ODS V3 Pre-Release: Public testing and refinement ahead of the official V3 launch. Turn your PC, Mac, or Linux box into a private AI server.
Pixel enhancement priorities for ODS
Reviewed September 7, 2026 against the current Pixel issues, ODS PR #3385, and native ODS work on Tower1, Tower2, Tower3, Strixy and an 8 GB VRAM laptop. This is an adoption plan, not a claim that the proposed features are installed or qualified. Pixel remains a core ODS feature; the other applications remain available alongside it. PR #3385 subsequently merged on 2026-09-24. That merge does not establish completion of every proposed capability or user-acceptance requirement.
The issue numbers below identify historical proposals in the private Pixel repository. They are plain reference identifiers, not public evidence links. Public readers should use the acceptance criteria here and the current ODS Portal guide.
Recommended order
| Priority | Pixel proposal | Value for ODS | First useful increment and acceptance evidence |
|---|---|---|---|
| First | #201: artifact-backed context (private proposal) | Reduce repeated large outputs in the prompt while retaining the complete evidence needed to finish and recover work. Tower3 reached a context threshold while repeatedly reading HTML during a real CSV repair. | Store complete tool output once; return a readable excerpt and a session-scoped retrieval handle with bounded selectors. Complete a real multi-step repair across context compaction and restart, retrieving the exact retained source and preserving the original task and failed checks. Measure context use, latency and peak memory on the laptop and larger hosts. |
| First, alongside context | #203: exact model provenance and capability-aware routing (private proposal) | Distinguish the configured label from the provider, model, backend and settings that actually served each request. This is necessary to diagnose slow runs and memory/output limits accurately. | Record the resolved route and observed reasoning/context/output settings per call, then use measured hardware fit for defaults. Exercise an owner-selected model change and an explicit fallback; show the actual serving identity and preserve the memory envelope and owner overrides. Keep ordinary model choice open. |
| Next | #199: stale-safe edits (private proposal) | Make focused edits dependable when a formatter, another agent or the owner changes a file after Pixel reads it. This is preventive protection; the inbox concurrency experiment does not prove this particular defect. | Add a read digest and an atomic compare-and-apply operation using Pixel's own edit format. Demonstrate an intervening edit, zero mutation on rejection, fresh reread and successful retry. Cover Unicode, LF/CRLF, repeated text and cancellation. A separate hash check followed by an unlocked write is insufficient. |
| Next, where supported | #200: fresh post-edit diagnostics (private proposal) | Give prompt feedback on syntax, types and unresolved imports without asking the model to rediscover them through repeated reads. | Bind diagnostics to the edited file version and return a concise delta. Demonstrate a broken edit, fresh diagnostics, correction and project tests. A missing or stale language server must not be reported as a clean result. Keep debugging/execution permissions distinct from diagnostics. |
| Incremental foundation | #198: engine-neutral executor contract (private proposal) | Reduce coupling to one engine's tool envelopes, progress events, cancellation and recovery behavior. Engine identity is distinct from inference-model identity. | Specify and test the lifecycle seam around an existing working route. Verify actual work, terminal delivery, cancellation and resume through it. Artifact retrieval, edit integrity and current-route fixes can proceed independently; a multi-engine rewrite is not a prerequisite. |
| After lifecycle and edit integrity | #202: isolated subagent patch capsules (private proposal) | Let Pixel use parallel workers without shared mutable checkouts or unreviewed integration. | Start with one child, an isolated workspace, explicit resource ownership and a patch returned to the parent. Verify independent work, stale-patch rejection, parent cancellation, descendant cleanup and reviewed integration. Do not infer child authority from model prose. |
| Optional measured canary | #204: advisory shadow reviewer (private proposal) | May catch repeated low-value actions or wrong hypotheses before a long run expires. The laptop's repeated generator rewrites are a useful evaluation case, not proof a reviewer would fix them. | Use a separate, bounded context at sparse checkpoints. Compare matched tasks with and without review: unique useful findings, completion quality, latency, tokens and primary-model disruption. Advice remains visible and cannot grant permissions or control execution. It must remain optional on constrained hardware. |
Existing repairs to qualify before adding replacements
- #209 (private proposal) describes duplicate embedded/CLI compaction and a lost response. ODS already contains a scoped, exact-version completion-recovery adaptation documented in runtime repairs. Verify its installed coverage and truthful continuation. Tower3's September 7 context-overflow event was followed by fresh tool calls, so it is evidence of context pressure and ongoing recovery, not a reproduced duplicate-compaction race or a proven lost request.
- #140 (private proposal) documents an
offreasoning value becoming truthy. PR #3385 carries defensive configuration. Measure actual requests and visible output; a slow laptop run alone does not prove thinking was accidentally enabled. Do not increase limits to conceal a protocol, tool-call or generated-code defect. - #139 (private proposal) and #141 (private proposal) concern execution surviving cancellation. Existing ODS cancellation mechanisms need descendant and restart qualification on each supported surface before expanding child jobs.
- #223 (private proposal) is useful for a genuinely read-only audit mode: retain capable inspection while enforcing the selected write restriction mechanically. It must not silently narrow the owner's ordinary or Full Access modes.
What these proposals do not fix automatically
The generated inbox application's temporary-file collisions and ledger races need application-level locking and atomic writes. Stale-safe source editing does not repair that application algorithm. Language-server diagnostics do not prove decimal totals, document layout, preview publication or working controls. An advisory model cannot replace accurate tool results or deterministic permission enforcement.
Keep live workflows as acceptance: create and revise an app; import/export real synthetic files; recover after interrupted work; manage an ODS extension; inspect an authorized host or LAN peer; and finish with a result the owner can use. Rotate tasks and wording. Compare each mechanism against the existing route under the same model and memory settings. Record OS, GPU/backend, model, effective limits, failures, latency and memory separately; five busy sessions do not establish five qualified installations.
Design provenance and original implementation
This review and plan are original. No OMP source code, prompts, patch grammar, tests or assets were imported. The Pixel proposals already distinguish their own contracts from the systems that informed them.
Conceptual sources reviewed at OMP revision
eea5628f13043286e17c4a2ea4fc28b15fda33ca include
artifact storage,
compaction,
and the issue-linked documentation for
Hashline,
RPC,
tasks
and advisory review.
The pinned OMP license
is MIT and names Mario Zechner, Can Bölük and Stencil Labs, Inc.
Implement these concepts through original Pixel/ODS interfaces and tests. Retain precise conceptual-source links where they influenced the design. Existing bundled third-party materials continue to carry their actual notices; the upstream attribution snapshot remains unchanged. Conceptual attribution must not be used to imply incorporation of code or endorsement by another project.