- tc-tools: Effect declarations -> §15 GatedCategory mapping, deny-by-default external reach, taint invariant property-tested (tainted external effects are NEVER auto-allowed) - tc-safety: pending approvals with exact payload+preview, CAS decide with audit + single-use grant in one tx, checkpoint suspend/load, exclusive resume claim, expiry sweep, decided-unresumed work queue (migration 0004 adds the outbox the gated email.send tool writes) - tc-runtime: resumable LoopState checkpointed to agent_runs; gated tool -> approval row -> approval_required/run_suspended events -> suspend; resume consumes the grant BEFORE executing (spent grant = no execution), rejection feeds a structured refusal in-band; durable resume sweeper; continuous journal seq across suspension (tested). ContentPart::Text became a struct variant — internally-tagged newtype primitives don't serialize - tc-api: GET/decide approvals endpoints (409 double-decide, tenant isolation), decision triggers in-process resume; full chain proven over HTTP incl. gateway resumeFrom continuation - frontend: approval_required/run_suspended events, suspended reply state, inline ApprovalCard (§10: summary, category, exact payload preview, approve/reject -> decide + stream re-attach), /approvals queue page, nav - E2E (14 journeys, workers:1 to serialize the shared backend): gated email blocks with disabled composer -> approve -> continuation + ✓ step + reload replay; reject -> ✗ step, nothing executed; queue page decides pending 106 Rust + 61 frontend tests + 14 Playwright journeys green. Co-Authored-By: Claude Fable 5 <[email protected]>
40 lines
1.1 KiB
TOML
40 lines
1.1 KiB
TOML
# Scenario script for the deterministic `scripted` LLM provider.
|
|
# A prompt containing a `marker` plays that scenario's turns; each tool
|
|
# round-trip advances to the next turn. Unmatched prompts get an echo.
|
|
|
|
[[scenario]]
|
|
marker = "[[scenario:hello]]"
|
|
|
|
[[scenario.turns]]
|
|
events = [
|
|
{ type = "text", text = "Hello! I'm Scout, your research analyst." },
|
|
]
|
|
|
|
[[scenario]]
|
|
marker = "[[scenario:tool-time]]"
|
|
|
|
[[scenario.turns]]
|
|
events = [
|
|
{ type = "text", text = "Let me check the clock." },
|
|
{ type = "tool_use", name = "clock.now", input = {} },
|
|
]
|
|
|
|
[[scenario.turns]]
|
|
events = [
|
|
{ type = "text", text = "Done — I checked the current time for you." },
|
|
]
|
|
|
|
[[scenario]]
|
|
marker = "[[scenario:gated-email]]"
|
|
|
|
[[scenario.turns]]
|
|
events = [
|
|
{ type = "text", text = "I'll send that email once you approve it." },
|
|
{ type = "tool_use", name = "email.send", input = { to = "[email protected]", subject = "Q2 numbers", body = "Revenue is up 14% quarter over quarter." } },
|
|
]
|
|
|
|
[[scenario.turns]]
|
|
events = [
|
|
{ type = "text", text = " The email step is finished." },
|
|
]
|