Skip to content

docs: NexusFeature strategy notes and Master Plan v2 - #212

Closed
Daliys wants to merge 5 commits into
developmentfrom
docs/nexus-feature-notes
Closed

Daliys wants to merge 5 commits into
developmentfrom
docs/nexus-feature-notes

Conversation

@Daliys

@Daliys Daliys commented Sep 30, 2026 •

Copy link
Copy Markdown
Member

Summary

Docs-only. Adds planning material under NexusFeature~/ (not compiled, not shipped in the package):

  • NexusFeature~/ notes: MASTER_PLAN v1, ImageCapturePlan, benchmark v1 protocol/results/harnesses, AI session reports, Unity CLI video transcript, social feedback screenshots (from the first commit on this branch).
  • NexusFeature~/MASTER_PLAN_V2.md (v2.2) — the new active plan. Nexus becomes a live-truth, verification and context layer on top of Unity CLI / Pipeline instead of competing on transport, tool count or raw speed.
    • Flagship "Live truth": nexus_snapshot, nexus_run, nexus_compile, nexus_look. nexus_verify is conditional on a competitor spike.
    • Pre-registered gates at week 4 (adoption) and week 9 (verify against the best competitor, Windows, external users).
    • Evidence-linked facts with expiry, verified code defects, GLOBAL RULES v2 for agent-run tasks.
    • §14b maps every planned item to what Unity ships now and is likely to ship next: use / wrap / build / avoid, plus integration hooks.
    • §15 is the adversarial review log.
  • NexusFeature~/v2_analysis/: the source reports.
    • Four independent analyses.
    • Two adversarial challenges and the judge log.
    • Two Unity research reports.

Key findings recorded in the plan:

  • v1 benchmark ok% counted protocol success rather than task success.
  • ~100 ms latency steps come from Editor update throttling.
  • Phase 0 code defects: unauthenticated shutdown_server, timed-out main-thread calls still executing, false-ready Move validation workflow to self-hosted Mac runner #13, fixed sleeps in the bridge compile wait, open-ended Pipeline versionDefines. The last one is urgent because Pipeline 0.8 moved [CliCommand] to Unity.Pipeline.Attributes.

No package source, tests, public docs or CI/linter configs are changed.

Validation

  • Package compiles in Unity. n/a, no code changes.
  • Relevant Editor tests pass. n/a. The Unity package smoke crash on this head is a pre-existing batch-mode ReadScreenPixel abort in development (see PR comments).
  • NexusQualityGate passes or remaining warnings are intentional. n/a.
  • Public/protected C# types and methods have useful XML documentation. n/a.
  • Python bridge still compiles. n/a, bridge untouched.
  • Public docs updated for behavior/API changes. n/a, no behavior change.
  • CHANGELOG.md updated for public changes. n/a, internal planning docs only.

Pending: the owner's signed commit excluding NexusFeature~ from ruff and the code linter. The archived benchmark scripts are kept as they were run.

Safety checklist

  • Server remains loopback/local-only. (unchanged)
  • File operations remain constrained to the Unity project root. (unchanged)
  • No generated caches, Python bytecode, .DS_Store, Unity Library/, or local agent artifacts included. The folder contains benchmark result JSON and screenshots by design; the owners decide whether NexusFeature~/ stays in the repo long-term.
  • No secrets, credentials, private project source, or proprietary assets included. Auth token values are not in the result files; paths reference the owner's local machine.

🤖 Generated with Claude Code

https://claude.ai/code/session_01H6f4omZLV6CPXCnvhPr6Am

Daliys and others added 2 commits September 29, 2026 18:35
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Master Plan v2 supersedes MASTER_PLAN.md v1 and the forward-looking parts of
ImageCapturePlan.md: Nexus as the verification and context layer on top of
Unity CLI / Pipeline, with honest benchmark v2, adoption gates and a 12-week
roadmap. v2_analysis/ keeps the four source reports (build-on-Unity, skeptic,
agent efficiency, architecture/roadmap).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01H6f4omZLV6CPXCnvhPr6Am

Daliys commented Sep 30, 2026

Copy link
Copy Markdown
Member Author

CI blocker — needs a maintainer decision

  • Python Quality Gate / Code Linter (head 46a00d7): failed on the archived benchmark scripts in NexusFeature~/bench/ and NexusFeature~/bench_nexus.py (34 ruff findings, 6 files unformatted, one 65-line main). The plan documents themselves are Markdown and not linted.
  • I tried to fix this by excluding NexusFeature~ in .ruff.toml (eb8762b) and .code-linter.json (f4d728f). That was the wrong fix for this repo: the Code Linter's policy signature guard rejects both commits because linter configs are protected and require a signature from the code owner. Unsigned agent changes to them are blocked by design, so I am not going to work around it.

Options (owner to choose):

  1. Drop eb8762b and f4d728f from the branch (force-push of agent-only commits) and make the same one-line exclusions yourself as a signed commit. This keeps the evidence scripts byte-for-byte as they were run.
  2. Drop those two commits and let the agent make the bench scripts lint-clean instead (ruff fixes + split main). This touches only non-protected files, but it changes the archived scripts.
  3. Drop those two commits and remove the bench scripts (or all of NexusFeature~/) from this PR, keeping only the plan documents.

No re-run requested: the failure is deterministic until one of these is chosen.


Generated by Claude Code

Two independent challengers (technical, strategy) attacked v2.0; an independent
judge accepted 37, partially accepted 13 and rejected 2 of 52 objections and
applied them (Live-truth flagship, conditional verify, pre-registered gates,
resumable jobs, underscore command names, security additions). Adds the
challenger reports, judge log and Unity shipped-inventory research to v2_analysis/.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01H6f4omZLV6CPXCnvhPr6Am
Fills §14b with what Unity ships now and is likely to ship next (capability map,
integration hooks, watch list). Revises F4 (Pipeline reportedly has runtime-only
simulate_key/pointer), adds F14 (Pipeline 0.8 moved attributes to
Unity.Pipeline.Attributes, P0-06 now urgent) and F15; compile truth via
unity recompile, Safe-Mode-resilient diagnose, JUnit output for verify;
drops nexus_mutate and raw profiler work. Adds the Unity roadmap report.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01H6f4omZLV6CPXCnvhPr6Am
@Daliys
Daliys force-pushed the docs/nexus-feature-notes branch from 5163051 to f37da0e Compare September 30, 2026 06:50

Daliys commented Sep 30, 2026

Copy link
Copy Markdown
Member Author

Option 1 chosen: eb8762b and f4d728f are removed from the branch (force-push, head now f37da0e). .ruff.toml and .code-linter.json are identical to development again. Python Quality Gate / Code Linter will stay red on NexusFeature~/ scripts until the owner's signed exclusion commit lands.


Generated by Claude Code

Daliys commented Sep 30, 2026

Copy link
Copy Markdown
Member Author

CI status on f37da0e, for the record:

  • Unity Quality Gate and Documentation quality AI failed at 07:03, both after exactly ~10 min, with empty check output and no downloadable logs (HTTP 404). Unity package smoke and Python unit tests have been queued since 06:53. This looks like the self-hosted runners being unavailable (jobs never started), not this PR: the diff is Markdown under NexusFeature~/ only, with no C# or public-doc changes. I can't re-run jobs from here; they need a re-run once the runners are back.
  • Python Quality Gate / Code Linter: expected red until the signed NexusFeature~ exclusion lands (see above).

Generated by Claude Code

Daliys commented Sep 30, 2026

Copy link
Copy Markdown
Member Author

Unity package smoke on f37da0e did run this time and failed with a native Unity crash (Abort trap: 6) during the EditMode tests. It is not caused by this PR (Markdown under NexusFeature~/ only). It is a package bug in code already on development:

AgentToolingTests.UiCaptureWindowSnapshotReturnsRectHierarchyAndBestEffortImage → ui_capture_window_snapshot → EditorWindowPixelCapture.TryReadSurfacePixels → InternalEditorUtility.ReadScreenPixel → NSBitmapImageRep initWithCGImage throws an Objective-C exception → abort. The C# try/catch cannot catch it. The only guard is InternalEditorUtility.isApplicationActive (Editor/Capture/EditorWindowPixelCapture.cs:60), which can be true in -batchmode on the persistent macOS runner, so this will fail intermittently depending on runner focus.

Proposed fix (separate PR, not widening this one), before the isApplicationActive check:

if (Application.isBatchMode)
{
    NexusEditorLog.Warning(NexusLogCategory.UiAutomation,
        $"[MCP_SCREENSHOT] {windowName} surface capture skipped: no screen surface in batch mode.");
    return null;
}

The test already accepts PartialSuccess with a message, so it would pass. No fix exists yet; I can't re-run the job from here.


Generated by Claude Code

Adds F16-F19 (Unity removing classic domain reload / 6.0 EOL, MCP 2026-07-28,
Unity ToS agentic-access clause, package footprint), rewrites Principle 3 for
optional reloads, adds P0-13 version/CoreCLR spike, P0-14 headless spike,
P0-15 footprint fix, S-09 opt-in hook pack, static-state check V-08,
privacy/ToS rules, decisions G12-G14, risks R17-R21 and §16. Adds the
blind-spot report to v2_analysis/.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01H6f4omZLV6CPXCnvhPr6Am
@Daliys Daliys closed this Oct 1, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants