AutoForge/prompts/ROADMAP.md

55 KiB

AutoForge development roadmap

Status and authority

  • Initialized: 2026-10-28.
  • Current phase: corrected deterministic dynamics accepted; first bounded live dashboard next.
  • Active assignment: Step-05, ready for manual worker execution.
  • Last accepted milestone: Step-04 with Step-04A/Step-04B, accepted 2026-10-28.
  • Last accepted correction: Step-04B; all Step-04 findings closed.
  • Outstanding corrections: none.
  • Step-05 status: READY FOR WORKER; no dashboard implementation/report yet.
  • Current state and handoff: STATE.md.
  • Authority: the master charter, repository/handoff amendment, and the owner's 2026-10-28 architectural realignment. The latest decision cancels the dual-shell requirement and adopts one MAUI Blazor Hybrid executable plus vertical slices. The amendment's layout, naming, protected files and manual workflow still apply.
  • This document describes the intended architecture and future capabilities. Planned projects, interfaces, milestones, and demonstrations are not claims of implemented or verified functionality.

Product goals

Build one extensible, local-first .NET 10 platform combining:

  1. Deterministic vehicle simulation and virtual electronic control modules.
  2. Diagnostics, telemetry, visualization, calibration management, and eventually explicitly authorized module-specific programming.
  3. Local AI planning and vehicle knowledge with independently validated runbooks.

Support arbitrary vehicle module inventories, not a fixed ECU/TCU list. Vehicle, module, protocol, and calibration definitions must be data-driven and applicable to specific hardware/software identities. Unknown or restricted capabilities must remain explicit. Simulation, manual diagnostics, manual calibration work, and tests must function without an AI service or cloud account.

Preserve the complete product vision through incremental, integrated vertical slices. Do not finish one major subsystem in isolation before integrating it.

Repository and manual workflow

The only top-level project directories are docs, src, and prompts. Necessary root configuration files are allowed. Preserve Git metadata and existing protected repository infrastructure; these are not project directories.

  • All application projects and automated test projects belong under src.
  • The primary solution will be src/AutoForge.slnx.
  • Tests belong under src/tests, never a top-level tests directory.
  • Generated test evidence, if retained, belongs in an ignored location under src, not a top-level artifacts directory.
  • Technical/product documentation belongs under docs.
  • All file-based orchestrator/worker communication belongs under prompts.
  • Do not create a nested AutoForge repository directory.
  • Do not initialize a repository, automatically commit, or push.
  • .github/copilot-instructions.md, if present, is read-only. If absent, do not create it or an alternative intended to override it.

The orchestrator writes assignments and maintains this roadmap and state. The worker alone writes application code, tests, technical documentation, and the root README; executes builds/tests; and writes the corresponding report. The orchestrator reviews the assignment, report, relevant source and tests, Git changes, documentation, and actual reported command evidence. It does not execute application builds/tests or implement fixes.

Handoffs are manual. No agent invocation, messaging automation, polling, watchers, automatic prompt pickup, or unattended progression is permitted.

Assignments use Step-01.md, Step-02.md, etc. Reports use Step-01-Report.md, Step-02-Report.md, etc. Corrective assignments use the original number plus sequential uppercase letters, for example Step-01A.md and Step-01A-Report.md, then Step-01B.md. Accept all necessary corrections before assigning the next major step. Do not use the superseded numeric-dot prompt/report naming convention.

Assignment sizing and handoff policy - revised 2026-10-28

The owner requested an orchestration-level reassessment to reduce repeated corrective handoffs. Longer prose and larger passing test counts have not been enough. The changes below are prospective; they do not add acceptance criteria retroactively to completed assignments.

Evidence and causes

  • Step-01A was an owner-directed architecture change, not a preventable implementation defect. Keep that separate from quality failures.
  • Steps 02-04 combined many new contracts: parsing, state/capability semantics, protocol validation, physics, async ownership and UI integration. The default paths could pass while invalid/partial/reentrant paths remained untested.
  • Assignments described broad responsibilities but did not always supply decisive input/outcome and transition matrices before coding. Review was discovering those contracts too late.
  • Orchestrator review also surfaced some parent-milestone cases in successive passes. Future review must consolidate the whole assigned contract, not rely on passing counts or accept only an attractive demonstration.

Changes effective for the next handoff

  1. Use WORKER-QUALITY-GATE.md for all new assignments, beginning with Step-04B. It supplements the bounded assignment and existing manual/protected-file rules; it does not authorize agents or automation.
  2. One observable user outcome and one primary new behavioral contract per assignment. Do not combine several new formats/state machines/engines merely because they belong to the same milestone. Integration remains end-to-end.
  3. Before implementation, supply a compact contract matrix: normal, rejected, boundary/partial, lifecycle/interleaving and UI/result outcomes where relevant. Name exact invariants and test vectors; omit irrelevant rows with a reason.
  4. The worker tests those cases, then performs its own source-path self-review after the first green run. It fixes findings inside the active assignment before declaring COMPLETE. Test totals are evidence of execution, not coverage.
  5. Reviewer and worker use the same frozen criteria. Do not turn optional improvements or new feature preferences into corrective requirements. Genuine defects, regressions, safety problems and unmet existing criteria still block acceptance, with one consolidated narrowly scoped correction.
  6. Preserve manual handoff and truthful BLOCKED/INCOMPLETE reporting. No prompt can guarantee zero defects; the goal is fewer preventable handoffs, not weaker review or concealed failures.

Unassigned future rows remain milestone envelopes, not a requirement to implement their entire scope in one worker turn. If an envelope is too large, split and update the future major-step numbering/dependencies before handoff; letters remain reserved for actual corrections. Do not renumber completed history.

The next dashboard milestone is narrowed to the first real configurable live dashboard: numeric and bar displays, explicit binding/range/unit/error states and one acquisition owner. Charts, drag/drop, complex layouts and advanced analytics remain in the later visualization milestone. Calibration/runbook and hardware envelopes must likewise be risk-sliced before assignment.

Architecture baseline

Existing accepted foundation, preserved

Step-01 established four projects, all preserved by Step-01A. Their current references:

Project Responsibility Allowed project references
src/AutoForge.Core Platform-neutral foundational value types, invariants, and shared contracts None
src/AutoForge.Application Use-case coordination and application-facing services; target/profile features and pending-review diagnostic sessions Core, Diagnostics
src/tests/AutoForge.UnitTests Domain/application behavior, generic component, protocol and trace tests Core, Application, Components, Diagnostics, Simulation
src/tests/AutoForge.IntegrationTests Real application/RCL/diagnostic composition and project-graph tests, without hardware/cloud Core, Application, Components, UI, Diagnostics, Simulation

The original foundation was headless. Accepted Step-01A added Client/UI/Components without discarding these projects or their 54 cases. The solution now has seven projects and 65 worker-verified tests (39 unit, 26 integration). The old deferred dual-host plan remains canceled.

Corrected application architecture

AutoForge is a modular monolith with Vertical Slice Architecture for application use cases. Its sole initial application executable is src/AutoForge.Client, a .NET 10 MAUI Blazor Hybrid application, initially Windows. Simulation, diagnostics, calibration and AI services can run in-process. No standalone WebAssembly shell, separate web application, or backend executable is required. Additional processes/hosts need a concrete requirement and an approved decision; do not scaffold speculative alternatives.

Accepted Step-01A established:

Project Responsibility Allowed project references
src/AutoForge.Client Thin native MAUI startup, lifecycle, configuration, BlazorWebView hosting and DI/platform composition UI, Application, Simulation composition and justified platform adapters; no test projects
src/AutoForge.UI Complete application routes, pages, navigation, layouts, feature presentation and interaction workflows Components, Application/Diagnostics contracts and justified domain contracts; never Client/Simulation
src/AutoForge.Components Generic reusable status controls, inputs, gauges/charts/layout building blocks as actually needed Framework UI dependencies; no UI, Client, hardware or application feature implementation

Primary UI direction: Client -> UI -> Components. The shell contains only platform/bootstrap code, not routed application pages, business workflows, diagnostic protocols or simulation algorithms. Application UI lives in the UI RCL; vehicle-independent reusable widgets live in Components. No circular references.

Application remains necessary and owns the working target-selection slice at Features/ExecutionTargets/SelectExecutionTarget, with its public namespace preserved. Future slices group operations under actual feature areas such as Vehicles, Diagnostics, Simulation, Calibration, Telemetry and AI. Do not create their empty directories now. Slice-local request, handler/service, validation and result types are added only when useful. MediatR, generic repositories/managers and uniform boilerplate are not required.

Step-02 has added Core vehicle/profile models, an Application Features/Vehicles/Profiles import/catalog/select/export feature and the existing UI's /vehicle-profiles page. Step-02A closed the timestamp and optional evidence constraint gaps; the corrected Step-02 milestone is accepted. Worker verification now covers 103 unit and 37 integration cases (140 total).

UI feature presentation follows the same feature areas where practical; it calls application contracts without duplicating business validation. Reusable engines, protocols, calibration and telemetry capabilities remain independently testable libraries, extracted only when needed. They must not depend on UI, Client or the application slice that coordinates them. Do not force all technical libraries to reference Application.

Planned boundaries, added only when used

Boundary / likely project Responsibilities and dependency direction
Core Vehicle/module identities, capability states, provenance, units, versioned domain schemas; no UI, inference, transport implementation, or storage dependencies
Application Feature/use-case slices coordinating reusable capabilities; existing target selection retained; depends inward on domain and needed capability abstractions
Diagnostics Addressing, protocol/service handling, decoding, transport interfaces and communication trace; depends on domain/transport abstractions, not UI or coordinating slices
Simulation Virtual clock, coupled dynamics, virtual modules, simulated transport, deterministic faults; depends on shared contracts/Diagnostics, never on an LLM or UI
Persistence / telemetry infrastructure Recording, replay, calibration/version storage, audit and knowledge repositories implementing application ports; introduce a project only when justified
AI / knowledge integration In-process replaceable inference/research/retrieval capabilities behind restricted planning contracts; no arbitrary hardware-driver access
Hardware adapter Verified OBDX Pro GT/native integration behind shared diagnostic ports; Windows-specific integration composed by Client and isolated from reusable UI
AutoForge.Components Generic data-bound gauges, charts, controls and status/layout components, introduced only as used
AutoForge.UI Application feature pages/routes/navigation/layouts; composes Components and calls application features
AutoForge.Client Sole initial executable: Windows-first MAUI Blazor Hybrid, thin startup/lifecycle/BlazorWebView and in-process DI composition

These are logical responsibilities, not an instruction to create all assemblies now. Keep related responsibilities together until an actual dependency or platform boundary justifies extraction. Avoid premature microservices.

Step-03 has added two actual capability boundaries: AutoForge.Diagnostics owns diagnostic/provider/transport contracts, a clearly synthetic protocol and decoding; AutoForge.Simulation implements virtual modules and simulated transport against those contracts. Application coordinates Diagnostics, not Simulation internals. Client composes Simulation in-process. Simulation depends on Core/Diagnostics, never Application/UI/Client or an LLM. The current solution now has nine projects. Step-03A closed the frame/reply/trace/ clear-verification findings, and Step-03B separates successful probe evidence from identity availability. The corrected Step-03 milestone is accepted. Worker verification now covers 111 unit and 72 integration cases (183 total).

Step-04 has extended the existing Simulation implementation with coupled dynamics and a fixed-step virtual clock. Shared platform-neutral configuration/control contracts belong in Core or an appropriate existing shared-contract location; Application still does not depend on Simulation implementation. No additional project is assigned. Static synthetic diagnostics remain available, separate from an explicitly selected dynamics provider/run configuration. The dynamics code is accepted with Step-04A/Step-04B: reference handling, drivetrain invariants, terminal/progress outcomes and controlled owner-drain rejection have been reviewed. Worker verification now covers 126 unit and 81 integration cases (207 total).

Planned dependency flow:

flowchart TD
    Client[AutoForge.Client - MAUI Blazor Hybrid] --> UI[AutoForge.UI - application RCL]
    UI --> Components[AutoForge.Components - reusable RCL]
    Client --> App[Application feature slices]
    UI --> App
    App --> Core[Core domain and contracts]
    App --> Capabilities[Reusable simulation / diagnostics / calibration / telemetry / AI capabilities]
    Capabilities --> Core
    Client --> Native[Platform adapters]
    Native --> Ports[Appropriate shared capability ports]
    App --> Ports
    Ports --> Core

Ports in the diagram are logical contracts, not an additional required project. Client is the native composition root; headless tests compose application services without Client. Capabilities in this diagram are logical responsibilities, not instructions to create all assemblies. Dependency injection does not grant AI or UI unrestricted programming authority.

Invariants

  1. Explicit target: simulation, recorded replay, and physical hardware are distinct execution targets. Never infer, silently switch, or fall back between them. Descriptor selection is not a connection or a successful operation.
  2. Layered diagnostics: separate physical transport, vehicle network, addressing, diagnostic protocol/services, decoding, calibration operations, and application commands. Generic OBD-II does not imply access to all modules.
  3. Determinism: fixed virtual time and reproducible inputs/seeds drive simulation. Do not use wall-clock timing, rendering frequency, or LLM sensor generation as the physics source. Timeouts/faults must be reproducible.
  4. Physical consistency: RPM, gear ratio, speed, torque/load, fuel use, and temperatures derive from coupled state within documented model assumptions. Simplified predictions must not be marketed as validated real-world results.
  5. Independent oracle: tests may inspect simulation ground truth through an isolated test boundary; application and AI paths may not bypass diagnostics to access it.
  6. Truthful data: distinguish vehicle configuration, firmware, calibration, calibration definitions, diagnostic captures, and telemetry recordings. Version persisted schemas; preserve originals and applicability information.
  7. Telemetry lineage: preserve source/module identity, units, sampling, data quality, raw/decoded/derived/estimated/simulated classification, and both simulation and capture time as applicable.
  8. Separate operations: live simulation, recorded-session replay, and deterministic scenario re-execution are not interchangeable. Comparisons identify source, configuration, scenario, confidence, and limitations.
  9. Safe execution: AI proposes versioned logical runbooks. A deterministic validator/executor checks identity, applicability, limits, dependencies, approvals, preconditions, backup, readback, unrelated-data preservation, diagnostics, and supported recovery.
  10. Three verifications: backup verification, programming verification, and functional verification are distinct. Missing, interrupted, or uncertain verification cannot be reported as success.
  11. Evidence and authority: VERIFIED, UNVERIFIED, INFERRED, and SYNTHETIC information retain sources, applicability, confidence, verification status, and limitations. Similarity search is not authority for offsets or exact compatibility. Unknown calibration locations remain unknown.
  12. Physical gates: first hardware integration is read-only. Capability discovery, transport selection, simulation approval, or an AI recommendation never authorizes physical writes. Physical programming requires separate explicit authorization and module-specific controlled evidence.
  13. Prohibited functionality: no emissions defeat, anti-theft bypass, mileage falsification, safety-system disabling, or unrestricted safety-critical writes.
  14. Local-first and replaceable AI: allow local OpenAI-compatible endpoints, initially considering LM Studio and Qwen3.5-9B around a practical 16 GB model memory budget. Verify actual availability, licensing, context, and memory use at implementation time; do not make a model or fine-tuning a core dependency.
  15. Honest errors and support: no fake providers, silent defaults, fabricated manufacturer protocol behavior, fabricated backups, or unverified platform claims. Protect vehicle identifiers, recordings, model assets, and local service endpoints. Never commit secrets or private runtime datasets.

Milestones and acceptance gates

Step-01, Step-01A, corrected Step-02 and corrected Step-03 remain accepted. Corrected Step-04 is accepted with original and corrective evidence preserved. Step-05 is the only active assignment and uses the revised bounded-scope/ contract-matrix/worker-self-review policy. Dashboard and later capabilities remain unimplemented. Dependencies indicate acceptance prerequisites, not permission for automatic handoff. Later assignments refine these scopes using actual implementation evidence; they must not silently discard required capabilities.

Step Milestone Depends on Status Principal acceptance gate
01 Repository and solution foundation Repository inspection ACCEPTED 2026-10-28 Four-project .NET 10 foundation, explicit target-selection behavior, real unit/integration tests, successful recorded build/test evidence, accurate README and architecture docs
01A Single-shell and vertical-slice architectural correction 01 accepted; owner realignment ACCEPTED 2026-10-28 Seven-project Client -> UI -> Components foundation; preserved Core/Application; 65 passing cases and successful solution/native build/startup reported; corrected ADR/docs inspected
02 Vehicle profiles and module model 01A accepted; assignment regenerated ACCEPTED 2026-10-28 AFTER 02A Immutable versioned profile metadata, strict interchange, provenance/capabilities, exact selection and real UI; correction evidence inspected
02A Profile timestamp and optional evidence validation correction 02 review findings ACCEPTED 2026-10-28 Explicit-offset parser and optional-constraint semantics inspected; actual pre-fix failures and 103 unit/37 integration passes plus native checks reported
03 Virtual modules and diagnostic communication 02 and 02A accepted ACCEPTED 2026-10-28 AFTER 03A/03B Real synthetic wire/provider/session/virtual-state path, truthful phase discovery and verified clear; complete corrected evidence inspected
03A Synthetic wire validation and verified-clear correction 03 review findings ACCEPTED FOR R-01/R-02/R-03 2026-10-28 Strict wire and semantic reply validation, matching trace and Application verified clear inspected; 111 unit/64 integration passes plus native baseline reported
03B Partial discovery and identification correction 03A review finding R-04 ACCEPTED 2026-10-28 Independent probe/identity outcomes, retained partial services, accurate counts and fail-closed actions inspected; 183-case/native evidence reported
04 Deterministic vehicle dynamics 03, 03A and 03B accepted ACCEPTED 2026-10-28 AFTER 04A/04B Coupled numeric scenario gates, strict config/run/v2 binding, typed failures and deterministic lifecycle ownership inspected
04A Dynamics input, drivetrain and lifecycle integrity correction 04 review findings ACCEPTED 2026-10-28 AFTER 04B Reference/kinematic/progress/failure/ownership corrections and independent/controlled evidence inspected
04B Side-effect-free lifecycle rejection during owner drain Existing 04A R-03/AC-04 ACCEPTED 2026-10-28 Frozen five-case overlap/retry contract satisfied; 126 unit/81 integration passes, clean Release/native evidence and worker self-review supplied
05 First configurable live dashboard 04 and corrections accepted; 01A foundation READY FOR WORKER Numeric/bar gauges, validated in-memory bindings/ranges/units, explicit data states and one session-owned acquisition feed; no new schema/host/charts/drag-drop
06 Telemetry recording and replay 05 PLANNED Persist raw messages and decoded signals with metadata; playback/seek/speed controls feed the same gauges without silently reconstructing missing data
07 Calibration management and simulated programming 06 PLANNED Versioned definitions/diffs, verified baseline and backup, approved simulated write, readback/unrelated-data checks, audit and supported restore; no physical writes
08 Local AI, knowledge, and validated runbooks 07 PLANNED Replaceable local inference, applicable knowledge retrieval, structured logical runbook, independent validation/approval and simulation-only execution
09 Comparative evaluation and complete synthetic demonstration 08 PLANNED Complete demonstration below plus deterministic fault/regression matrix and independent verification of intended and unintended changes
10 Verified OBDX Pro GT read-only integration 09 plus documented device research/access PLANNED / HARDWARE GATE Source-backed device support matrix, explicit physical target, real read-only identification/diagnostics/telemetry recording where supported; no write path
11 Advanced dashboards, analysis, and comparisons 09; 10 for real captures PLANNED Persist configurable layouts/thresholds/units, linked charts, raw-data inspection, multi-run/multi-vehicle comparisons with lineage and uncertainty
12 Broader diagnostic protocols and hardware adapters 10 PLANNED Conformance-backed additional CAN/ISO-TP/UDS/KWP2000/J2534 or legacy adapters where verified; same application ports and capability restrictions
13 Curated vehicle knowledge and research workflow 08; 12 where needed PLANNED Sourced profile/module/calibration knowledge, review/promotion workflow, exact applicability, licensing/access controls; research cannot authorize writes
14 Higher-fidelity simulation and virtual module coverage 09; 13 for verified datasets PLANNED Validated model extensions, richer transmissions/environment/module interactions and reproducible fault coverage; published fidelity limits
15 AI evaluation and optional specialization 08, 09, 13 and adequate datasets PLANNED Held-out deterministic runbook evaluations, refusal/uncertainty cases, memory/latency measurements; LoRA/QLoRA only if justified by measured benefit
16 Programming qualification without physical writes 07, 09, 10, applicable 12/13 evidence PLANNED Module-specific safety case, bench plan, voltage/state/identity checks, interruption/recovery tests in simulation, legal authorization and defined go/no-go criteria
17 Explicitly authorized controlled bench programming 16 and separate user authorization PLANNED / AUTHORIZATION GATE Authorized supported bench target only; verified backup completeness, transfer/readback/recovery evidence and audit; no general vehicle-write enablement
18 Approved vehicle/module-specific programming 17 and separate vehicle-operation authorization PLANNED / AUTHORIZATION GATE Explicit per-operation approvals, compatibility/operating-state/power/security checks and proven recovery limits; strict module-specific support matrix
19 Packaging, extensibility, and additional platforms 05, 09; feature-specific gates remain PLANNED Offline-friendly reproducible releases, documented supported hosts, feature manifests/DI, compatibility/migration regression evidence
20 Optional MCP and firmware-emulation research 08, 13; isolated research approvals OPTIONAL / PLANNED Concrete benefit and separate ADR; restricted tools retain validator/approval boundary; emulation limitations and proprietary-data rights documented

Scope details for Steps 02-10

  • 01A accepted prerequisite: one thin MAUI Client, two RCLs, feature-organized target selection, actual headless rendered tests and worker native build/smoke evidence. All 54 original cases remain in the expanded 65-case suites.
  • 02: the regenerated assignment explicitly supersedes the earlier unexecuted version. Organize profile load/select/import/export as cohesive application slices with reusable Core contracts. Add only a thin profile metadata page in the existing UI, using those same services; no host restructuring or new engine. Retain vehicle specifications and configuration/tune selection, arbitrary module identities, supported/unsupported/unknown/inaccessible/restricted capability states, firmware/calibration identity, relationships, provenance, validation, schema versioning, and multiple synthetic variants. Never imply a diagnostic capture contains a complete firmware image.
  • 02A accepted correction: strict full-date/time/offset parsing, optional module evidence constraints, atomic import/UI regressions and aligned docs. Original reports remain unchanged; acceptance uses the corrected implementation.
  • 03: synthetic module state/storage, identification, telemetry definitions, DTC retrieval and simulated clearing, request/response semantics, addressing, session state where appropriate, delays, invalid requests, unsupported services, interruption and negative responses. Authentic protocol semantics require sources; use an explicitly synthetic versioned protocol first rather than inventing manufacturer behavior. Keep the existing profile schema intact; separate validated simulation endpoint/service definitions from profile metadata. A real Simulation provider is registered explicitly; update former empty-provider UI wording. No physical adapter or vehicle dynamics is implemented in this step.
  • 03A correction: reject incomplete/ambiguous frames before service dispatch, validate decoded replies before positive results/trace, and distinguish a clear acknowledgement from verified empty readback. Preserve actual virtual sessions, state isolation, static signals, controlled faults and previous regressions. Do not introduce dynamics, physical adapters or an arbitrary raw-frame UI.
  • 03B correction: preserve a validated positive probe and its advertised services when identity is unsupported, fails or is cancelled. Distinguish probe responders from identified modules; do not fabricate identity or relax existing identity-dependent action gates. Keep each wire exchange's truthful trace.
  • 04: start/stop/pause/resume/reset/step, configurable pacing independent of simulation results, save/load configuration, seeded scenarios, engine idle, throttle/load/torque/fuel, transmission/shift scheduling, mass/drivetrain/road load, temperatures, and a stationary-to-45-MPH-to-stop scenario with a five-minute hold. Bind model version/initial conditions/environment/scenario to explicit run identity, with versioned configuration import/export and no diagnostic-read driven physics. Preserve static v1 diagnostics and use an explicit versioned/ negotiated dynamic telemetry extension with separate model and diagnostic time. Report mathematical assumptions and repeatability tolerances.
  • 04A correction: reject invalid references through normal typed/UI failures, enforce initial and runtime drivetrain consistency rather than independent RPM clamping/false locked flags, and make progress/terminal failures/async ownership truthful. Add independent force/boundary checks and controlled pacing/lifecycle tests; preserve the existing warm/45-MPH/full-hold/stop numerical gates.
  • 04B correction: only the existing pending-owner overlap gap: validate Start/Resume eligibility before any lifecycle mutation while pause/stop/reset is still draining its old task. No physics, wire, schema or UI feature expansion.
  • 05: first bounded dashboard slice: reusable numeric/bar components bound to actual diagnostic signals, validated display configuration, explicit units and unavailable/stale/error states, with one source/session-aware acquisition owner. Supply a contract matrix for invalid ranges, missing signals and session replacement before coding. No charts, drag/drop or new clock/protocol/backend. Larger dashboard composition remains Step-11 scope. Client stays thin and no alternate host is introduced. Dashboard configuration is in-memory only for this slice; persistence/schema work is not bundled into the feed state machine.
  • 06: versioned recording format, module/source/units/time/quality metadata, raw plus decoded retention, interrupted/corrupt session handling, replay play/pause/stop/restart/seek/navigation/speed. Acquisition must not depend on visible gauges. Mark captured versus interpolated/reconstructed samples.
  • 07: scalar/table definitions with ranges/scaling/units and dependencies; original/proposed/tested/approved versions, structured diffs, verified backup completeness, approval binding, audit, readback, unrelated-data preservation, restoration where supported, explicit failure/recovery outcomes.
  • 08: local inference provider configuration, structured queries plus optional semantic retrieval, source-backed applicability checks, versioned runbook schema and restricted tools. Include natural-language diagnostics, driving scenarios, and calibration planning; AI cannot emit arbitrary hardware frames/offset writes. Missing knowledge and disconnected inference produce explicit outcomes without disabling non-AI functionality.
  • 09: equivalent baseline/modified scenarios and initial state, independent calibration oracle, behavioral/fuel/shift comparisons with confidence labels, fault reports, replay/raw-trace inspection and an auditable verification report. Include timeout, unsupported service, malformed response, wrong identity/version, interruption, corruption/checksum, partial write, reset, incompatible config, DTC and failed-verification cases.
  • 10: verify OBDX Pro GT protocols, drivers, OS/architecture constraints, manufacturer libraries/SDKs, actual J2534 support and C# integration using attributable sources before implementation. Missing device/driver access means real-hardware verification is blocked, not passed by mocks. Restrict services to verified read-only allowlists; diagnostic access does not imply every module is discoverable or programmable.

Longer-term scope preservation

Steps 11-20 retain richer gauges/clusters/charts and dashboard editing, comparison and analysis, more vehicle/module/protocol coverage, independently validated research, curated datasets, advanced dynamics and transmission modeling, local model evaluation/specialization, authorized programming, and supported release packaging. New MAUI platforms require platform-specific evidence. Additional hosts, including any Avalonia investigation, require a concrete use case and approval; none are currently required or to be scaffolded speculatively. Optional MCP does not bypass runbook safety. Firmware emulation is separate from the initial behavioral simulator, never an implied capability of synthetic modules.

Bench and vehicle programming milestones may not proceed on roadmap status alone. If authorization, applicable definitions, safe test equipment, licensing, recovery evidence, or manufacturer procedures are missing, retain a blocked gate and continue only independent safe work through an explicit new assignment.

Required synthetic end-to-end demonstration

Target acceptance: Step-09. Use the explicitly synthetic AutoForge test truck, initial warm idle 650 RPM and proposed warm idle 1,250 RPM. These values and any memory layout are synthetic, not manufacturer calibration data.

  1. Load the synthetic vehicle; inspect its modules and current calibration.
  2. Start simulation; display original idle telemetry; record and stop the run.
  3. Request the idle change in natural language. The local AI identifies the applicable module/parameter and produces a versioned logical runbook.
  4. Deterministically validate identity, definition, range/dependencies, execution target and approval; back up the original virtual calibration and verify it.
  5. Apply only the approved simulated change. Verify the operation, read calibration back, confirm the target value and independently confirm unrelated values.
  6. Re-run the same scenario/initial state using the modified calibration; observe the changed idle behavior and record the second run.
  7. Replay either recording; compare both runs, preserving their source, time, configuration, fidelity, and confidence metadata.
  8. Inspect underlying diagnostic communications and a report separating backup, programming, and functional verification with supported recovery limitations.

Evidence must cover all these actions without privileged application/AI access to the test oracle. Add a table-based shift-schedule case later; do not imply all calibration requests are scalar edits. Simulation fuel-use changes are estimates, not proven real-world economy improvements.

Test and evidence strategy

  • Worker selects maintained .NET 10-compatible test packages after inspecting the installed SDK and package compatibility. No speculative SDK/package signatures.
  • Unit tests cover invariants, validation, decoding, algorithms, and failure semantics. Integration tests exercise real composition across boundaries, serialization/persistence, transports, and execution as those features arrive.
  • Add simulation-specific test projects only when they have real responsibilities. Keep all tests in src/tests and separate privileged oracle access.
  • Use deterministic fixtures/seeds/virtual time; define reproducibility and numeric tolerance contracts. Keep test infrastructure independent of AI/cloud/hardware.
  • Adapter contract and hardware tests distinguish stub evidence from real devices. AI evaluations include held-out invalid/unsupported/uncertain requests and deterministic validators, not only LLM-as-judge.
  • Record exact commands, target framework/platform, configuration, exit outcomes, test counts and failures, and durable evidence references in each worker report.
  • Run the smallest relevant builds/tests plus required regression coverage. Missing prerequisites or unexecuted tests are BLOCKED/NOT EXECUTED, never PASSED.
  • Every assignment requires appropriate real unit/integration tests, documentation and README review. No empty "integration" tests or success-shaped placeholders.
  • RCL and application feature tests must run independently of MAUI. Native Client build/startup evidence remains separate; headless tests cannot prove BlazorWebView integration or satisfy a missing workload. Document exact native prerequisites and successful/failed/not-executed checks without architectural substitution.
  • The orchestrator inspects evidence without claiming independent runtime execution.

Decisions, completed work, remaining work, and blockers

Initial decisions

  • D-001: amendment layout and Step-XX naming supersede the original examples.
  • D-002: one .NET 10 solution; minimal headless Core/Application foundation first.
  • D-003 (SUPERSEDED 2026-10-28): the previous dual-shell/deferred-host decision is retained as history only. It is canceled by D-009/D-010 below.
  • D-004: common diagnostic/application interfaces across simulation, replay and physical providers; explicit execution target, no fallback.
  • D-005: deterministic simulation and independent verification are prerequisites for AI-assisted calibration execution and physical integration.
  • D-006: local-first replaceable inference; no initial fine-tuning/vector/MCP/cloud dependency. Storage implementations are chosen when real use cases arrive.
  • D-007: initial hardware is read-only; bench and vehicle writes have separate authorization/evidence gates and retain prohibited-operation restrictions.
  • D-008: handoffs remain manual; only the worker builds/tests and implements.
  • D-009: sole initial executable AutoForge.Client, .NET 10 MAUI Blazor Hybrid, initially Windows; complete UI in AutoForge.UI RCL and generic widgets in AutoForge.Components RCL. No required WebAssembly/web/backend executable.
  • D-010: modular monolith with application Vertical Slice Architecture; UI -> application contracts, Client -> composition; engines/protocols remain independent. Retain existing Application immediately and organize its actual use case by feature.
  • D-011: Step-01A corrects the accepted foundation before Step-02. Do not rewrite the final Step-02 until 01A is implemented, reviewed and accepted; regenerate it using actual results and explicitly supersede the old unexecuted assignment.
  • D-012: with 01A accepted, Step-02 adds a minimal metadata-only profile consumer to the existing UI alongside headless feature services. This advances one cohesive profile slice, not the later simulator/dashboard milestone.
  • D-013: profile timestamps must contain their full date/time and explicit offset, never ambient date/time-zone defaults. A null optional module applicability identity is an absent constraint, not a requirement that the target identity also be null; supplied constraints remain validated and grant no authority.
  • D-014: Step-03 introduces Diagnostics and Simulation as purposeful separate libraries, composed by the existing Client. Diagnostic application features use provider-neutral ports; virtual state/oracles do not leak into Application/UI. Initial telemetry is explicitly static synthetic module state, not vehicle physics.
  • D-015: Step-04 keeps the nine-project architecture. Simulation owns physics and clock implementation; Core/shared contracts expose only needed run inputs and controls. UI/Application read vehicle telemetry through diagnostics, not the oracle.
  • D-016: dynamics uses an explicitly selected provider/configuration, separate from the existing static provider. Run identity binds the exact configuration/model; reset preserves that binding. Dynamic wire/origin/time changes are versioned explicitly rather than weakening static protocol v1 or relabeling physics as static.
  • D-017: fixed virtual steps determine physics; wall pacing, UI refresh and diagnostic request counts cannot change results at a given model time. Full acquisition/ recording/dashboard work remains assigned to later milestones.

The worker records the relevant accepted decisions as technical ADRs when creating or changing the implementation. Future changes require a reason and roadmap/state history, not repeated redesign without evidence.

Completed

  • Read both supplied instruction sets and inspected the existing repository.
  • Established the initial planned boundaries, sequence, and first worker assignment.
  • Accepted Step-01 after reviewing the assignment, report, all production/test source, project/configuration files, root-file diffs and documentation.
  • Confirmed immutable target descriptors, explicit/no-fallback resolution, eager DI validation, snapshot isolation, and the intended four-project dependency graph.
  • Reviewed worker evidence: .NET SDK 10.0.402, successful Release build with zero warnings/errors, 35 unit plus 19 integration tests passed, zero failed/skipped. Source case counts and existing restore/Release outputs corroborated the report; the orchestrator did not run application builds or tests.
  • Prepared Step-02 under the prior architecture; it was not executed. It was held during correction and is now superseded by the regenerated version.
  • At correction planning, inspected the then-four-project solution: there were no obsolete hosts to delete; the missing UI foundation and obsolete plans/docs were assigned rather than discarding useful code.
  • Adopted the new single-shell/RCL/vertical-slice decisions and prepared Step-01A.
  • Accepted Step-01A after inspecting shell bootstrap, RCL route/component ownership, feature code, all affected tests, project/package graph, documentation and ADRs. Worker evidence records successful seven-project Release/native builds with zero warnings/errors, 39 unit + 26 integration cases passed, zero failed/skipped, and actual Windows home/selection/not-found and RCL asset smoke observations.
  • Static source counts, seven-project acyclic references, restored assets and native/portable Release outputs corroborate the report. The orchestrator did not execute application builds/tests or native startup.
  • Regenerated Step-02 against that accepted implementation, with explicit supersession, feature/RCL boundaries, regressions and worker verification duties.
  • Reviewed the completed Step-02: immutable Core metadata, strict object/schema parsing, profile catalog/DI, UI, fixtures, new tests, Git changes and documentation. Worker evidence reports 59 unit + 34 integration cases passed, zero failed/skipped, zero-warning/error solution/native builds and actual profile UI smoke.
  • Source counts, restored assets, Release outputs and fixture shape corroborate those reported checks, not independent runtime execution. Static review found two untested validation defects; Step-02 was not accepted at that initial review.
  • Prepared Step-02A. Original Step-02 assignment/report remain unchanged; the worker was required to correct and report before a Step-03 handoff.
  • Reviewed and accepted Step-02A and corrected Step-02: strict timestamp lexical/ calendar/offset checks, null-as-absent evidence constraints, retained mismatch/ unknown-target failures, actual direct/codec/DI/UI regressions and documentation.
  • Reviewed worker-reported pre-fix failures, corrected focused runs, sequential zero-warning/error complete Release build, 103 unit + 37 integration passes, zero failed/skipped and native positive/error/original-fixture smoke. Source case counts, seven-project graph and existing outputs corroborate the report. The orchestrator did not run builds/tests or launch the native app.
  • Prepared Step-03 for manual handoff; no Step-02B is required.
  • Reviewed the completed Step-03 protocol/transport, definitions/session state, application/UI, oracle/tests, dependency graph, docs and report. Worker evidence records 108 unit + 53 integration passes (161 total, zero failed/skipped), clean complete nine-project/native Release build and actual native workflow.
  • Static source counts, project graph, restore/Release/native outputs corroborate the reported normal-flow checks, not independent runtime execution.
  • Found three uncovered correctness gaps and prepared Step-03A. At that review Step-03 was not accepted and Step-04 was withheld; the original report remains preserved.
  • Reviewed Step-03A source/tests/docs and actual reported pre-fix failures: R-01/R-02/R-03 are closed. Required/unique fields, semantic validation before positive trace, and the Application clear-and-verify rule now satisfy that scope.
  • Worker reports sequential nine-project Release build with zero warnings/errors, 111 unit + 64 integration passes (175 total, zero failed/skipped) and native normal-flow smoke. Injected failed-clear evidence is correctly identified as real DI/rendered integration, not native evidence.
  • Static source counts, graph and existing outputs corroborate the report. Review also found R-04: positive probe evidence was lost when the subsequent identity read was unavailable. Prepared Step-03B; Step-04 was withheld pending that correction.
  • Reviewed and accepted Step-03B and corrected Step-03: independent probe/identity outcomes, retained services, accurate expected/probed/identified counts, partial failures/cancellation, fail-closed actions and raw exchange truthfulness.
  • Worker reports actual pre-fix failure, corrected real-DI/rendered matrix, complete nine-project Release build with zero warnings/errors, 111 unit + 72 integration passes (183 total, zero failed/skipped) and native baseline smoke. Partial-identity fault coverage is explicitly rendered real-DI, not native.
  • Source counts, graph and existing restore/Release/native artifacts corroborate that evidence; the orchestrator did not run builds/tests or launch the app.
  • Prepared Step-04; no Step-03C is required.
  • Reviewed completed Step-04 configuration/model/controller/provider, v2 integration, UI, tests, documentation and worker report. Worker evidence records a clean nine-project Release build, 115 unit + 75 integration passes (190 total, zero failed/skipped), numerical baseline metrics and rebuilt native smoke.
  • The warmed default reports 3,361 reproducible samples, 10 s idle, 10.3 s acceleration, all 3,000 cruise samples inside the required band, 5.8 s braking and 10 s return idle. These are worker results, not orchestrator test execution.
  • Source counts, graph and existing outputs corroborate the reported run. Acceptance was withheld at that review for the input/kinematic/lifecycle findings; prepared Step-04A without changing implementation or original evidence.
  • Reviewed Step-04A's reference/coupling/progress/failure/controller/UI corrections, independent equations and controlled scheduler tests. Worker reports 122 unit + 80 integration passes (202 total, zero failed/skipped), clean full Release build, unchanged numerical gates and native invalid-reference/fault/baseline smoke.
  • R-01/R-02 and the progress/failure/reentrancy fixes satisfy their inspected scope. One existing R-03 overlap remains: Start/Resume mutates the engine before StartRunnerLocked rejects an old task that has not drained.
  • Prepared a single-issue Step-04B. Also reassessed assignment sizing, frozen contract matrices and worker self-review; added the prospective shared quality gate rather than adding new criteria to Step-04A.
  • Reviewed and accepted Step-04B against C-01 through C-05: Start/Resume/scenario start all check owner eligibility before lifecycle mutation, rejected commands preserve full status/sample state, cleanup can finish and one retry starts once.
  • Worker reports actual three-transition pre-fix state failures, corrected controlled unit/real UI-service tests, explicit self-review, complete Release build zero warnings/errors and 126 unit + 81 integration passes (207 total, zero failed/skipped). Native baseline is distinct from controlled race proof.
  • Source counts, graph and existing outputs corroborate that evidence; the orchestrator ran no application builds/tests/native commands.
  • Accepted corrected Step-04 with Step-04A/04B and prepared bounded Step-05 using the shared quality gate and frozen data/feed matrix.

Remaining

  • Worker execution of Step-05, then one consolidated review against its frozen contract matrix and shared quality gate. Complete Step-05A only for genuine unmet requirements/regressions before Step-06; do not expand criteria in review.
  • Profile/module metadata, synthetic virtual-module communication and deterministic dynamics/v2 are accepted. Dashboard and later capabilities remain planned. Preserve the complete sequence/safety gates.

Known blockers and uncertainties

No blocking Step-04 deficiency remains. The owner-drain finding below is historical and closed by accepted Step-04B:

  • StartOrResumeAsync changes engine lifecycle to Running before checking whether the old advancement task has finished. During pending pause/stop/reset cleanup, StartRunnerLocked can throw "already active" after that mutation. The old cancelled task then exits, leaving a rejected command's Running state without an owner. Existing tests await transition completion before starting/resuming, so they do not cover the interleaving.

This was within Step-04A's existing concurrent lifecycle/no-stale-owner criterion, not a retrospective quality-gate requirement. Step-04B supplied the controlled pre-fix reproduction and validated owner eligibility before mutation.

The original Step-04 findings below are retained as history; R-01/R-02 and the other R-03 corrections have been inspected as fixed:

  • Step-04 R-01: missing profile/module references can escape configuration import as catalog/lookup exceptions that the simulation import UI does not handle.
  • Step-04 R-02: accepted initial RPM/speed/gear inputs can contradict a locked drivetrain. Runtime RPM is independently capped while reporting zero slip; supported launch settings can also report zero slip despite uncoupled RPM.
  • Step-04 R-03: cumulative progress is cleared at phase boundaries; faulted scenarios and subsequent reads can publish success status; pending scenario cancellation ownership and late async results need guards and controlled proof.

Step-04A supplied before/after, independent force/ratio, controlled pacing and UI/native evidence. The orchestrator did not run application code or reproduce these vectors. The new shared exit gate applies only to future work.

No Step-03 correction deficiency has reopened. R-04 below is historical and closed by Step-03B, alongside Step-03A's R-01/R-02/R-03:

  • A module can return a valid positive probe but not support read.identity, or its identity request can fail. The discovery client replaces the successful probe with the identity result and discards its advertised services. DiscoveryResult.Responders then counts only non-null identities, and UI calls that count "actual ... endpoint(s) responded". A reachable endpoint is therefore reported as not responding. The corrected result now preserves both phase outcomes, retained probe services, explicit unavailable identity and separate counts without relaxing identity-dependent action guards.

The worker supplied an actual pre-fix reproduction and corrected coverage for the valid probe-only and identity-failure cases. The orchestrator did not run them.

The original Step-03 R-01/R-02/R-03 findings below are historical and closed:

  • R-01: missing request version is silently defaulted to 1, missing logical tick can become zero, and duplicate dispatch properties are not rejected before deserialization. An invalid/ambiguous frame can reach virtual service dispatch.
  • R-02: positive service payloads are not fully validated (for example negative signal ticks or blank identifiers/units), and a mismatched probe module is detected only after its trace has already recorded PositiveResponse.
  • R-03: the UI treats any successful DTC readback as proof of clearing, including readback that still contains codes. A positive acknowledgement is not verified state change.

Step-03A supplied actual pre-fix reproductions, corrected wire/service/UI regressions, documented contracts and native baseline smoke. The orchestrator did not execute those cases.

The original Step-02 findings below are historical and resolved by Step-02A:

  • R-01 originally used permissive DateTimeOffset parsing. The replacement requires complete date/time/seconds and explicit Z/numeric offset, validates its calendar and offset and preserves exporter precision without ambient defaults.
  • R-02 originally required both evidence and target identities null. The corrected helper treats an absent constraint as absent, while every supplied constraint still requires a matching known target identity.

The worker supplied actual pre-fix failures and post-fix unit/DI/UI/native evidence. Neither before/after input case was executed by the orchestrator.

The Step-01A report records SDK 10.0.402, runtime 10.0.13, Windows target 10.0.26100.0/win-x64, MAUI Windows manifest 10.0.110 and WebView2 155.0.4283.62. Native execution is worker evidence, not an orchestrator rerun. Other platforms, inference and automotive hardware remain unverified.

Non-blocking deviations are retained explicitly:

  • The worker mistakenly invoked workload restore during inspection, reported manifest updates (workload version 10.0.403 and iOS/MacCatalyst changes), stopped the action and did not roll back. Required Windows manifest remained 10.0.110; final Windows builds/tests/smoke passed according to the report. Broader environment impact is unverified. This is not retroactive authorization. Do not repeat workload-changing commands or attempt rollback without approval.
  • Step-02 resolved the working README CRLF issue: inspection found 149 LF and zero CRLF lines. A command-local CR-at-EOL-aware diff check passes. No global Git setting was changed; the existing LF-to-CRLF notice remains non-blocking.
  • The pending-transition rejection policy is now documented. Step-05 must align remaining acceptance-pending lifecycle text while preserving the original and corrective numerical/native reports, including the 207-case Step-04B run.

No major environment installation, update, repair or rollback is automatically authorized. Physical definitions/access and programming approval remain future gates.