Repository navigation
feat(research): automate deterministic CAD decoding and evidence - #158
jdharcourt wants to merge 47 commits into
Conversation
|
Automatic reviews are disabled for this repository. Configure in Code Review Settings. |
|
/devin review |
|
root owns exactly this is the requested guidance/change-entry update on the existing automation PR; no other code path is claimed. Root will publish one normal commit and verify its new-head CI independently of prior-head results; no native dispatch, qualification, extra review request, queue or merge authority is implied current verified head is |
Diff intelligence · update #2 · 193,428 changed lines · 18,900 reviewablePR #158 ·
+193,393 additions · −35 deletions · 661 files · 74 binary files
The bar follows the table order. Shares use additions + deletions and reconcile to 100% when text changes exist. Binary files have no fabricated line counts. Review focus18,900 reviewable lines; 174,528 lines excluded by configured policy. This metric is not a guarantee of low risk.
Classification: 661 files matched deterministic rules or Git metadata; 0 unclassified files remain reviewable. Rule matches express configuration confidence, not semantic or security certainty. Composition changesPrevious Net snapshot change: +11 total lines, +2 reviewable lines, -0.0 percentage points excluded.
What changed in this pushActual head-to-head diff: +11 / −0 across 2 files (0 binary).
Net PR snapshot changes are not the lines edited during this push. Largest changed files
Unclassified, renamed, deleted, binary and metadata-only files
Reviewable size over recent pushes
Base |
|
The requested decoding-challenge instruction is committed verbatim to AGENTS.md in Mandatory decoding automation change entry Exact source comparison, Markdown formatting, local link resolution, whitespace and the two-path repository policy check passed. No local tests, collection, pipeline jobs, SDK or native applications were executed. The new head's PR checks are rerunning; the earlier |
|
Published A07 and A06b to PR #158 under the user's new explicit push instruction. Current head: The update adds four task commits and39 files over previous head A07 commits retain their identities:
A06b was integrated without rewriting its original branch:
Both original A06b commits are also published on Recorded verification remains scoped to its original source: A07 focused process/native lifecycle evidence and A06b process/decoder/admission checks are retained; current combined remote CI is a separate result. The39-file exact-blob integration check and diff whitespace check passed. Publishing these fixes does not establish independent saved automatic policy, recover the missing original Windows raw custody, or qualify any new cross-format pair. A07 remains parked and A06b qualification remains not evaluated/zero qualified pairs. No historical capture, receipt, source pin or P02/P05 reservation was rewritten to claim otherwise. The full PR remains subject to its existing CI, configured review and native acceptance requirements. Earlier passing checks or reviews do not validate this new head. Full-scope reviewer timing remains deferred while the already documented independent native acceptance requirements are open. |
Deterministic-code follow-up ownership planAfter the user's requested five-minute wait, the isolated integration checkout was fast-forwarded to current PR head Initial partition (audit/read-only until a concrete defect and exact edit claim are posted):
Existing SOLIDWORKS/Fusion native owners retain all application sessions and research claims. The A15a frozen source, runtime, private store, observation window and prior evidence remain unchanged; this coding assignment does not wait for the two-week pilot. P02/P05 reservations, original custody and historical receipt identities remain intact. Missing raw captures or native policy proof will not be replaced by invented evidence, weaker matching, direct ledger edits or qualification promotion. Exit condition: finish genuinely actionable deterministic code with real process-level evidence, formatting/policy checks and coherent publication; distinguish concrete external research/review prerequisites from code still needing implementation. No merge, queue enrollment, model research calls or review bypass is authorized. Additional existing-file ownership will be posted before edits. |
Deterministic follow-up — exact repair ownership before editsThe new coding goal is active at unchanged head Engine agent — A07d direct dependency-result binding
A consistently fingerprinted ordinary dependency selector currently admits bytes whose SHA differs from that dependency's accepted result. Bind that direct selector to its result pin. Preserve the distinct artifact/candidate selector semantics; verify a real controller-to-worker positive dependency chain and refusal before publication. Evidence agent — semantic worker session validation
The current comparison worker publishes artifacts for re-fingerprinted missing/undeclared/malformed dependency snapshots. Reuse the shared session validator owned by the engine agent; do not duplicate or edit that shared file. Retain comparison semantics and all unqualified/native-not-established outcomes. Verification agent — isolated Git provenance at startup/readback
The real installed bootstrap starts normally but fails with inherited GIT_DIR. Isolate inherited Git controls in startup and downstream committed-byte reads, retaining exact-source, dirty-tree, replacement/graft and installation checks. Use real subprocess/installed-bootstrap regression coverage; no licensed CAD is needed. Root — Windows disposable-clone cleanup and integration
All seven admission assertions pass but teardown fails on read-only Git objects. Retry only the task-owned checkout's read-only file removal after resolved-root containment validation; unrelated errors must remain failures. Retain actual process coverage and sandbox-refusal provenance. The initial independent 71-case process audit has passing observations after narrowly permitted sandbox retries. The new39-file incoming range is UTF-8 Python/Markdown only, with no vendor software addition. Root will independently review the Git repair; agents will cross-review other fixes so each patch has a non-author verifier. No overlap with existing native owners, frozen pilot source/runtime/store, shared register ownership, historical receipt identities or P02/P05 reservations. No qualification promotion or gate bypass. |
|
Deterministic follow-up checkpoint (ownership unchanged)
The refreshed remote head remains |
|
Scope refinement before edits: fresh GitHub review-thread inventory also found two existing static-analysis findings in |
Diff intelligence · update #18 · 327,876 changed lines · 60,314 reviewablePR #158 ·
+326,555 additions · −1,321 deletions · 2,248 files · 83 binary files
The bar follows the table order. Shares use additions + deletions and reconcile to 100% when text changes exist. Binary files have no fabricated line counts. Review focus60,314 reviewable lines; 267,562 lines excluded by configured policy. This metric is not a guarantee of low risk.
Classification: 2,247 files matched deterministic rules or Git metadata; 1 unclassified files remain reviewable. Rule matches express configuration confidence, not semantic or security certainty. Composition changesPrevious Net snapshot change: +1,045 total lines, +366 reviewable lines, -0.1 percentage points excluded.
What changed in this pushActual head-to-head diff: +1,102 / −57 across 23 files (0 binary).
Net PR snapshot changes are not the lines edited during this push. Largest changed files
Unclassified, renamed, deleted, binary and metadata-only files
Reviewable size over recent pushes
Base |
|
Published all deterministic follow-up work to PR 158: accepted head Reproducible source-bound report Root independently verified 61 focused process cases: inference 24/24, installed bootstrap 14/14, semantic sessions plus both controller/reader artifact chains 16/16, and Fusion admission/Windows clone teardown 7/7. The agent's broader comparison run had 113 passes and seven pre-submission sandbox binding refusals; every refused case passed the permission-enabled reruns. Non-author source/AST reviews cover every repair. All 13 changed Python blobs match the actual verified bytes; their sorted SHA-256 manifest is The two existing FreeCAD assertion-side-effect findings are addressed in Formatting, focused lint/steering, Mypy, rationale, full index path policy and relative documentation links passed. Focused Pyright reported zero source errors with its three existing unsupported-config notices retained. Full incoming PR history plus index vendor-DLL/DMG guard passed; all 23 follow-up files are UTF-8 Python/Markdown text with no embedded executable/archive signatures. Fresh remote CI is now starting; earlier green checks do not validate this accepted head. Existing independent native evidence, configured review approval and merge-group acceptance remain separate requirements. Full-scope reviewer timing remains deferred while native prerequisites are outstanding. Frozen A15a state, parked jobs, native owners and historical custody remain untouched; research model calls remain off. No merge or queue enrollment. |
|
Final deterministic coding verification at unchanged accepted head
The scoped deterministic coding goal is fulfilled: four published commits include all three agents' work, 61 independently exercised focused process cases, the reproducible source-bound report, and verified absence of vendor DLLs. This does not complete or restart the separate two-week pilot observation or native owners' research. No merge or queue enrollment was performed. |
Diff intelligence · update #19 · 328,044 changed lines · 60,482 reviewablePR #158 ·
+326,639 additions · −1,405 deletions · 2,249 files · 83 binary files
The bar follows the table order. Shares use additions + deletions and reconcile to 100% when text changes exist. Binary files have no fabricated line counts. Review focus60,482 reviewable lines; 267,562 lines excluded by configured policy. This metric is not a guarantee of low risk.
Classification: 2,247 files matched deterministic rules or Git metadata; 2 unclassified files remain reviewable. Rule matches express configuration confidence, not semantic or security certainty. Composition changesPrevious Net snapshot change: +168 total lines, +168 reviewable lines, -0.0 percentage points excluded.
What changed in this pushActual head-to-head diff: +84 / −84 across 1 files (0 binary).
Net PR snapshot changes are not the lines edited during this push. Largest changed files
Unclassified, renamed, deleted, binary and metadata-only files
Reviewable size over recent pushes
Base |
Diff intelligence · update #20 · 328,044 changed lines · 60,482 reviewablePR #158 ·
+326,639 additions · −1,405 deletions · 2,249 files · 83 binary files
The bar follows the table order. Shares use additions + deletions and reconcile to 100% when text changes exist. Binary files have no fabricated line counts. Review focus60,482 reviewable lines; 267,562 lines excluded by configured policy. This metric is not a guarantee of low risk.
Classification: 2,247 files matched deterministic rules or Git metadata; 2 unclassified files remain reviewable. Rule matches express configuration confidence, not semantic or security certainty. Composition changesPrevious Net snapshot change: +0 total lines, +0 reviewable lines, +0.0 percentage points excluded.
The merge base changed; snapshot deltas also reflect the updated base. What changed in this pushActual head-to-head diff: +410 / −17 across 7 files (0 binary).
Net PR snapshot changes are not the lines edited during this push. Largest changed files
Unclassified, renamed, deleted, binary and metadata-only files
Reviewable size over recent pushes
Base |
|
@odinglyn0 ops? |
Diff intelligence · update #21 · 328,424 changed lines · 60,795 reviewablePR #158 ·
+326,951 additions · −1,473 deletions · 2,254 files · 83 binary files
The bar follows the table order. Shares use additions + deletions and reconcile to 100% when text changes exist. Binary files have no fabricated line counts. Review focus60,795 reviewable lines; 267,629 lines excluded by configured policy. This metric is not a guarantee of low risk.
Classification: 2,252 files matched deterministic rules or Git metadata; 2 unclassified files remain reviewable. Rule matches express configuration confidence, not semantic or security certainty. Composition changesPrevious Net snapshot change: +380 total lines, +313 reviewable lines, -0.1 percentage points excluded.
What changed in this pushActual head-to-head diff: +323 / −79 across 12 files (0 binary).
Net PR snapshot changes are not the lines edited during this push. Largest changed files
Unclassified, renamed, deleted, binary and metadata-only files
Reviewable size over recent pushes
Base |
Diff intelligence · update #22 · 328,829 changed lines · 60,795 reviewablePR #158 ·
+327,355 additions · −1,474 deletions · 2,256 files · 83 binary files
The bar follows the table order. Shares use additions + deletions and reconcile to 100% when text changes exist. Binary files have no fabricated line counts. Review focus60,795 reviewable lines; 268,034 lines excluded by configured policy. This metric is not a guarantee of low risk.
Classification: 2,254 files matched deterministic rules or Git metadata; 2 unclassified files remain reviewable. Rule matches express configuration confidence, not semantic or security certainty. Composition changesPrevious Net snapshot change: +405 total lines, +0 reviewable lines, +0.0 percentage points excluded.
What changed in this pushActual head-to-head diff: +404 / −1 across 3 files (0 binary).
Net PR snapshot changes are not the lines edited during this push. Largest changed files
Unclassified, renamed, deleted, binary and metadata-only files
Reviewable size over recent pushes
Base |
Diff intelligence · update #23 · 330,070 changed lines · 61,214 reviewablePR #158 ·
+328,596 additions · −1,474 deletions · 2,264 files · 83 binary files
The bar follows the table order. Shares use additions + deletions and reconcile to 100% when text changes exist. Binary files have no fabricated line counts. Review focus61,214 reviewable lines; 268,856 lines excluded by configured policy. This metric is not a guarantee of low risk.
Classification: 2,262 files matched deterministic rules or Git metadata; 2 unclassified files remain reviewable. Rule matches express configuration confidence, not semantic or security certainty. Composition changesPrevious Net snapshot change: +1,241 total lines, +419 reviewable lines, -0.1 percentage points excluded.
What changed in this pushActual head-to-head diff: +1,286 / −45 across 21 files (0 binary).
Net PR snapshot changes are not the lines edited during this push. Largest changed files
Unclassified, renamed, deleted, binary and metadata-only files
Reviewable size over recent pushes
Base |
Diff intelligence · update #24 · 330,317 changed lines · 61,346 reviewablePR #158 ·
+328,843 additions · −1,474 deletions · 2,266 files · 83 binary files
The bar follows the table order. Shares use additions + deletions and reconcile to 100% when text changes exist. Binary files have no fabricated line counts. Review focus61,346 reviewable lines; 268,971 lines excluded by configured policy. This metric is not a guarantee of low risk.
Classification: 2,264 files matched deterministic rules or Git metadata; 2 unclassified files remain reviewable. Rule matches express configuration confidence, not semantic or security certainty. Composition changesPrevious Net snapshot change: +247 total lines, +132 reviewable lines, -0.0 percentage points excluded.
What changed in this pushActual head-to-head diff: +253 / −6 across 8 files (0 binary).
Net PR snapshot changes are not the lines edited during this push. Largest changed files
Unclassified, renamed, deleted, binary and metadata-only files
Reviewable size over recent pushes
Base |
Change
Automate CAD decoding challenges through the research pipeline: supported capture, durable jobs and immutable artifacts, application-worker ownership and recovery, deterministic semantic comparison and inference, affected-file rescans and traceable evidence.
The feature includes the restored SOLIDWORKS and Fusion decoding, native inference and verification, writer research, and FreeCAD shape/edit-history work. Decoder rules must come from recorded evidence and deterministic validation; a matching pattern is a hypothesis rather than a proven rule. AGENTS.md requires use of the supported automation and explicit reporting of missing capabilities.
The A01–A16 register records each milestone's actual scope and remaining acceptance. Including this work does not establish complete native writing, lossless conversion or qualified cross-format routes. Models remain off. Original corpus/capture identities, held-out reservations, receipts and historical release authority remain unchanged. Original PR #94 is closed and marked superseded by #158. Its branch, full original description and historical evidence remain preserved; its old vendor-DLL commit ancestry is not inherited by this clean-main-based branch.
Documentation
Review-call contracts and source verification
Native test handoff and source review
Directory-layout repair
Programme and current step register
Research controller
Supervisor and resource ownership
Deterministic inference
Original split and restored feature scope
CI phase boundaries
Capture history binding
Mandatory decoding automation
Clean-history lineage gotcha
Git identity gotcha
Validation and testing handoff
Current source is
a6cc279f629d088c526a654149bb9dc99a8b2bb6, retaining all fifteen restored phase commits plus the scoped test-preparation and review-contract repairs. Earlier green checks ata920bf7dand4bad1bf9do not validate these later changes.The directory-density failure is repaired: both affected directories now contain 32 direct tracked files. Focused moves preserve every collected test ID, assertion, source hash and historical receipt identity. The initial test-preparation path check passed 2071 changed destinations, and the review repair passed its four-path check; binary-history, rationale, focused typing/formatting and documentation-link/JSON checks passed. Current-head path policy, CI Agg, Ruff, Black, rationale, actionlint and Zizmor passed. The separate Code Quality run also completed successfully on this exact head. Applicable non-review PR checks are green; heavy queue acceptance remains deferred.
The user has no local Fusion or SOLIDWORKS and explicitly requested stopping before licensed application testing or merging. The Windows SOLIDWORKS-machine test handoff is prepared from the actual retained commands and guards. Replace every
REVIEWED_PR_HEADplaceholder witha6cc279f629d088c526a654149bb9dc99a8b2bb6before running it;00ea6bberemains only the historical command-review baseline. The Windows supervisor uses its direct research entry point; the installed CLI does not support that operation on Windows. The retained SOLIDWORKS readback worker requires build34.4.1.1011. A SOLIDWORKS-only machine cannot prove Fusion-specific startup/restart or native acceptance.No local tests, collection, SDK, generators or native applications have run for this preparation. Heavy Linux tests, platform process/OOM witnesses, security/type gates and the completeness report remain in the native merge queue. No queue enrollment, auto-merge or merge was performed.
Remaining acceptance
Independent Fusion capture source admission, real owned application exit/restart with accepted fresh-phase evidence, native qualification and the two-week pilot remain open. Source restoration does not supply missing original handoff objects or authorize the pending historical 25-member CI batch.
Both configured review dashboards verified the earlier
1e87ddf32359e8b0186024871f6c31e3fed363d0size-limit failure, covered by the user's size-only waiver. The requested rereview ona6cc279f629d088c526a654149bb9dc99a8b2bb6reports failure without a detailed current-head comment; no passing reviewer report or approval is inferred. Independently enforced GitHub approvals and queue requirements remain separate from test readiness.Current ownership and testing-stop coordination
Contributions are provided under the MIT License.
Review repair checkpoint
Commit
a6cc279f629d088c526a654149bb9dc99a8b2bb6addresses the three inline static-analysis findings with explicit report arities, separate constructor/renderer refusal calls and explicit validated dimension selection. All six preview refusal controls, the Boss guard, report order, counterexamples and error handling are retained. All three threads have accepted-fix replies and are resolved; rereview was requested with the configured exact comment. No licensed or local tests were run. Current-head Code Quality completed successfully for all three languages. A fresh review-comment inventory after completion found no new source findings, and all three original threads remain resolved. Managed analysis reports a 300-file diff API limit; successful job completion does not establish full heavy-test or native acceptance.