Make Carl a working household clerk, accountant and advisor through reusable agent tools #33

New Issue

2026-10-02T12:12:34Z

KofTwentyTwo commented

2026-10-02 12:12:34 +00:00

(Migrated from github.com)

Carl must operate as the household clerk, accountant and advisor within one agent application. The owner expects normal conversation to inspect and summarize records, reconcile imports, correct local bookkeeping, maintain property/debt relationships, create and revise plans, and track execution. Reusable domain tools must let the model choose the necessary steps; a separate hard-coded workflow for each sentence is insufficient.

Actual failures and current state

The owner reports never receiving an LLM answer through normal chat. A monthly-spending question is retained as UNKNOWN, with no completed answer or financial artifact. Nine audited reads and nine provider-usage events exist; independent authorized replay confirms every read succeeds. Seven transaction page reads do not provide a deterministic whole-period category/merchant ranking.
The owner reports Check Status does not work. Diagnose native process/dock refresh versus the saved terminal outcome; a stale PENDING conversation label does not establish current backend state.
The running private local package remains the sealed 563-test application, with actual owner OIDC, actual Anthropic and retained Flyway V62/core V8. Eight real-data browser checks cover stated actions/dashboard/document/export views, not complete chat or every process.
A bounded readiness-summary repair passes 50 focused controlled SDK/PostgreSQL tests. A new full 566-test run against published foundation artifacts FAILED: one HTTP assertion failure and nine migration-connection errors in that suite. Diagnosis is in progress; no failed build is installed.
Exact grouped spending reads and native status/display repair are in progress. Real-model/browser acceptance is still required. No release or complete-product claim follows from controlled tests.

Required capability

Current permission-scoped search, entity reads, history, evidence and deterministic aggregates over ALL matching authorized records. Relative reporting periods use the household time zone and explicit resolved dates. Separate observed source movements, transfers, possible refunds and unqualified coverage.
Reusable typed local action tools to create/update/link/classify records and maintain plans through the same authoritative services as QQQ. Preserve verified caller identity, record/field/audience permissions, source immutability, attribution, revisions, idempotency, concurrency and recoverable outcomes.
Execute clear authorized reversible local bookkeeping changes; investigate available records first. Ask focused questions for ambiguity or missing evidence. Bulk/uncertain changes need an explicit review. Do not manufacture ownership, APRs, budgets, statement components, payment status, agreements or deadlines.
Rental-mortgage acceptance: use the owner's actual private example, identify the account and property, preserve the confirmed relationship, find candidate payments, and apply supported coding. Preserve unknown ownership/account identity. Mortgage principal, interest and escrow must not be collapsed into a claimed deductible expense. The real account label is deliberately omitted from this public issue.
External financial/vendor sends, payments, loans, purchases, transfers and commitments remain unavailable. Separately designated Synology authority is owner-deferred for local QA.
Source documents and prior model output are evidence, never identity or action authority. Private exports, account labels, financial values, credentials and raw model diagnostics stay outside GitHub/source/CI.

Acceptance

An actual verified owner sends an ordinary monthly-spending question in native/docked chat; actual Anthropic uses the aggregate capability and the answer renders with exact independently checked facts, accessible citations, dates and coverage.
Check Status reads that same saved request and updates the displayed state without another inference submission.
The actual rental-mortgage request is completed through general domain capabilities; QQQ shows the same records and attributable corrections. Unknown payment components remain visible.
Reusable local action governance is qualified through the foundation; no raw-executor, arbitrary SQL or impersonation bypass.
Plans and check-ins persist and remain revisable; reports/exports use current permissions and source provenance.
Full consumer gates, actual PostgreSQL negative/concurrency tests, packaged checks, scans and actual browser/model acceptance pass for the installed package. Retain original failed requests and evidence.
Inventory all tables/processes with actual acceptance or explicit input/permission/owner-deferred status. Do not claim all 150 processes work from metadata enumeration alone.

Related: #28 (real conversational knowledge/readiness), #29 (native actions), #30 (bulk corrections), #31 (documents), #3 (account identity), #5 and #6 (delivery). Financial priming also requires current debt terms, account/alias/bucket clarification, budget/reserve and current rental evidence. These gaps do not explain or excuse ordinary chat/status defects. Current work is local on feature/cicd-liquibase-gitops; integration and deployment states remain separate.

Shared dependencies: KofTwentyTwo/kof22-agent-foundation#60 (standalone governed local action authority) and KofTwentyTwo/kof22-agent-foundation#61 (dock status refresh). Native user-facing status acceptance: #34.

Carl must operate as the household clerk, accountant and advisor within one agent application. The owner expects normal conversation to inspect and summarize records, reconcile imports, correct local bookkeeping, maintain property/debt relationships, create and revise plans, and track execution. Reusable domain tools must let the model choose the necessary steps; a separate hard-coded workflow for each sentence is insufficient. ### Actual failures and current state - The owner reports never receiving an LLM answer through normal chat. A monthly-spending question is retained as UNKNOWN, with no completed answer or financial artifact. Nine audited reads and nine provider-usage events exist; independent authorized replay confirms every read succeeds. Seven transaction page reads do not provide a deterministic whole-period category/merchant ranking. - The owner reports Check Status does not work. Diagnose native process/dock refresh versus the saved terminal outcome; a stale PENDING conversation label does not establish current backend state. - The running private local package remains the sealed 563-test application, with actual owner OIDC, actual Anthropic and retained Flyway V62/core V8. Eight real-data browser checks cover stated actions/dashboard/document/export views, not complete chat or every process. - A bounded readiness-summary repair passes 50 focused controlled SDK/PostgreSQL tests. A new full 566-test run against published foundation artifacts FAILED: one HTTP assertion failure and nine migration-connection errors in that suite. Diagnosis is in progress; no failed build is installed. - Exact grouped spending reads and native status/display repair are in progress. Real-model/browser acceptance is still required. No release or complete-product claim follows from controlled tests. ### Required capability - Current permission-scoped search, entity reads, history, evidence and deterministic aggregates over ALL matching authorized records. Relative reporting periods use the household time zone and explicit resolved dates. Separate observed source movements, transfers, possible refunds and unqualified coverage. - Reusable typed local action tools to create/update/link/classify records and maintain plans through the same authoritative services as QQQ. Preserve verified caller identity, record/field/audience permissions, source immutability, attribution, revisions, idempotency, concurrency and recoverable outcomes. - Execute clear authorized reversible local bookkeeping changes; investigate available records first. Ask focused questions for ambiguity or missing evidence. Bulk/uncertain changes need an explicit review. Do not manufacture ownership, APRs, budgets, statement components, payment status, agreements or deadlines. - Rental-mortgage acceptance: use the owner's actual private example, identify the account and property, preserve the confirmed relationship, find candidate payments, and apply supported coding. Preserve unknown ownership/account identity. Mortgage principal, interest and escrow must not be collapsed into a claimed deductible expense. The real account label is deliberately omitted from this public issue. - External financial/vendor sends, payments, loans, purchases, transfers and commitments remain unavailable. Separately designated Synology authority is owner-deferred for local QA. - Source documents and prior model output are evidence, never identity or action authority. Private exports, account labels, financial values, credentials and raw model diagnostics stay outside GitHub/source/CI. ### Acceptance - [ ] An actual verified owner sends an ordinary monthly-spending question in native/docked chat; actual Anthropic uses the aggregate capability and the answer renders with exact independently checked facts, accessible citations, dates and coverage. - [ ] Check Status reads that same saved request and updates the displayed state without another inference submission. - [ ] The actual rental-mortgage request is completed through general domain capabilities; QQQ shows the same records and attributable corrections. Unknown payment components remain visible. - [ ] Reusable local action governance is qualified through the foundation; no raw-executor, arbitrary SQL or impersonation bypass. - [ ] Plans and check-ins persist and remain revisable; reports/exports use current permissions and source provenance. - [ ] Full consumer gates, actual PostgreSQL negative/concurrency tests, packaged checks, scans and actual browser/model acceptance pass for the installed package. Retain original failed requests and evidence. - [ ] Inventory all tables/processes with actual acceptance or explicit input/permission/owner-deferred status. Do not claim all 150 processes work from metadata enumeration alone. Related: #28 (real conversational knowledge/readiness), #29 (native actions), #30 (bulk corrections), #31 (documents), #3 (account identity), #5 and #6 (delivery). Financial priming also requires current debt terms, account/alias/bucket clarification, budget/reserve and current rental evidence. These gaps do not explain or excuse ordinary chat/status defects. Current work is local on feature/cicd-liquibase-gitops; integration and deployment states remain separate. Shared dependencies: KofTwentyTwo/kof22-agent-foundation#60 (standalone governed local action authority) and KofTwentyTwo/kof22-agent-foundation#61 (dock status refresh). Native user-facing status acceptance: #34.

KofTwentyTwo commented

2026-10-02 12:25:44 +00:00

(Migrated from github.com)

Owner-requested pause / independent LLM takeover — October 2, 2026

Implementation is stopped. Normal owner chat remains FAIL. Latest integrated consumer full build:566 tests,1 failure,9 errors,0 skips; the subsequent isolated HTTP conversation suite passes29/29. Both logs are preserved; isolated success does not erase the full-suite failure. Existing running563-test/FlywayV62 package and retained private records remain unchanged.

Bounded readiness retrieval is integrated but not fully qualified. The spending aggregate is an unintegrated four-file guarded candidate with50 focused tests passing. Shared Check Status candidate reproduces the UI defect and passes focused/frontend checks, but packaged GREEN, full foundation qualification and a identified revocation regression remain outstanding. Generic local-record authority is a saved design proposal only. No arbitrary writes, private data publication or external financial action was introduced.

A detailed docs/HANDOFF-NEXT-LLM.md and .continue-here.md exist in each local repository. Current source snapshots, hashes, diffs, isolated candidates and failed/passing evidence are saved in a durable protected local handoff. Background workers are stopped; no cleanup commit, push, merge, migration, release or deployment. This issue stays open for the next LLM. Related Carl#34, foundation#60/#61; existing Carl#28–31 details remain scoped history. Private source facts and credentials are not included here.

Owner-requested pause / independent LLM takeover — October 2, 2026 Implementation is stopped. Normal owner chat remains FAIL. Latest integrated consumer full build:566 tests,1 failure,9 errors,0 skips; the subsequent isolated HTTP conversation suite passes29/29. Both logs are preserved; isolated success does not erase the full-suite failure. Existing running563-test/FlywayV62 package and retained private records remain unchanged. Bounded readiness retrieval is integrated but not fully qualified. The spending aggregate is an unintegrated four-file guarded candidate with50 focused tests passing. Shared Check Status candidate reproduces the UI defect and passes focused/frontend checks, but packaged GREEN, full foundation qualification and a identified revocation regression remain outstanding. Generic local-record authority is a saved design proposal only. No arbitrary writes, private data publication or external financial action was introduced. A detailed docs/HANDOFF-NEXT-LLM.md and .continue-here.md exist in each local repository. Current source snapshots, hashes, diffs, isolated candidates and failed/passing evidence are saved in a durable protected local handoff. Background workers are stopped; no cleanup commit, push, merge, migration, release or deployment. This issue stays open for the next LLM. Related Carl#34, foundation#60/#61; existing Carl#28–31 details remain scoped history. Private source facts and credentials are not included here.

This repo is archived. You cannot comment on issues.

1 Participants

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: KofTwentyTwo/carl#33