AutoForge/prompts/Step-12A.md

274 lines
16 KiB
Markdown

# Step-12A - Local-only Ollama setup and real-model qualification
## STEP IDENTIFIER
Step-12A, explicitly requested by the owner on 2026-10-31.
This is the active worker follow-up for still-unaccepted Step-12. It replaces
[Step-12-Completion.md](Step-12-Completion.md) as the execution handoff.
The owner requested the lettered identifier; this is not a claim that a new
application defect was found or that Step-12 has been accepted.
Read the frozen [Step-12 assignment](Step-12.md), both
[original](Step-12-Report.md) and [completion](Step-12-Completion-Report.md)
reports, the [previous Step-12A return](Step-12A-Report.md), [STATE.md](STATE.md)
and [WORKER-QUALITY-GATE.md](WORKER-QUALITY-GATE.md).
Preserve those reports unchanged. Return only
`prompts/Step-12A-Execution-Report.md` under `prompts`, then STOP for manual review.
Do not execute the superseded completion handoff.
Owner-directed execution revision, 2026-10-31 14:28 -04:00:
the repeated approval-form/quiescence-confirmation gate is removed.
The owner wants the existing local model to run and no Ollama cloud. The worker
is directed to perform the necessary documented cloud-disable configuration
and normal restart of this existing Ollama instance as needed, preserving the
model/store/version and avoiding known active shared work.
Do not request another approval form for these scoped actions or return the
same user-unavailable permission blocker. Tool AllowAll alone is not consent;
the authority here is the owner's direct local-only execution direction.
No LM Studio, new runtime/model download or alternate backend is needed.
All model-output/validator/native/regression criteria remain unchanged.
## STEP TITLE
Establish verified cloud-disabled Ollama and qualify the installed 8B model
through the real native runbook-preview workflow.
## OBJECTIVE
Complete the original Step-12 local-only prerequisite and real-model/native
evidence. Execute the work, rather than returning another general plan.
The owner requires **local models only; no Ollama cloud or cloud fallback**.
The worker owns runtime setup, application/native checks and any evidenced
original-contract integration fixes. The orchestrator only reviews evidence
and maintains orchestration documents. No automatic worker/agent handoff.
## ARCHITECTURAL CONTEXT
Preserve all nine projects, one Windows .NET 10 MAUI Blazor Hybrid Client,
existing Application inference/context/generation owner and authoritative
Step-11 validator. No alternate web host or canned native model response.
Generation proposes a read-only simulation calibration preview only.
It cannot save a proposal, approve/open/write a target, execute a runbook,
apply values to dynamics or perform physical operations.
Advertised model tools/thinking do not enable application tools.
## CURRENT IMPLEMENTATION STATE
- Step-11 accepted. Step-12 code/control tests present but Step-12 unaccepted.
- All three previous reports are INCOMPLETE/BLOCKED: the old approval form
was unavailable. No actual chat or model-intent qualification occurred.
That obsolete form gate is superseded by the owner direction above.
- Worker-reported baseline: clean nine-project Release build; 179 unit +
415 integration = 594 passes, zero failed/skipped. Preserve all cases.
- Installed selection: `huihui_ai/qwen3-abliterated:8b-v2`, local GGUF/qwen3,
8.2B, Q4_K_M, approximately 5 GB; completion/tools/thinking capabilities,
Apache License metadata.
- Last verified catalogue digest:
`543f9deff86d8347fd96ea1f3e54e73861ea70d52def71d31ccaeaf013de6ba6`.
- Last daemon: Ollama 0.40.13 at `http://127.0.0.1:11434`,
model store `D:\AI\ollama\models`. Shell OLLAMA_MODELS differed; preserve
the actual daemon store, not the shell's candidate.
- Last inspected `C:\Users\g3\.ollama\server.json` was absent and the current
serve owner's startup log said `Ollama cloud disabled: false`.
- Native model selection/full synthetic context and ordinary manual preview
work. Real generated exact-goal output and pending-chat lifecycle do not
yet have evidence. RTX 4090/24 GB VRAM and 64 GiB RAM were observed, not
measured as model throughput/deadline qualification.
Recheck current state. Do not reuse old PIDs or treat old logs as fresh proof.
## REQUIRED IMPLEMENTATION
### Contract/evidence matrix
| ID | Required action | Passing result |
|---|---|---|
| C-01 | Recheck current runtime/config/store/model | Exact installed local selection and minimal documented setup; no redownload, cloud/context dispatch or unrelated changes |
| C-02 | Admit owner-directed setup after inspecting current owners and known active work | No repeated approval form; avoid concrete active-job conflicts, exact existing instance only, preserve unrelated resources |
| C-03 | Perform scoped local-only configuration/restart and verify running server | Fresh responsive owner/startup with cloud-disabled:true and same installed model/store/version |
| C-04 | Submit exact original goal through real native workflow | Original model content reaches same validator unchanged; exactly +600 rpm/-10 N*m and every unrelated value unchanged |
| C-05 | Exercise actual pending generation and source/UI isolation | Conflicts reject, Cancel/unmount drain, no late preview, fresh owner, manual/stale/units work and zero inference-caused business mutations |
| C-06 | Complete regression/build/docs/report/cleanup | All 594 regressions and nine projects pass; evidence distinguishes controlled/native, history preserved and manual STOP |
### 1. Establish the local-only prerequisite
Use bounded loopback metadata probes, current listener/serve/tray ownership and
targeted config/startup-log inspection. Do not dump unrelated logs, environment
variables or private data. Verify exact model name/digest/local GGUF/completion,
positive installed size and no remote/cloud metadata.
If the running server is already verifiably cloud-disabled, record current
startup proof and continue. Otherwise perform this owner-directed setup:
1. Prepare the documented minimal configuration:
`{"disable_ollama_cloud":true}` in the actual supported
`C:\Users\g3\.ollama\server.json` location, merging any existing unrelated
settings. Recheck installed-version support with the
[official FAQ](https://docs.ollama.com/faq#how-do-i-disable-ollamas-cloud-features).
Use the installed runtime's documented setting; do not make machine-wide
environment changes or add automatic runtime management to the app.
2. Discover current serve/tray owners and inspect running-model metadata and
known active jobs before restart. An empty `/api/ps` is not a universal
attestation, and a loaded model alone does not prove an active job.
If concrete active shared work is observed, let it finish or report the
exact conflict; do not cancel/unload that work. Do not demand an unavailable
global-idle checkbox or send another permission form.
3. Preserve the specific prior config outside the repository if needed, make
only the cloud-disable change and restart this existing instance through its
normal tray Quit/relaunch mechanism. Do not select "Restart to update",
upgrade, reinstall or stop unrelated jobs.
If normal shutdown cannot be automated, a fallback may target only freshly
verified exact serve/tray PIDs belonging to this instance, with no known
active-job conflict. No process-name killing or unrelated process targets.
4. Preserve `D:\AI\ollama\models` and the existing installed version/model.
If starting a child process, prevent the worker shell's different model-
store environment from redirecting it. Do not relocate/delete/re-pull.
5. Verify endpoint readiness, new current owner/startup evidence reporting
`Ollama cloud disabled: true`, and the exact model/digest/local details.
A changed model/digest needs explicit operator review, not substitution.
If a concrete active-job conflict, missing access or inability to establish
running cloud-disabled proof prevents execution, report that actual blocker
and attempted remedy. Do not invent an idle assurance, send chat while cloud
is enabled or claim model failure. An unanswered approval form is no longer
an assigned prerequisite. Proceed without it within the scoped owner direction.
No further model/runtime downloads or different endpoint are authorized.
### 2. Qualify original model output in the actual native app
Build the real Client, launch it using existing
[native guidance](../docs/build-and-test.md), and verify rendered readiness.
Explicitly create/select the supported warm synthetic baseline, with no
diagnostic/dynamics/rehearsal programming operation:
- Warm-idle: 650 rpm.
- Torque axis: [650,1000,2000,3200,4200,5000] rpm.
- Torque cells: [240,300,360,350,300,180] N*m.
Capture actual workspace/head/applicability, full before values and initial
business-source state. Through `/runbooks`, explicitly refresh/select the exact
installed model, inspect actual synthetic source/knowledge context and only
then acknowledge independently verified cloud-disabled operation.
Submit the exact original goal:
`Propose warm-idle target 1250 rpm and set torque at 2000 rpm to 350 N*m; keep every other value unchanged.`
Capture original assistant `message.content`, model/digest/context/goal,
visible same-validator result, latency and every attempted outcome.
Passing intent means exactly 1250 rpm and the 2000-rpm cell 350 N*m:
two diff entries, +600 rpm and -10 N*m; all other cells/axes/values unchanged.
Parseable in-range wrong-intent JSON is not a pass.
Verify no authoring Save/head/history change, approval/target write, dynamics
action or archive/replay mutation was caused by generation.
Keep the existing 180-second whole-operation deadline, 4096 prediction-token
budget and all request/response/schema/source/preview bounds. No stripping
thinking/Markdown, JSON repair, injected IDs, clipped values or hidden retries.
Thinking traces are not report content or evidence. Retry only explicitly
against fresh current context; report attempts honestly.
If the real model/runtime exposes an actual integration defect, diagnose it
with evidence and make only a narrow fix that satisfies original Step-12.
Test the producer/consumer path and self-review after green. Do not weaken the
validator, alter source semantics, swap models or create a rule-based generator.
Unresolved material changes return as an explicit blocker for the orchestrator.
### 3. Finish native lifecycle and existing contract verification
Observe actual real-model work pending before testing repeated Generate/
refresh/selection rejection, Cancel/drain and separate navigation/unmount.
Verify no second owner/request, no late preview/UI publication, prior successful
preview retained and fresh entry with no old model/acknowledgement/owner.
An Idle or already-finished request followed by Cancel is not in-flight proof.
Do not substitute a controlled HTTP reply for this real-model native path.
Verify manual invalid input retains prior success, deliberate source/head
change marks preview Stale without rebasing, units change display only and
ordinary routes still work. Keep preparatory baseline/staleness actions separate
from inference-caused mutations. Use real before/after evidence and existing
feature-boundary counters for zero unassigned diagnostics/dynamics/archive/
authoring/rehearsal effects. Preserve canonical values, approvals and raw data.
## FILES OR PROJECTS EXPECTED TO CHANGE
The owner-directed external local-only config and directly related
README/docs verification status; only `prompts/Step-12A-Execution-Report.md` under `prompts`.
Application/UI/Client/tests only for an evidenced narrow original-contract fix.
No new project/package/host or automatic runtime management.
Preserve original Step-12, superseded completion prompt, all three blocked reports,
this assignment, roadmap/state/shared gate and all prior history.
## INTERFACE AND DEPENDENCY REQUIREMENTS
Keep original Step-12 C-01..C-10/AC-01..AC-05 and Step-11 strict validator.
Same explicit source/head/model/knowledge/preview/owner guards; no cloud,
proxies/redirects/authentication fallback, tool execution, arbitrary files/
code/credential upload, physical programming or private simulation oracle.
No agents, automatic handoff, staging, commit/push, Git backup persistence,
SDK/workload/model install/update or unrelated recording/navigation cleanup.
## TESTING REQUIREMENTS
After establishing local-only operation, run focused tests for any changed integration
and execute required complete commands sequentially:
- `dotnet build .\src\AutoForge.slnx --configuration Release --no-restore`
- `dotnet test .\src\AutoForge.slnx --configuration Release --no-build`
All nine projects/native Client and all 594 existing cases must pass; report
any additions separately. Restore code dependencies only after a genuine
chosen-command missing-dependency failure, never SDK/workload/model installation.
Separately execute C-04/C-05 real-model/native checks; green controlled tests
do not prove actual generated intent or pending native cleanup.
Do not repeat suites merely to restate the obsolete permission-form blocker.
If an actual access/active-job/setup failure occurs, report it honestly.
Cleanup only exact task-owned processes/profile/temp files/listeners. Release
test barriers in finally; no sleep-based race proof. Do not stop/unload the shared
runtime as native cleanup, delete the model, reset user preferences/archives,
remove approved cloud-disable settings or perform broad rollback.
## DOCUMENTATION REQUIREMENTS
Update directly related README/local-generation/build/index/runbook docs with
actual setup/model/native evidence and limitations. Keep older blocked/empty-
catalogue reports unchanged and clearly historical.
Worker COMPLETE is pending manual acceptance, not authority to start Step-13.
## ACCEPTANCE CRITERIA
| ID | Step-12A criterion and original Step-12 obligation |
|---|---|
| AC-01 | C-01..C-03: scoped setup/current cloud-disabled proof, exact installed local metadata/resources and concrete active-work handling; original AC-01 |
| AC-02 | C-04: original real native goal output, unchanged validator/exact two-entry diff/no mutation; original AC-02 |
| AC-03 | C-05: real pending conflicts/Cancel/unmount/fresh owner, manual/stale/units/isolation plus controlled guards; original AC-03/AC-04 |
| AC-04 | C-06: complete nine-project Release build and 594 regressions plus actual native-model proof; original AC-04 |
| AC-05 | Truthful C/AC evidence, own source-path self-review, docs/preservation/cleanup and manual report/STOP; original AC-05 |
## OUT-OF-SCOPE WORK
New model selection/downloads, cloud inference, policy debate, application
tool use/MCP/RAG/agents, proposal Save/approval/programming, applying changes to
dynamics, Step-13 demonstration, runtime upgrades/automatic managers, new hosts/
packages/projects, persistent chat/training systems and unrelated fixes.
This scoped local-only setup/restart is directed by the owner; do not interrupt
known active shared work or extend that direction to unrelated operations.
## REPORTING REQUIREMENTS
Save `prompts/Step-12A-Execution-Report.md` entirely inside exactly one plain triple-backtick
fence, no language tag/outside content/nested fences.
Include Step-12A/status; C-01..C-06/AC-01..AC-05 mapping to original obligations;
owner-directed scope, actual active-work inspection/setup and fresh startup proof; model/digest/
license/resources; original real response/attempts/latency; canonical before/
candidate/diff; actual pending/native versus controlled evidence; zero business
mutations; changed files/tests/docs; exact commands/exits/warnings/per-project
counts; self-review; preservation, cleanup and precise remaining blockers.
No secrets or private thinking traces.
COMPLETE requires all original remaining checks, not metadata/build alone.
Otherwise report INCOMPLETE/BLOCKED with the exact unmet action/criterion.
Preserve all three blocked reports unchanged. Save and STOP for manual orchestrator
review; do not invoke another worker/agent or begin Step-13.