Curated-agent correction: observations

TDD and mutation observations

• Disabled plus explicit model initially returned a resolved plan without persona/tool policy; the new planner case now returns unknown_target.
• A configured Explore process override initially reached the RPC path; the engine now pins curated names after overlay and keeps custom-agent process mode unchanged.
• Senpi's ordinary bash created a probe file in the rejection reproduction. Curated sessions now receive the schema-constrained broker; unit and live mutation attempts are rejected.
• The original mtime checker failed all new content-freshness cases. The final five build tests cover mtime independence, body tamper, stale source marker, whitespace normalization, and transitive senpi-task inclusion.
• Mutating the mirrored fallback table, record parser, or root export made its corresponding regression test fail before restoration.
• The disabled same-name category case was RED at 10 pass / 1 fail and GREEN at 11 pass / 0 fail.

Final gate observations

• Both scoped typechecks exited 0.
• bun test packages/senpi-task packages/omo-senpi: 1,014 pass, 0 fail, 2 snapshots, 3,422 expectations.
• bun run test:senpi: 259 pass, 0 fail, 1 snapshot, 813 expectations.
• Markdown link audit: 16 pass, 0 fail.
• Build verifier tests: 5 pass, 0 fail.
• Final build-extension.mjs --check: exit 0; the generated bundle also passes the 700,000-byte package budget.

Final live observations

• Curated driver: 18 checks PASS, including bash_read_succeeded, bash_mutation_rejected, unchanged probe bytes, absent direct/bash forbidden files, and no QA-attributed real Senpi path.
• Baseline driver: 12 checks PASS with zero leaked PIDs and no QA-attributed real Senpi path.
• Concurrent host sessions changed during both snapshots. Their paths are retained in the JSON artifacts and classified separately; the whole-directory digest is therefore honestly recorded as changed.