plus a short notice. Pages use those same character offsets without splitting surrogate pairs. Cache I/O has a three-second timeout; the original tool call is not retried or cancelled by that timeout. Tool errors keep their error status.
Saved text uses the existing secret-redaction rules and is capped at 128 Ki UTF-16 code units; omitted tails are explicitly marked, never described as fully saved. Redaction is best-effort, not a guarantee that arbitrary sensitive prose can be recognized. Results expire after one hour without read-based renewal, on app restart, or earlier under cache pressure. They are ephemeral, not another persistent history/backup store. The cache holds at most 128 entries and 16 MiB of UTF-8 text total, with at most 16 entries and 2 MiB per bot/conversation; expired entries are swept on access. JavaScript string memory and bookkeeping can exceed the UTF-8 accounting size, but remain bounded.
External MCP trimming, structured shared-computer results, shell command semantics, automatic retries, provider fallback, model/account selection and permissions are unchanged. The new read tool is annotated as a local read, but still requires the live turn's scoped capability. This fixture verifies the real proxy/server path with a fake provider; it does not prove every live provider's UI or Windows/Linux packaging. CI must cover those platforms.