AutoForge/prompts/Step-12.md

22 KiB

Step-12 - Local AI and applicable knowledge integration

STEP IDENTIFIER

Step-12. Step-11 was accepted on 2026-10-31 with no correction. This is the sole active implementation assignment. Step-13 demonstration is blocked.

STEP TITLE

Generate one synthetic calibration runbook with an explicitly selected local-only model and applicable context, then use the existing deterministic preview validator.

OBJECTIVE

One user outcome: an explicit bounded natural-language calibration request produces either a model-generated validated read-only preview or a precise error. One primary new contract: a local-only, source/model/context-bound cancellable generation whose untrusted output cannot bypass validation or mutation approval.

No authoring Save, target open/write, runbook execution, cloud inference or physical operation is assigned. Read STATE.md, ROADMAP.md, runbooks v1 and WORKER-QUALITY-GATE.md. Establish the frozen matrix before coding, self-review after first green, write prompts/Step-12-Report.md, and STOP.

ARCHITECTURAL CONTEXT

Keep all nine projects, one Client and independent Components. Application owns a small replaceable inference port, matched context provider and generation orchestration against actual authoring/preview contracts. An Ollama HTTP implementation may live in the existing Application feature with Client configuring the real transport; UI owns explicit selection/generation/ cancel/context/output/error and page lifetime. No new project/package/host, generic agent/tool/runtime framework or vector database.

Reuse RunbookPreviewService, the actual authoring service and existing synthetic definition/diff/units. The deterministic validator stays authoritative; do not implement a weaker model-only schema/source validator. Source/tests remain under src, tests under src/tests.

CURRENT IMPLEMENTATION STATE AND PREREQUISITES

Accepted manual runbooks allow one simulation/proposeCalibration step, exact source/head/full scalar/table values, strict schema/limits, immutable preview/ diff and Stale status. They save/approve/execute nothing. Worker baseline: 179 unit + 368 integration passes (547 total, zero failed/ skipped), clean nine-project Release build and native manual-preview checks. The orchestrator did not run those commands.

First provider: Ollama local HTTP at http://127.0.0.1:11434. Use an existing operator-approved runtime and installed local model. Do not install/pull/update models or runtimes, sign in, start/restart a shared service or alter user/system settings automatically. Real model availability was not verified by the orchestrator. Missing prerequisites block the assigned real-generation/native checks; do not substitute a stub or cloud and declare COMPLETE.

Verified public contracts, recheck against the installed runtime before coding:

  • Chat API: POST /api/chat, explicit model/ messages, stream:false, structured format; read message.content.
  • Model catalogue: GET /api/tags, local model name/digest/size and optional remote_model/remote_host.
  • Model details: POST /api/show, completion capabilities and model metadata.
  • Local-only configuration: operator disables cloud with documented server configuration or OLLAMA_NO_CLOUD=1, then verifies the runtime's cloud-disabled status.

Use the actual documented anchor/path the installed docs expose if necessary; do not invent unsupported APIs. A loopback URL alone does not prove the Ollama model will not delegate to cloud. No API key/authentication is needed or allowed.

REQUIRED IMPLEMENTATION

Frozen contract matrix

ID Input/precondition Required result
C-01 App/route entry or ordinary manual workflow without a model No automatic network/generation/model/source action; manual validator/authoring/diagnostic/replay/rehearsal still works. Explicit catalogue refresh and opt-in/model selection are required.
C-02 Eligible installed local model versus remote/cloud/ambiguous/stale metadata, absent runtime/model or redirect/proxy Only loopback local-only eligible selection can generate; precise visible rejection before goal/context reaches an ineligible destination. No model substitution, pull/install/cloud/tool call.
C-03 Actual current supported authoring source/head and matched synthetic knowledge Capture exact immutable source/head/values/constraints, model name/digest, knowledge ID/revision/applicability and bounded goal. Missing/mismatched source/knowledge rejects before chat.
C-04 Controlled valid response and real native goal from warm baseline Original model content passes existing Step-11 validator; exact 1250 rpm and 2000-rpm torque 350 N*m yield +600/-10 diff with unchanged other cells. No saving/approval/write; real model evidence distinguishes intent accuracy from mere parse success.
C-05 Invalid/refused/empty/markdown/truncated/tool-call/wrong-source/wrong-unit/unknown-key model output Typed generation/validation error, no repaired/rebound/default candidate; previous successful preview remains intact. Unvalidated content is safely labeled/limited, never actionable authority.
C-06 Source/head changes, newer manual preview or model selection/digest superseded while generation is pending No stale overwrite or rebasing. Publish only for the current owner/source/head/selected model and captured prior-preview identity, through atomic validator publication checks.
C-07 Repeated Generate, competing refresh/selection, Cancel/close/re-entry and delayed response One owner; conflicts reject before mutation/request, cancellation/close drains outside required gates, no late preview/Ready state, fresh page lifetime does not reuse closed completion.
C-08 HTTP errors/timeouts/invalid metadata/duplicate critical response fields/chunked oversized body Typed visible failure; bounded actual reads/request sizes, no success-shaped empty catalogue/content or partial JSON publication. Preserve prior source/preview/core workflows.
C-09 Prompt/rationale attempts to alter policy/source, global units/renders/hide Quoted goal/context/model output remain untrusted data; deterministic allowlist/source/template still applies. No tool/write/model control, validation on render or unit-based source mutation.
C-10 Real generation/coexistence and complete verification All 547 baseline cases/new coverage and nine-project Release build pass; real installed local-model/native preview runs with cloud disabled; counted existing transport/model/archive/rehearsal/authoring mutations remain zero.

A. Explicit local-only provider and transport

Use a small replaceable typed inference port; implement Ollama first using built- in .NET HTTP/JSON, not a new SDK dependency. Production base URI is fixed loopback above; tests may inject a controlled loopback HTTP peer/handler below the provider. Reject non-loopback destinations. Disable redirects, system proxies and default credentials; requests cannot carry API keys or resolve an arbitrary model-supplied URL. Allowed calls are catalogue/details/chat only, never pull/create/delete/push/ cloud/search/tool execution. Do not stop or unload unrelated runtime/model work.

Require explicit local-only operator acknowledgement before sending a goal or source context. Native verification must record actual operator-configured cloud- disabled runtime evidence, not a boolean checkbox claimed as cryptographic proof. Do not modify the runtime configuration yourself without separate approval.

Select an exact reported installed model, not freeform fallback/first model. Eligibility requires usable name/digest, positive installed size, local model details and completion capability. Reject remote_model/remote_host indicators, cloud-tagged models and malformed/ambiguous eligibility data. Show rejected models/reasons, not a misleading healthy empty catalogue. A failed refresh may retain last-known display but never authorizes generation without fresh checks. Recheck selected catalogue identity/digest/details before each explicit chat. After completion, recheck that same local eligibility and catalogue digest before validator publication. Changed/unavailable metadata rejects the result; a model name in the chat response alone does not attest its digest or local execution.

Use one non-streaming POST /api/chat per explicit Generate: stream:false, structured JSON format for the accepted runbook shape, bounded generation options such as temperature 0 and a finite output-token budget. The format is a generation hint, not validation or a guarantee of determinism. Supply the application-generated runbook/source template and require source/ head/keys/units/axes unchanged. Do not grant tools or feed previous unbounded chat history. Ignore reasoning fields as content; never interpret tool calls.

Validate the completed response envelope: explicit requested model, assistant message string content, actual finished state and non-truncated outcome. Reject empty/refused/malformed/truncated or tool-call completion. Handle additive API metadata compatibly, but reject duplicate/invalid critical identity/completion/content/error fields. No silent defaults for those fields.

Fixed bounds, not new user settings: goal <=2000 UTF-16 code units, complete UTF-8 chat request <=131072 bytes, each catalogue/details/chat response <=1 MiB (1,048,576 bytes), HTTP JSON depth <=64 and operation deadline <=180 seconds. Read bodies bounded even without Content-Length; a one-byte-over/chunked response cannot be buffered unboundedly before rejection. Distinguish timeout/cancel/ HTTP/metadata/completion/validator failures; do not return fake success. The decoded runbook must also meet all unchanged Step-11 limits.

B. Small applicable local knowledge and exact source snapshot

Provide one local synthetic engine-authoring policy pack with explicit ID/revision and exact model/definition applicability. Its authority is the accepted synthetic template/owned contracts, not manufacturer or real-vehicle truth. Retrieve by exact source model/version/definition/revision; no generic "nearest" or role-based fallback. Missing/mismatched pack is a visible pre-chat failure.

Context includes copied actual authoring workspace/head/applicability, canonical whole values, scalar bounds, fixed rpm axis, torque-cell bounds/units, supported operation and the pack's relevant synthetic restrictions. Keep the pack provenance and included facts visible independently of anything the model claims. Use a simple local provider/data definition, not embeddings/web retrieval/RAG/ filesystem crawling or another schema/persistence subsystem.

Capture the bounded explicit goal, selected model/digest and prior successful preview ID at request start. Show source and selected context for deliberate user review. Do not include arbitrary files, source code, raw diagnostic archives, credentials, vehicle personal data or undocumented operating instructions. Only bounded explicitly selected synthetic authoring data is supplied.

System guidance restricts generation to synthetic warm-idle/fixed-axis torque proposals, with no hardware programming, emissions/anti-theft/odometer/safety bypass instructions, external tools or claims of approval/testing. Treat the goal/pack/output as quoted untrusted data; model obedience is not the safeguard. Existing deterministic schema/source/template validation is still mandatory.

C. One generation owner and unchanged validator publication

Explicit Generate captures an immutable request/context; it never saves an authoring revision or opens/approves/mutates a target. At most one catalogue/generation operation is active in the page-owned feature. Reject competing Generate/refresh/model-change actions visibly before starting another request. Cancel/Close share a drainable cleanup as appropriate; no gate is held across an await needed by cleanup. New page entry creates a fresh owner.

After HTTP completion, enforce current owner/generation/cancellation/model/ source/head and prior-preview identity before publication. A newer manual preview must not be overwritten even if its authoring head is unchanged. Make these checks atomic with accepted validator publication: minimally extend the existing service with expected-preview/cancellation/current-request admission if needed, reusing its strict parser/source/head/template/diff logic. Do not merely precheck then call an unguarded publish that can race. Preserve all existing manual Validate behavior and tests.

Pass the original decoded message.content through the authoritative validator, without stripping markdown fences, fixing syntax, injecting absent IDs, changing unit/source/head, clipping values or replacing invalid output with a template. No auto retry/correction loop. A retry is an explicit new request against fresh current context. Valid JSON alone does not establish correct user intent; show the original goal and exact candidate/diff for human review.

Successful output is only an AI-generated read-only validated preview, still unverified for intent/physical applicability/testing/approval. Retain model/digest/ context/goal provenance separately from source facts; model rationale is safe text, not evidence or policy. Existing staleness still applies after publication. Keep only bounded latest generation evidence/prior success, not hidden history.

Failures/cancellation/close/source or manual-preview replacement preserve prior success and source state. Late callbacks cannot publish after owner close or a new request. No fault becomes healthy "empty model result", Ready, Saved or Applied. Manual/no-model workflows remain usable and errors are visible.

D. Real UI and real-model evidence

Extend the existing runbook area with explicit local-only acknowledgement, catalogue refresh, installed model selection/eligibility, bounded goal, source/ knowledge preview, Generate/Cancel and separate model-output/error/provenance. No network request on route entry/render/unit change.

Show valid output through existing candidate/diff/current-Stale presentation. Invalid output is labeled unvalidated/not an actionable preview, bounded and escaped. Separate old successful preview from rejected/changed request data. Global units only project human-readable values; requests/content/diffs stay canonical and unchanged. No Save/Approve/Execute/Apply or tool control. Page unmount cancels/drains its own request without touching shared Ollama.

Native decisive goal on the actual warm baseline: Propose warm-idle target 1250 rpm and set torque at 2000 rpm to 350 N*m; keep every other value unchanged. Require a real local model response whose validated canonical candidate and diff match that exact goal, not merely any parseable in-range result. Report all attempts honestly. No request repair, canned native response or hidden rule-based generator counts as local model success.

FILES OR PROJECTS EXPECTED TO CHANGE

Focused Application inference/provider/context/generation and narrow guarded preview publication, UI runbook generation/lifetime, Client HTTP DI only if needed, tests and related README/docs. No new project/package/host or model installation, existing template/authoring-save/rehearsal/physics/protocol change. Only prompts/Step-12-Report.md under prompts.

INTERFACE AND DEPENDENCY REQUIREMENTS

Keep nine projects/references, one Client, independent Components, immutable source/runbook/authoring values and all template/schema/approval/CAS limits. No public network/cloud/model pull, proxies/redirect fallback, source metadata backfill, model tool execution, source code/credential upload or arbitrary raw command. Generated knowledge cannot promote authority or bypass approval. Do not fix the separate recording navigation issue or add Git/durable backups. No SDK/workload/runtime/model/system install/update/repair/rollback, staging/ commit/push or destructive Git; preserve existing work and protected history.

TESTING REQUIREMENTS

Map C-01..C-10 to meaningful named tests before handoff:

  • Real provider over a controlled local HTTP peer/handler: catalogue/details/ completion fields, model eligibility/digest, remote/cloud/redirect/proxy exclusion, one actual chat request, structured options/context and error classification. This proves HTTP integration, not that a real LLM generated the test response.
  • Current-source/knowledge/no-match vectors and original validator negative outputs, with exact previous-preview/source/head/history preservation. Hold real workflow response/read publication to exercise cancel/source changes/ newer manual preview/closed/new-owner races, with atomic rejection assertions.
  • Fixed byte/depth/goal/deadline and missing/invalid/duplicate critical metadata/ chunked response bounds; no broad catches or count-only claims. Release owned helpers/barriers in finally; do not use sleep/polling as race proof.
  • Actual rendered selection/opt-in/generation/cancel/error/stale/units controls, no generation on render, and real diagnostics/recording/replay/approved- rehearsal coexistence with counted zero unassigned business-source calls.
  • Keep all 547 baseline cases (179 unit, 368 integration), including manual validator/authoring/source/physical gates. Compare canonical request/output/ source/config/archives and prove no mutation/execution.

Run focused cases, then sequentially:

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

Build all nine projects/native Client. Install/restore code dependencies only after a genuine chosen-command missing dependency; this does not authorize model/runtime installation or pulling. Recheck actual installed API behavior.

Required native/real-model smoke: operator-approved local-only runtime already running, verified cloud-disabled status and installed eligible model/digest; explicit actual source/context, decisive goal above, original real response, same-validator exact diff, no source/approval/target/dynamics-configuration/ recording changes, manual invalid/source-stale handling, cancel/unmount and old routes/units. Record runtime/model/version/license/resource context honestly, without secrets or thinking traces. Controlled faults are not native platform observations. Absent runtime/model/local-only evidence or failure to produce the assigned valid native response is BLOCKED/INCOMPLETE, not a waiver or stub-success. Stop only task-owned test HTTP/native helpers; do not terminate a shared model runtime or reset user data/preferences. No installation/download is automatic.

DOCUMENTATION REQUIREMENTS

Update README/index/architecture/build guide and local-generation/runbook usage: explicit local-only opt-in/eligible model, operator prerequisites, verified public API contracts, cloud delegation caveat, matched synthetic pack/source/context, strict unchanged-validator output, current-owner/manual-preview CAS/cancel/drain, real-model versus stub evidence and no write/tool/physical authority. Step-11 is accepted; Step-12 remains pending manual review. Broader diagnostics/ scenario research/executors, vector knowledge and demonstration are future.

ACCEPTANCE CRITERIA

ID Criterion Evidence
AC-01 C-01..C-03 explicit local-only transport/model eligibility and exact applicable context Real HTTP boundary/metadata/remote-negative tests and operator/runtime local-only prerequisite evidence
AC-02 C-04/C-05/C-09 original model content uses authoritative validation without repair/authority/side effects Exact controlled outputs, unchanged negative-parser/source tests and actual real-model decisive goal
AC-03 C-06..C-08 one owner, atomic source/model/manual-preview publication and bounded error/cancel/close handling Controlled interleavings/limits/HTTP failures and rendered command/lifetime assertions
AC-04 C-10 all 547 regressions/new coverage, real-model/native workflow and full nine-project Release build pass Exact worker commands/exits/warnings/counts, zero business-source mutation and native/stub distinction
AC-05 Matrix/self-review/docs/history/protection/manual gate accurately describe limits/prerequisites Own post-green producer/consumer review, named evidence and prompts/Step-12-Report.md

OUT-OF-SCOPE WORK

Cloud/public inference, model/runtime installation/pull/update/agent hosting, tool/runbook execution, proposal Save/approval/programming, broad diagnostic/ scenario planning, embeddings/vector/RAG/web research, persistent chat/model/ runbook/Git backup storage, new templates/unit families/project/package/host, hardware/physics changes and unrelated cleanup. No future assignment, agents, automatic handoff, orchestration/shared-gate edit, staging, commit or push.

REPORTING REQUIREMENTS

Save prompts/Step-12-Report.md wholly inside exactly one plain triple-backtick fence, no language tag/outside text/nested fences. Include identifier/title/status, functions/files/interfaces, C-01..C-10/AC-01..AC-05 mapping with named executed evidence, exact local-only/model/digest/context/response/validator/publish/cancel/ negative/limit/no-side-effect outcomes, commands/exits/warnings/per-project counts, actual real-model/native versus stub/control evidence and prerequisites, docs/README, own post-green source-path self-review, limitations/deviations/ blockers and preservation. COMPLETE requires all assigned checks; otherwise report INCOMPLETE/BLOCKED with the precise unmet criterion. Save and STOP; do not begin Step-13.