Story: Run an independent local AI assistant

Provide the independent agent/MCP application described in issue #4295, backed only by NATS and a separately deployed CUDA LFM3 model.

• [X] Combined independent agent/MCP process with NATS-only backend access.
• [X] Shared authenticated read-only tools, standard external MCP client and bounded agent loop.
• [X] Official LFM3 QAD model with full CUDA offload on native Windows.
• [X] Independent checksummed packages, installation and lifecycle, tested on Windows/Linux.
• [X] Real currency/missing-observation answers; main backend remains running.
• [ ] Validate the native inference deployment on a Linux NVIDIA host.
• [ ] Validate restricted broker subject ACLs and a hosted client-enrollment profile before claiming production isolation.

• Scope correction, 2026-11-07: current work is agent and agent evaluation only. Do not change reporting/main-server/ORE source, schemas, business data, broker configuration or deployments. Retain existing deterministic client composition and read-only NATS access; normal login/logout is permitted. Proposed financial-rule migration is deferred for project-lead discussion and separate future PRs.
• Active Stage 4 is the evaluation runner/baseline, 5 agent-only behavioral improvements and 6 repeatable local evaluation/release. Most quality cases may fail; publish failures, blocked backend/data/identity cases and harness errors honestly. Deterministic safeguards and harness-integrity checks remain required; a complete report is not a claim of business fulfillment or production readiness.
• Future proposals F1–F10 are in projects/ores.agent/docs/architecture-future-prs.md, including query ownership, packaging/lifecycle, hosted security and commands. They do not block active agent/eval work and are not authorised here. The historical backend addition cc917e0d36 has since been reverted at the user's request; see the removal record below. Shared history is retained and no PR is opened. See the active projects/ores.agent/README.md#design-and-current-scope for the current sequence; earlier numbers remain historical.
• Business Stage 3 raw Greeks preserve cash finite-difference units at exact target shifts, factor/tenor/scenario/currency separation and per-bucket coverage. Missing/suppressed rows do not prove zero; only complete buckets receive a total. Signed ORE rate Delta at a 1bp target bump and cash volatility Delta are labelled explicitly, with no derivative normalization or sign reversal.
• Business Stage 3 NPV uses historical booking membership at the exact run-start timestamp, preserving microseconds. Portfolio totals follow the booking hierarchy, with complete population/coverage and explicit same-row base currency metadata; direct trade assignments are not mixed in.
• Measures whose units, bumps or factor/scenario contracts cannot be verified remain unavailable. The read-only release does not enable writes.

• Code 753a54b4550b9b73a602fc19a7b42b9393e773ad; separate release 0.1.5-cf24e13db443. npm test passes 77 tests on Windows and pinned Linux; package manifests match. Native Linux NVIDIA execution remains open.
• scripts/reconcile-greek-reference.py independently joins all 510 stored Delta/Gamma figures to 255 actual engine CSV rows and reconstructs historical scope membership. scripts/smoke-greeks.mjs reconciles every contributor through external MCP for 16 FX-book and 16 parent buckets; complete GBP 1Y/0.0001 FX-book Delta is GBP -6502.44 and Gamma GBP 0.14.
• Real GPU Delta/Gamma/DV01, run clarification, cross-party denial, errored-run withholding and missing-not-zero behavior pass. Positive live vega is not demonstrated by the non-option dataset; selector behavior has fixture coverage and missing data is explicit.
• The successful historical run 134b921d-eb6a-4277-964f-13d4b33429f5 completed clean. Earlier failed/errored attempts remain recorded. Fresh unscheduled prepared-input runs were set up through existing authenticated operator APIs; no agent write tool or native/broker deployment was added.
• https://github.com/alexhock/OreStudio/blob/f2dd9f89c88c45f270cf42cea43e99cc09cf5bc1/projects/ores.agent/docs/greek-validation.md records the existing normal-packager simulation/sensitivity gap and prepared-input lifecycle issue. These remain automatic-execution prerequisites; a version-checked operator start was used only on new validation runs, never to relabel terminal results.
• 19 inspected native/broker/inference PIDs and private agent config hash stayed unchanged across rollout. Main app and agent health/readiness returned HTTP 200. Story remains STARTED for subsequent business/security/Linux gates.

• Code 558452611601852436ccb3820efc1b10e6b8d7a0; installed independent release 0.1.4-754c262be033, identical Windows/Linux artifact. npm test passes 68 tests on each platform.
• scripts/smoke-scope-reporting.mjs reconciles every paged contributor against a separate Python Decimal reference: FX book 63 trades/GBP 206318954.123122; loans book 10/GBP 50969977.080832; parent 73/GBP 257288931.203954.
• Real GPU book/portfolio questions, run clarification, cross-party denial, partial-run withholding and unsupported Greek handling pass. No Greek total is claimed.
• Context corrected from 4096 to 262144 with an inference-only restart, full 25/25 CUDA offload and a successful 7225-token prompt. Agent output allowance defaults to 2048, configurable; truncated responses remain errors.
• During the subsequent agent release switch, 19 inspected native/broker/inference PIDs and the private agent config hash stayed unchanged. Main app and agent health/readiness returned HTTP 200. Only existing read subjects were added to the agent allowlist; no C++/broker changes or main build were required.
• Evidence and remaining gates: projects/ores.agent/docs/validation.md and projects/ores.agent/docs/agent.md#scope-totals-and-greeks. Story remains STARTED.

• Stage 1: health/configuration, checksummed package and separate installation; three Windows tests passed. Linux/Windows CI workflow added (manual dispatch); remote CI not yet run.
• Model: user approved official LFM3-1.7B QAD Q4_0, approximately 1.05 GB, replacing Qwen.
• Stage 2: ad58eefcf4, authenticated NATS/session/tool adapter.
• Stage 3: 6520bca6ed, official SDK MCP endpoint and independent client smoke.
• Stage 4: da3599e08b, pinned official CUDA runtime and approved LiquidAI model.
• Stage 5: a3cc72d40b, bounded direct tool loop and separate UI.
• Stage 6: 88fe926d56, immutable release and independent lifecycle controls.
• Validation: 12 tests pass on Windows/Linux; identical artifact 512b6069701e.
• Live release: 88fe926d56fa64653a70e4c009cab53eaa3cbda3; all 25/25 model layers offloaded on the RTX 4070. Main native backend/broker PID set unchanged across AI lifecycle checks.
• Remaining: native Linux NVIDIA validation, hosted auth/client enrollment and broker ACL hardening. Task/story remain STARTED; no unverified completion is claimed.

• Commit: 7cc8511292a2416da54c25fe8bd54530326b37c7; release 0.1.1-86a407144a11.
• npm test: 24 tests pass on Windows and Linux; same artifact identity on both.
• scripts/smoke-market.mjs: external MCP, provider clarification and actual CUDA-backed historical FX lookup pass; sampled IR series correctly returns no observation.
• Before/after native backend, NATS and inference PID sets unchanged. No server build, broker configuration, new subjects or data mutation.
• Freshness, exact point matching, selected-party restrictions and bounded complete scans are documented in projects/ores.agent/docs/agent.md#market-data.
• Remaining broader story acceptance items (Linux GPU and hosted isolation) stay open.

• Validated implementation commit: 8e83cc9f7b99df31e6a1f77929294fe2eb67dd73.
• npm test: 28 tests pass on Windows and Linux; local Linux evidence in build/output/agent-business-linux-check.log.
• scripts/inspect-business.mjs: successful real NATS reads for selected-party trades, books, portfolios, report definitions/instances/results and IR catalogue observations.
• Live sample at 2026-11-06 20:33 UTC: 73 trades, 2 books, 3 portfolios, 25 report definitions, 61 runs and 584 result rows. Four runs were clean, but none established the required matching clean trade-NPV example. All 32 selected-party IR series were empty.
• Inspection delivered; clean-NPV and populated-IR data readiness remain open. Discovery can proceed; no reporting result acceptance is claimed from fixtures.
• Existing stopped backend services were restored for inspection. No server rebuild, broker configuration change, database recreation, report trigger or feed import was performed by the probe.
• Reproduction, concrete record IDs and next actions: https://github.com/alexhock/OreStudio/blob/f2dd9f89c88c45f270cf42cea43e99cc09cf5bc1/projects/ores.agent/docs/business-baseline.md. Scope and later stages: projects/ores.agent/README.md#design-and-current-scope.

• Implementation: 065e5dc857; final exact-identity guard 26217842d462e4f5b193b55d4e02ce366adb728d.
• npm test: 37 passing Windows/Linux tests. npm run package: identical artifact 0.1.2-adeb734cfeb9 on both systems.
• scripts/smoke-discovery.mjs: independent MCP, real book membership, FX-forward economics, portfolio hierarchy, counterparty code and CUDA chat/clarification pass. A separate selected-party login cannot read the first party's known trade.
• Native backend, broker and inference PID sets unchanged across standalone agent deployment. Existing market-data/currency regressions pass; no server/broker configuration changes or business writes were introduced.
• Scope limits: tenant-wide counterparties have explicit paged name searches; Bond economics expose reference/notional only. No valuation/reporting capability is claimed by this release. See projects/ores.agent/docs/agent.md#discovery-and-trade-values.
• Next: Stage 2 stored trade valuation results, with the clean-NPV data gate from Stage 0 still open.

• Implementation 5ff7160a7b; final unit-boundary correction bf08fba92c6cfe2da67ec172f6c207272501b4ff.
• npm test: 48 passing Windows/Linux tests. npm run package: identical 0.1.3-5b23d543cb79 artifact.
• scripts/smoke-reporting.mjs: actual independent MCP, report provenance, default withholding, explicit partial inspection and second-party isolation pass. GPU model selects NPV/status reads; exact explicit partial commands preserve IDs via validated direct execution.
• Discovery and market regression scripts pass. Native backend, NATS and inference PIDs remain unchanged; main web app HTTP 200. No report execution, server build, broker change or business data import was introduced.
• Clean live valuation acceptance remains open. Existing applicable runs completed with errors; fixture success is not substituted for that gate. Diagnose/prepare one approved clean run separately before claiming full Stage 2 business acceptance or reconciled aggregates.
• Reproduction and limits: projects/ores.agent/docs/agent.md#discovery-and-trade-values; evidence summary in projects/ores.agent/docs/validation.md.

• Passed on 2026-11-07 using installed agent 0.1.3-5b23d543cb79; no new agent deployment.
• Isolated definition 39eb7b57-db88-4669-ae7e-fbe3dbf40016; clean run 884d0bc9-932d-4c4c-9b66-ee97d2b810bb priced at 2026-11-06 and completed 2026-11-06 21:43:39.214032 UTC.
• Example trade 05dcc4f0-4af9-597f-8e20-94c3018c6b4c: NPV GBP -7,998,696.427339, source npv.csv. Direct authenticated NATS, independent MCP and GPU chat reconcile; second-party access is refused. Existing clean/partial/status reporting smoke passes.
• Original regulatory output failed on Bond SA-CCR/netting data. A first isolated attempt failed on absent scope and remains failed. The corrected report gathered 73 trades from two books.
• Declared backend prerequisite: reporting.v1.ops.copy_run_scope, ownership/permission gated, atomic, empty/unshared destination only. Five reporting compilation units and its DLL rebuilt with two jobs. A temporary instance verified unauthenticated/wrong-party/same-configuration/populated-shared/replay refusals, copied the scope and was stopped before the corrected run.
• Main fleet not restarted/replaced; no broker ACL/configuration changes or queue purge. Original fleet does not permanently serve the new subject; native Linux validation of this backend addition remains open. Agent remains read-only.
• The existing 48 Windows agent tests passed; the unchanged artifact had already passed 48 tests/packaging on Linux. No native Linux backend or least-privilege/cross-tenant proof is claimed.
• Evidence, reproduction, warnings and backend API observations: https://github.com/alexhock/OreStudio/blob/f2dd9f89c88c45f270cf42cea43e99cc09cf5bc1/projects/ores.agent/docs/valuation-validation.md. Next: Stage 3 scoped summaries and available Greeks, with explicit completeness and units.

• Implementation ebb7fe7849d2089da0444f250bbe32edddfbf405; corpus 1.0.0 contains 26 cases. All 85 component/harness tests pass on Windows and pinned Linux. Linux replay reproduces Windows outcomes with matching source fingerprints; its archive report records unknown Git revision.
• Replay handling passes 26/26, model/fixture handling 14/21 and live-read handling 13/18. Overall passes are 22/26, 12/21 and 11/18. Blocked cases stay in denominators and business fulfillment is separate. One repetition is baseline evidence, not a broad accuracy benchmark.
• Failed automated checks remain recorded. Booking refusal has a known lexical-oracle false negative; preserve the frozen result and review/version any correction during Stage 5. Missing IR/vega and report/booking capabilities remain explicit.
• Usage, safe aggregate and private evidence paths: projects/ores.agent/docs/evaluation.md and https://github.com/alexhock/OreStudio/blob/f2dd9f89c88c45f270cf42cea43e99cc09cf5bc1/projects/ores.agent/docs/evaluation-baseline.md. Normal completed reports return 0 despite quality failures; strict replay returns the expected 2.
• No production agent/backend source or business records changed. Existing read-only NATS plus normal session lifecycle only; installed 0.1.5, 19 native/broker/inference PIDs, agent PID and private config remain unchanged. Main app and agent health/readiness return HTTP 200.
• Next: agent/eval-only Stage 5; backend proposals remain separate future PRs. Task/story remain STARTED for later stages.

• Final implementation 17cb7bc4a07cb9108ab81fe7abf69d8afb7303be; separate release 0.1.6-4317a6cd1870. All 94 tests pass on Windows/Linux with identical manifests and replay source fingerprints. Native Linux NVIDIA validation remains open.
• Added typed bounded intent/catalogue selection, preservation of literal dates/factors/scenarios, deterministic command refusal, identity-bound expiring clarification and explicit withholding on unit/provenance validation failures. No financial semantics or backend migration was introduced.
• Original Stage 4 corpus/grader/model settings are unchanged. Two-repetition handling scores: 42/42 model/fixture, 33/36 live, versus Stage 4 14/21 and 13/18 with one repetition. Eight separate paraphrase/clarification cases retain their own denominator; small samples are not broad benchmarks. Full failures and blocked data/backend capabilities remain recorded.
• Earlier live findings at 0ca93f8aa3 remain preserved, including run-status substitution and a generation-limit response. The former prompted the final scope catalogue correction; generation allowance and unit checks were not relaxed. Existing trade base-NPV currency-unit validation still withholds the positive result.
• Installed authenticated chat smoke checks pass; 19 native/broker/inference PIDs and private config hash remain unchanged. Main app and agent health/readiness return HTTP 200. Only the standalone agent lifecycle changed; no new reporting/server/ORE, schema/business-record or broker changes.
• Evidence: https://github.com/alexhock/OreStudio/blob/f2dd9f89c88c45f270cf42cea43e99cc09cf5bc1/projects/ores.agent/docs/stage5-validation.md and reviewed evals/baselines/stage5-2026-11-07.json. Next: agent/eval-only Stage 6; backend architecture stays separate future PRs. Task/story remain STARTED.

• Implementation 5004b5e7c8e36f7c26e70788227ca92fdfa1675f; installed independent release 0.1.7-0a1d2f580ce4. All 101 tests and both replay handling gates pass on Windows/Linux with identical fingerprints/manifests. Remote CI is configured, not dispatched; native Linux NVIDIA and local Node 22 validation are not claimed.
• Added npm run verify, eval:doctor and eval:compare commands, strict/non-overwriting comparisons, explicit provenance checks and argument/tool/clarification/refusal/numeric-attribution/latency sample metrics. Private chat spans correlate HTTP request IDs across inference, intent/catalogue selection, tools, adapter reads and formatter attempts without payloads/identities/error text. No financial semantics changed.
• Frozen cases/grader/model settings are unchanged. Two-repetition handling is 42/42 model/fixtures and 32/36 live; separate paraphrase profiles pass 8/8 each. Strict live comparison returns 2 for scope-run-clarification (prior 1/2, now 0/2); native-NPV mixed units and existing base-NPV unit errors remain withheld. No oracle weakening or best-of reruns hide the finding.
• Standalone agent rollout and five installed checks pass, including 62 request-correlated timing events across all six phases. Main app and agent health/readiness return HTTP 200. All 19 native/broker/inference PIDs and private config hash remain unchanged; previous 0.1.6 remains for rollback. No new reporting/main-server/ORE, schema/business-record or broker changes.
• Evidence: https://github.com/alexhock/OreStudio/blob/f2dd9f89c88c45f270cf42cea43e99cc09cf5bc1/projects/ores.agent/docs/stage6-validation.md, projects/ores.agent/docs/evaluation.md#compare-and-measure and reviewed evals/baselines/stage6-2026-11-07.json. Active local Stages 0-6 are delivered with explicit gaps; broader feature/story remain STARTED for unimplemented business/hosted/Linux-GPU gates. Review future scope with the project lead; this does not authorise backend PRs or claim production readiness.

• At the user's explicit request, reverse cc917e0d36352a5265e862c66e358f5593c2cef7, including its reporting protocol/handler/repository/modeling/generated changes and operator-only copy-validation-scope.mjs helper. Reporting source matches the branch base; the original commit and reversal remain auditable without a force-push. No PR is opened.
• Agent/MCP/evaluation runtime does not call that operation. Existing stored NPV/Greek data and installed agent 0.1.7/native/NATS/inference services remain untouched. Fixture profiles need no live data; reproducing the historical live dataset on a new environment requires separately agreed preparation, not the removed API.
• Keep the negative evaluation mutation-blocking test. Historical setup evidence remains labelled historical, and future backend work stays outside the current agent-only scope.
• Validation: npm run verify passes all 101 tests and both offline replay handling gates; all 40 component documentation checks pass. The complete net branch diff is limited to agent files, its dedicated workflow and agile documentation. Reporting source is byte-for-byte identical to the branch base. No native rebuild or application redeployment is required.

• Final code 055af005e05aa5bdc59c3993609a4937baa3bf85; structural checkpoint 80b04ac08b; installed 0.1.8-c1371bb8918e. All 107 tests and three replay handling gates pass on Windows/Linux with identical source fingerprints/manifests. Native Linux NVIDIA/remote CI remain open.
• Fixed explicit trade/run substitution, negated/absent partial consent, provider phrase corruption and false native/base evaluation passes. Added a typed RequestPlan, one identity-bound pending-choice store, deterministic scope continuation, single presentation dispatch, cached catalogues and shared bounded pagination. Existing financial/identity/coverage checks remain in adapters; evaluation options/reporting are separated and readable.
• Frozen corpus/grader and legacy fixture bytes are unchanged. New review-2.0.0 has 11 immutable-fixture cases with explicit measure/value/currency/basis assertions. Actual fixture version/hash is recorded and compared. Two-repetition handling is 42/42 original model, 34/36 live and 22/22 review-model; paraphrases 8/8 each. The former continuation regression now passes 2/2. Existing live base-NPV unit and IR/vega/command gaps remain explicit, with no data/server repair.
• Seven installed chat/timing checks pass, including typed run continuation, provider wording and factor punctuation. Main/agent health/readiness return HTTP 200. All 19 native/broker/inference PIDs and private config hash are unchanged; prior 0.1.7 remains for rollback. The removed reporting addition stays reverted and no backend/broker/schema/business records change.
• Evidence: https://github.com/alexhock/OreStudio/blob/f2dd9f89c88c45f270cf42cea43e99cc09cf5bc1/projects/ores.agent/docs/review-validation.md, projects/ores.agent/docs/agent.md#api-mcp-and-request-handling and reviewed evals/baselines/review-2026-11-08.json. Broader feature gates remain STARTED; future architecture/server work is separately scoped with the project lead.