Connect ordinary chat to real household records and generate source-linked readiness plans #28

New Issue

2026-10-02T00:45:05Z

KofTwentyTwo commented

2026-10-02 00:45:05 +00:00

(Migrated from github.com)

Ordinary native chat cannot retrieve authoritative household records: CarlConversation.infer supplies no tools and the router only accepts closed workflow proposals. Natural questions consequently return clarification despite a connected model and retained real imports.

Provide caller-aware, bounded read tools for inventory, accounts, transaction history, source evidence, plans, and protected reports; use the foundation ToolGate and current conversation audience intersection. Preserve the explicit write workflows, finite budgets and excluded external financial actions. Verify actual native chat against the authorized private dataset without publishing private inputs or outputs.

The authorized private readiness audit also finds missing debt terms, budgets, properties, expenses, priorities and plans. Save confirmed owner priorities and generate a source-linked financial readiness plan, retaining missing facts. Never fabricate rates, payments, property facts, deadlines, offers or a payoff outcome to fill tables. Collect missing facts through usable native actions and qualify substantive plans after those facts arrive.

Acceptance: ordinary questions and follow-ups retrieve current permitted real records; imports remain faithful; private/shared/revoked access is enforced; a generated readiness plan is visible through the same QQQ state and protected exports. Full application gates must pass and the running package must match the verified source.

Ordinary native chat cannot retrieve authoritative household records: `CarlConversation.infer` supplies no tools and the router only accepts closed workflow proposals. Natural questions consequently return clarification despite a connected model and retained real imports. Provide caller-aware, bounded read tools for inventory, accounts, transaction history, source evidence, plans, and protected reports; use the foundation ToolGate and current conversation audience intersection. Preserve the explicit write workflows, finite budgets and excluded external financial actions. Verify actual native chat against the authorized private dataset without publishing private inputs or outputs. The authorized private readiness audit also finds missing debt terms, budgets, properties, expenses, priorities and plans. Save confirmed owner priorities and generate a source-linked financial readiness plan, retaining missing facts. Never fabricate rates, payments, property facts, deadlines, offers or a payoff outcome to fill tables. Collect missing facts through usable native actions and qualify substantive plans after those facts arrive. Acceptance: ordinary questions and follow-ups retrieve current permitted real records; imports remain faithful; private/shared/revoked access is enforced; a generated readiness plan is visible through the same QQQ state and protected exports. Full application gates must pass and the running package must match the verified source.

KofTwentyTwo commented

2026-10-02 12:15:23 +00:00

(Migrated from github.com)

October 2 owner QA: FAIL for ordinary native chat. The owner reports never receiving an LLM reply, and a monthly-spending request is retained as UNKNOWN. Actual Anthropic usage and nine audited reads exist; all nine independently replay successfully with current authorization. Seven raw transaction pages lack the whole-period ranking needed by the question. Connected inference and selected successful backend cases do not qualify the owner’s UI workflow.

Bounded readiness evidence has a meaningful controlled SDK/PG regression: ten-iteration exhaustion becomes two reads/three turns with unchanged limits. Fifty focused tests pass. A fresh full 566-test run against published foundation artifacts failed one HTTP assertion plus nine migration-connection checks; diagnosis is running, and the installed package remains563. A deterministic grouped spending read is being qualified independently, with exact currency/classification/transfer/refund/unknown coverage and resolved household-zone periods. Native Check Status/display investigation is #34 and shared foundation#61. No failed request is automatically replayed.

The broader required product is now #33: reusable clerk/accountant/advisor capabilities and verified local actions, with shared authority foundation#60. Completion requires actual owner native/dock send→real model→audited tools→visible answer, independently checked facts/citations and matching full-gate package. Private records and diagnostics remain outside this tracker.

October 2 owner QA: **FAIL for ordinary native chat**. The owner reports never receiving an LLM reply, and a monthly-spending request is retained as UNKNOWN. Actual Anthropic usage and nine audited reads exist; all nine independently replay successfully with current authorization. Seven raw transaction pages lack the whole-period ranking needed by the question. Connected inference and selected successful backend cases do not qualify the owner’s UI workflow. Bounded readiness evidence has a meaningful controlled SDK/PG regression: ten-iteration exhaustion becomes two reads/three turns with unchanged limits. Fifty focused tests pass. A fresh full 566-test run against published foundation artifacts failed one HTTP assertion plus nine migration-connection checks; diagnosis is running, and the installed package remains563. A deterministic grouped spending read is being qualified independently, with exact currency/classification/transfer/refund/unknown coverage and resolved household-zone periods. Native Check Status/display investigation is #34 and shared foundation#61. No failed request is automatically replayed. The broader required product is now #33: reusable clerk/accountant/advisor capabilities and verified local actions, with shared authority foundation#60. Completion requires actual owner native/dock send→real model→audited tools→visible answer, independently checked facts/citations and matching full-gate package. Private records and diagnostics remain outside this tracker.

This repo is archived. You cannot comment on issues.

1 Participants

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: KofTwentyTwo/carl#28