Runtime agent, run-1 transcript (2026-06-05)

Raw working document, published unedited as evidence. Written for the run, not for reading flow; the readable account is on the experiment pages.

Provenance note (added on commit): this is the consumer-side run cited on the Results page, the $2,288.00 + $1,099.00 = $3,387.00 NZD leg. It reconciled two invoices because it confirmed the seeded procurement (PR-1, $2,288.00) in addition to the order it placed itself ($1,099.00); later runs confirm only their own order, which is why other committed transcripts show a single $1,099.00. Evaluator caveat below is original and stands.

Evaluator: Claude (in a human operator’s session), via the deployed remote MCP. Caveat: the evaluator co-wrote the build spec, so this run is not a pure fresh-agent test. It validates functionality, error legibility, and the full loop; the official unbiased transcript comes from the clean-chat re-run after fixes. Verdict: PASS (findings below fed back for fixes).

Status note (editorial, added on publish): that clean-chat, fresh-agent re-run through the MCP itself is still open. An independent agent has since re-driven the full lifecycle through the HTML-only run’s built app (4/4 HMAC-verified), which confirms the API contract and lifecycle, but not the MCP’s tool-description legibility this run tested. See the Results page caveats.

Run log (condensed)

1. Orient + pre-warm (parallel): prepare_vehicle(vin JTDBR32E730012345) → accepted. get_job(CCC-2026-04817) → job J-4817, 2019 Corolla (plate MJL472 NZ), insurer Aotearoa Mutual, 8 parts: 3 ordered (bumper, bonnet, LH headlamp, under seeded procurement 10000000-…-0001, order_requested, Christchurch Toyota), 5 estimated. find_repairer_site → Canterbury Collision Centre, Riccarton. list_work_providers → Aotearoa Mutual.
2. Basket: recommend_basket → 12 offers, 3 suppliers (Christchurch Toyota OEM; Repco + PartsMaster aftermarket), per-part OEM-vs-aftermarket competition, by-supplier subtotals.
3. Induced error (deliberate): place_procurement mixing Toyota RH headlamp + PartsMaster grille → HTTP 422 refused; message named the actual violation (“all offer_selections must be sourced from a single supplier”) + recovery path (re-fetch basket, retry). Self-corrected without help.
4. Place: place_procurement, Toyota only: RH headlamp ($631) + radiator support ($468) → 35a8d1eb-…, order_requested, NZD $1,099.
5. Supplier confirms (both): confirm_procurement on 10000000-…-0001 and 35a8d1eb-… → both order_confirmed (terminal state of the faithful public lifecycle); deliver_before set.
6. Reconcile: reconcile_invoice(job) → 2 invoices, both reconciled, every line matched, $2,288.00 + $1,099.00 = NZD $3,387.00. is_extension: true correctly flagged on extension surfaces.

Pass criteria met

• Full lifecycle completed with no human help, oriented from tool descriptions alone (NEXT-chaining + seed IDs in descriptions made the path plannable).
• Induced error → message-driven self-correction.
• Final state correct; narration coherent.

Findings (fed back 2026-06-05; fixed before later runs)

1. Seed coverage: only Christchurch Toyota’s credential can confirm, but Toyota lacked offers for 3 of 8 parts (grille, A/C condenser, bonnet strut) → the whole job could never reach reconciled. Resolution: the constraint is now surfaced legibly in the tools.
2. 422 ordering: generic cause-list preceded the actual violation; now leads with the specific cause.
3. recommend_basket returned currency: null, resolved the faithful way: currency is a procurement-level concept and the basket carries none (the documented no-currency-on-basket rule).

What’s deliberately good (keep)

NEXT: chaining; intent-level tools; append-vs-replace merge semantics on identify_parts (with the read-merge-rewrite explanation of the replace-the-list API); honest extension labels + is_extension: true; one-credential-drives-both-sides demo design; LLM-recoverable error envelope.