AutoForge/prompts/Step-12-Completion.md

16 KiB

Step-12 completion - Qualify the installed local model in the native application

STEP IDENTIFIER

Step-12, completion of its original outstanding requirements. This is NOT Step-12A, a new milestone, or acceptance of Step-12. Read the frozen Step-12 assignment, original blocked report, STATE.md and WORKER-QUALITY-GATE.md.

This handoff supplies the current model selection and an executable sequence for the remaining prerequisites/evidence. It does not replace C-01..C-10 or AC-01..AC-05. Preserve the original and existing completion reports unchanged; write fresh resumed evidence to prompts/Step-12-Completion-Followup-Report.md, then STOP for manual orchestrator review.

STEP TITLE

Verify cloud-disabled Ollama, qualify the selected installed 8B model through the real MAUI application, and complete the original Step-12 evidence.

OBJECTIVE

Resolve the outstanding AC-01/AC-02/AC-04 prerequisite and real-model/native checks, preserving the implementation's already supported controlled contracts. Use actual installed model output and the authoritative application validator, not a canned response, hand-edited output, direct validator-only substitute or an explanation of what a future worker should do.

The worker executes the task. The orchestrator reviews returned evidence and maintains orchestration documents; it does not install/configure/test for you. No automatic agent handoff or Step-13 execution.

ARCHITECTURAL CONTEXT

Keep the existing nine-project .NET 10 solution and sole Windows MAUI Blazor Hybrid Client. Reuse the actual Application inference/context/generation owner, shared deterministic runbook validator and UI /runbooks workflow. Core operations remain available without inference.

Generation is simulation calibration preview only: no authoring proposal Save, target open/approval/write, runbook execution, dynamics application or physical operation. Tools/thinking advertised by the model do not grant application tool access or change the assignment.

CURRENT IMPLEMENTATION STATE

  • Step-11 is accepted; Step-12 is BLOCKED/unaccepted.
  • Step-12 code and controlled HTTP/UI/lifetime/isolation tests are present. Original worker evidence: clean nine-project Release build, 179 unit + 415 integration = 594 passes, zero failed/skipped. Preserve all 594 cases, including the original 547-case baseline.
  • The original report observed no installed models and cloud-disabled:false. These are historical observations, not a reason to repeat model installation.
  • Subsequent setup installed exactly huihui_ai/qwen3-abliterated:8b-v2. Last verified metadata: local GGUF, qwen3, 8.2B, Q4_K_M, approximately 5 GB, completion/tools/thinking capabilities and Apache License metadata.
  • Last verified catalogue digest: 543f9deff86d8347fd96ea1f3e54e73861ea70d52def71d31ccaeaf013de6ba6. Persistent model/config/template/license/parameters/original-manifest blobs passed size and SHA-256 checks.
  • Last observed daemon store: D:\AI\ollama\models. The shell's OLLAMA_MODELS candidate differed; do not relocate or redownload.
  • Last runtime version: 0.40.13 at http://127.0.0.1:11434. Its latest inspected startup log still said cloud-disabled:false. No actual model generation or native qualification has been performed.
  • The completion report confirms native installed- model selection/context and 594 regressions pass, but specific setup/restart/ quiescence approval was unavailable. No chat or runtime change occurred.

Owner directive, 2026-10-31 12:00 -04:00: local models only; no Ollama cloud. This is mandatory, not an optional backend choice or permission to use cloud while waiting. Do not ask whether cloud inference is acceptable or substitute a cloud model. Specific shared-configuration/restart approval and quiescence remain separate; an unavailable approval response is not affirmative consent.

Recheck current state; these observations are not permanent attestations. Read local generation contracts and build/native guidance.

REQUIRED IMPLEMENTATION

1. Verify prerequisites and obtain any required configuration approval

Use bounded, read-only runtime/version/catalogue/details probes and targeted runtime configuration/startup-log inspection. Verify the exact selected model, digest, positive installed size, local GGUF/architecture/completion eligibility and absence of remote/cloud delegation indicators. Inspect relevant license and actual RAM/GPU/VRAM/runtime resource context. Do not dump logs/environment variables containing unrelated private data.

Cloud-disabled status must be verified for the current running server, not a stale log from another process or a checkbox. Use the installed version's documented local-only setting, such as disable_ollama_cloud:true in its supported server configuration or OLLAMA_NO_CLOUD=1, and fresh startup proof. Recheck the official FAQ against the installed runtime.

If cloud is already disabled, record the evidence and continue. If it is not:

  1. Discover the actual runtime owner/configuration and prepare the exact minimal cloud-disable change and deliberate restart procedure.
  2. Obtain explicit operator approval for that specific shared-runtime setting change/restart before executing it. The earlier model-download approval and this prompt do not themselves authorize interrupting shared work.
  3. With approval, preserve unrelated settings and model-store paths, retain the specific prior configuration outside the repository if needed, and perform only the approved change when the operator confirms the runtime is quiescent. Prefer its normal shutdown/start mechanism; never reuse historical PIDs or kill processes by name. No machine-wide environment changes or broad reset.
  4. Verify the runtime is responsive at the assigned endpoint and fresh startup evidence says Ollama cloud disabled: true; recheck exact installed model metadata before sending any goal/source context.

If approval, access or proof is unavailable, report the precise blocker and prepared operator action. Do not merely copy the old report, check the UI acknowledgement anyway, use cloud, or claim a successful native check. Do not install/update/pull another model/runtime or switch to a different endpoint/model. A changed selected digest requires explicit reselection/ operator review, not a silent fallback.

2. Exercise actual local inference through the native application

Build/start the existing real Client using documented guidance, with task-owned native instrumentation only if needed. Verify readiness; do not substitute a web host or controlled HTTP peer for this native test.

Create/select the actual warm synthetic authoring baseline explicitly through the supported workflow: warm-idle 650 rpm, fixed torque axis including 2000 rpm with its baseline cell 360 N*m. Record actual workspace/head/applicability and whole before values. Do not create a physical target or program a rehearsal.

On /runbooks:

  1. Verify entry/render causes no inference request or automatic selection.
  2. Explicitly refresh the installed catalogue and select exactly huihui_ai/qwen3-abliterated:8b-v2; verify displayed model/digest/eligibility.
  3. Inspect actual source/head/knowledge/canonical context. Only after the running server's cloud-disabled proof, acknowledge local-only generation explicitly.
  4. Submit the original decisive goal exactly: Propose warm-idle target 1250 rpm and set torque at 2000 rpm to 350 N*m; keep every other value unchanged.
  5. Capture the actual original assistant message.content, request/model/digest/ context provenance and visible application result. It must reach the same strict validator unchanged and produce an actual current read-only preview.
  6. Verify exactly 1250 rpm and the 2000-rpm cell 350 Nm: scalar delta +600 rpm, torque delta -10 Nm, exactly those two changes and every other value unchanged. Parseable in-range JSON with different intent is not a passing result.
  7. Verify authoring head/history/configuration, source values, target approvals/ writes, dynamics and recording/replay data are unchanged by generation.

Report every attempt, latency and failure honestly. An explicit retry is a fresh single request against captured current context, not an automatic repair loop. Observe the existing 180-second operation deadline, 4096-token generation budget, request/response bounds and strict runbook limits; do not loosen them to force acceptance. Reasoning text is not output/evidence for validation and must not be displayed as a plan. If model/runtime thinking handling or another actual integration issue prevents valid original JSON, diagnose it with real evidence.

No planned code changes are required. Narrow compliant fixes needed to fulfill the original Step-12 contract may be made within this still-open step, with focused regression tests and source-path self-review. Do not replace the model with a rule-based generator, remove thinking/Markdown from returned content, inject IDs, convert units, clip values, bypass validation or weaken guards. If a material contract change or unresolved defect is needed, report it precisely for orchestrator decision; do not invent Step-12A yourself.

3. Finish native lifecycle, failure and coexistence evidence

Exercise a real in-flight selected-model request, not just no-model Idle:

  • Cancel and drain while actual generation is pending; observe the pending state before cancellation and verify no late preview replacement.
  • Navigate/unmount while actual generation is pending; verify cleanup drains, no late UI/preview publication and fresh entry has no old owner/selection/ acknowledgement. Do not stop/unload the shared model server as cleanup.
  • Repeated Generate/refresh/selection during the pending owner must not dispatch another operation. Preserve current source and prior successful preview.
  • Verify ordinary manual invalid input retains prior success, changed authoring source marks old preview Stale, unit changes project display only, and old routes remain usable.
  • Establish zero unassigned business-source effects with the existing real feature-boundary counters/state comparisons and native before/after evidence. Preparatory explicit baseline creation or deliberate staleness tests are separate actions, not mutations attributed to inference.

Timing evidence must actually show a pending real request; sleeps or a finished request followed by Cancel do not prove in-flight behavior. Existing controlled barriers remain complementary tests, not native-model substitutes. Do not inject a fake model reply into the native success path.

FILES OR PROJECTS EXPECTED TO CHANGE

Expected repository changes: directly related README/docs verification status and only prompts/Step-12-Completion-Followup-Report.md under prompts. Preserve Step-12.md, this handoff, the original report, shared gate, roadmap/state, the existing completion report and all older assignments/reports. Application/UI/Client/tests change only for an evidenced narrow original- contract integration fix; no new project/package/host or unrelated cleanup.

INTERFACE AND DEPENDENCY REQUIREMENTS

All original Step-12 schema/source/head/preview/model/owner/local-only/bounds/ approval restrictions remain authoritative. Use existing DI/inference port/ validator rather than a parallel native-only path. No application tool calls, cloud fallback, output repair, unrestricted network/file access, programming, Git backup persistence or direct access to a private simulation oracle.

Any operator-approved runtime configuration is external setup, not permission to add automatic runtime management/pulls/settings to the application. No staging/commit/push, agents, automatic handoff or protected-file edits.

TESTING REQUIREMENTS

Use existing focused generation/preview/HTTP/owner/UI/isolation tests for any affected code, then execute the original required commands sequentially:

  • dotnet build .\src\AutoForge.slnx --configuration Release --no-restore
  • dotnet test .\src\AutoForge.slnx --configuration Release --no-build

Report exact exits/warnings/per-project case counts; keep all 594 current regressions and the nine-project native build. Dependency restore is permitted only after a genuine chosen-command missing dependency; no SDK/workload/model installation/update/repair. Separately execute and report the real-model/native checks above. Build/test success alone does not satisfy AC-01/AC-02/AC-04.

Release controlled barriers in finally. Clean only exact task-owned native processes/profile/temp files/listeners; preserve existing user preferences, recordings and shared runtime/model resources. Do not delete/unload the model, remove the operator's approved cloud-disable setting or make a broad rollback.

DOCUMENTATION REQUIREMENTS

Update directly related README/index/local-generation/build/runbook docs with new actual prerequisite/model/native evidence. Qualify older empty-catalogue observations as historical; do not rewrite the original report. Clearly distinguish advertised capabilities, controlled tests, actual local generation and observed native behavior. Step-12 remains pending manual orchestrator acceptance even if the worker reports COMPLETE. Step-13 remains blocked/unassigned.

ACCEPTANCE CRITERIA

These evidence rows resolve the original criteria; they add no new feature:

Original criterion Required completion evidence
AC-01 / C-01..C-03 Verified current cloud-disabled runtime, exact eligible installed model/license/resource context, explicit selection/acknowledgement and actual matched source/knowledge
AC-02 / C-04,C-05,C-09 Original real-model/native decisive-goal response, same-validator exact +600/-10 two-entry diff and all other values unchanged; existing rejection/grounding contracts preserved
AC-03 / C-06..C-08 Controlled source/model/manual-preview/bounds/failure/owner regressions remain green; any integration fix has direct tests and no weakened publication/cleanup
AC-04 / C-10 All nine Release projects/594 regressions plus actual in-flight native Cancel/unmount/re-entry, manual/stale/units/coexistence and zero inference-caused business-source mutation
AC-05 Accurate completion report, own source-path self-review, aligned docs, preserved original evidence/history and manual stop

OUT-OF-SCOPE WORK

Step-13 demonstration, applying authoring values to dynamics, proposal Save/ target approval/programming, tools/MCP/agents/RAG, new model downloads/switches, runtime upgrades, automatic service/configuration management, new packages/ projects/hosts, persistent chat/training/benchmark systems and unrelated fixes. No ethics/model-preference reconsideration: the owner selected the installed non-stock 8B-v2 model. Qualification measures the actual frozen contract.

REPORTING REQUIREMENTS

Save prompts/Step-12-Completion-Followup-Report.md entirely inside exactly one plain triple-backtick fence, no language tag, outside content or nested fences. Identify Step-12 completion, status, original C/AC mapping and prior report. Include runtime/model/digest/license/resources, operator approval/exact setup change if any, fresh cloud-disabled proof, actual original response, canonical before/candidate/diff, every attempt/error/latency, native pending/Cancel/ unmount/re-entry evidence, explicit controlled-versus-native distinction, zero-business-mutation results, changed files/tests/docs, exact command exits/ counts, self-review, preservation, cleanup and any precise remaining blocker. Do not include secrets or private thinking traces.

COMPLETE requires every original remaining check. Otherwise use INCOMPLETE/BLOCKED and identify the exact unmet criterion/operator action. Preserve both existing blocked reports unchanged. Save and STOP; return manually to the orchestrator, with no Step-13 or automatic worker communication.