Golden itinerary proof — regularizar-atrasos through the installed plugin¶
Verification record for the claude-ecosystem-packaging campaign (plan step W05.P13.S46), 2026-07-03.
Two-layer evidence¶
Scenario logic (R7 harness, pre-plugin): the
2026-07-03-agent-harness-operability-followup-auditlive-measured theregularizar-atrasospath PASS with a real model persona operating the console end-to-end against golden scenarios (real server, real MCP session, trajectory scored on observed calls).Through the INSTALLED PLUGIN (this proof): a live headless session of the Claude desktop app’s embedded runtime, restricted to the installed plugin’s MCP tools (
mcp__plugin_aeat_aeat__*), was given the late-filer ask (“Estoy atrasado con el IVA trimestral — ¿qué me he perdido?”) against a fresh isolated storage root. Observed, in order: the harness floor loaded first (operating rules + R9 disclosure); the console surface was discovered through the contract/search tools; the orientation verb ran via the execute meta-tool; theoverview backlogcall hit the honest CLI boundary (REFUSED_CLI_BOUNDARY— no active profile in the empty workspace); and the assistant reported the refusal HONESTLY with the exact instructive next step (create the profile, then re-runaeat app overview backlog). No live-submission attempt at any point.
The empty-profile refusal is a designed, correct itinerary outcome: the operator-facing gate is instructive, the assistant relays it faithfully, and the next-step handoff is the CLI’s own suggestion — the honest-declaration and safety-handoff operating rules working through the delivered plugin, live.
Residual¶
Re-run this itinerary against a SEEDED profile through the plugin after
the first PyPI publish (one command swap in .mcp.json back to the
published pin) — tracked with the operator-gated follow-up. The scenario
logic itself is already PASS-measured under the R7 harness.