[Compatibility]: Qualify OMP through canonical autogenesis Skill + stdio MCP

Client and support state

Client: OMP
Qualified Preview path: OMP 18.18.3 on native macOS arm64
Windows: pending final-artifact acceptance on a native Windows interactive host; not passed and not supported yet
Compatibility seam: one canonical autogenesis Agent Skill plus the existing AutoGenesis Runtime over standard stdio MCP

User outcome

A user installs the canonical Skill, keeps their existing MCP configuration intact, and asks OMP to perform a real AutoGenesis task. OMP discovers the existing stdio Runtime tools, uses its native approval UI for protected calls, observes the real application result, and cleans up task-owned state.

Observable conformance journey

Scenario: Discover the canonical Skill and preserve MCP configuration
  Given a clean supported host with OMP and the Preview artifacts
  And the client may already contain unrelated MCP entries
  When the user installs AutoGenesis through the documented Agent Skills route
  And starts or reloads OMP through its native lifecycle
  Then OMP discovers exactly the canonical autogenesis Skill
  And the existing AutoGenesis stdio MCP entry loads
  And unrelated MCP configuration remains byte-equivalent
  And a second unchanged install or activation reuses the working state without duplicates

Scenario: Execute one real direct task
  Given OMP has discovered the canonical autogenesis Skill and loaded the existing Runtime
  When the user asks OMP to calculate 2 + 3 in Calculator on macOS or the approved Notepad task on Windows
  Then live AutoGenesis tools with callable input schemas are available through stdio MCP
  And OMP presents its native approval UI for protected calls
  And the Agent performs the task in the real application
  And the required result is observed rather than inferred from tool success
  And sessions, applications, managed processes, and task-owned temporary state are cleaned up
  And no plaintext secret is written to configuration, logs, prompts, generated content, or evidence

Common acceptance

Canonical autogenesis Skill discovery through Agent Skills; no autogenesis-setup or autogenesis-run Skill.
Existing unrelated MCP configuration is preserved.
The existing AutoGenesis stdio Runtime exposes live callable tools.
Protected calls use OMP's native approval UI; no parallel approval controller.
One real direct task completes with an observed application result.
Cleanup is bounded and observable.
Re-running install/activation is idempotent and does not duplicate or overwrite working configuration.
Evidence is sanitized and secret-safe.

Exact BDD Scenario execution remains a supported Skill route over the same Runtime. Explicit recording may produce a read-only preview. Generated-file write or commit is unavailable in this Preview and is not an acceptance requirement.

Accepted native macOS evidence

Reuse the accepted OMP 18.18.3 evidence; do not repeat it solely for this issue:

Final PR head e187985a50f215141c6ce59849ad4e7cf4afd06c and Contract 2.1.0 / 3768d673273fc82609a94a218a355622233e10892a9f2392688711aa4572d32c.
Installed canonical autogenesis Skill and existing task-owned MCP entry loaded.
31 callable live schemas were available.
OMP's native approval UI was used.
A real Calculator Mac2 session clicked Two → Add → Three → Equals; an intermediate assertion observed 2, the final assertion observed 5, and page evidence recorded 2+3.
app_close, session_close, managed OMP/Appium shutdown, Calculator absence, and closed loopback ports proved cleanup.
Accepted evidence: [Prompt]: Guide installation and environment readiness from user goals #65 (comment)

The separate macos: keys WebDriverAgent code-65 result remains a recorded transport limitation; the qualified semantic-click journey does not relabel it as passed.

Remaining acceptance and dependency

The macOS path is qualified. The issue remains open only for the native Windows final-artifact journey.

Dependency: final Preview artifacts and a native Windows interactive host.
Dependency owner: Human Product Owner provides/approves the host and final artifacts.
Execution owner: independent compatibility acceptor records exact OMP, Windows, Python, Runtime, Windows driver/runtime, and artifact identities and runs the same conformance journey with a real Notepad result.
Stage: Blocked until that host and final artifact are available; Windows must not inherit the macOS pass.

Explicit non-goals

No OMP-specific Skill, Runtime projection, compatibility adapter, or client patch.
No MCP prompt/resource or serverInfo exposure gate unless the direct journey actually depends on it.
No generated-file commit path in the current Preview.
No reinstall or repetition of already accepted macOS evidence merely to change issue wording.

Parent Epic: #25. Prerequisites: #18, #19, and #20.

Native macOS OMP 18.0.5 compatibility evidence from Issue #65 / PR #66 final Candidate e187985a50f215141c6ce59849ad4e7cf4afd06c:

Isolated minimal OMP inference passed: exact fixed token, exit 0, 10.6s.
Native OMP MCP connection test reports autogenesis-issue65 connected with 31 tools.
Native discovery advertises all three official AutoGenesis prompt commands.
Native resources list contains autogenesis://agent-contract/identity; the actual Agent read that URI and returned matching Contract 2.1.0 / 3768d673273fc82609a94a218a355622233e10892a9f2392688711aa4572d32c.
Final RPC v2 collector correlated and completed /mcp reload, list, test, resources, and prompts; native exit 0.

Remaining compatibility blockers:

40 bounded post-reload get_state polls did not expose documented dynamic mcp__<server>_<tool> names or live parameter schemas, despite connection test tool count 31.
Structured prompt command metadata supplied only generic arguments; canonical prompt invocation returned a sanitized native extension operation error and no expanded workflow receipt.
initialize serverInfo name/version was not observable in retained native client output.

This is Changes Requested / incomplete qualification, not a Runtime defect or blanket OMP unsupported conclusion. No AutoGenesis business tool, Appium server, Mac2 session, application, or user configuration was changed. Full evidence is attached to Issue #65 and PR #66; OMP client qualification remains blocked here.

Parallel isolated compatibility diagnosis update (native macOS 26.7.2 arm64, OMP 18.0.5; repository read-only):

Native connection remains healthy: one task server, 31 tools.
Root cause of the earlier missing dumpTools schemas: OMP defaults to tools.xdev: true, which presents the 31 MCP tools as xd:// mounts rather than flat mcp__... tools. A task-only tools.xdev: false overlay exposes all 31 live parameter schemas, including action, verification and generation schemas.
Root cause of the earlier prompt-get failure: get_available_commands.input.hint == "arguments" is generic UI text, not an MCP prompt argument name. Actual prompt metadata requires task_text for execute_natural_language_task; scenario prompts use scenario_text, with feature_file_path additionally required for generation.
Correct native invocation /autogenesis-issue21:execute_natural_language_task task_text="DIAG metadata only; no execution" retrieved the expanded canonical user message. Native and returned receipts matched at 4,830 bytes, SHA-256 0a04abfc6e722f77517a2504bbecd6234cc9e427ef63ec93bc32a06d660a2ef3.
OMP receives and retains wire-reported connection.serverInfo (appium-mcp-server / 1.30.0), but native RPC state/headless test output does not export it. This is the remaining observability gap; the installed TUI code presents it but was not exercised.

No shared Contract, Skill or Runtime change is indicated. Zero business tools executed; prompt turns were aborted. User config/credentials, repository and GitHub state were unchanged by the diagnosing agent. Full #21 qualification remains pending native initialize-metadata evidence and the remaining acceptance journeys.

Product Owner phase-two decision: OMP is the priority client. #21 is unblocked for the next narrow native acceptance step using the existing isolated workspace and diagnosis evidence. First prove the remaining initialize-metadata observability through OMP's documented native TUI/status surface; reuse the already valid 31-schema, canonical prompt-get and matching identity receipts. Do not reinstall, repeat valid evidence, run Calculator/Mac2 yet, modify shared Contract/Skill/Runtime, or weaken acceptance. If metadata proof passes, return with the exact protected-effect plan required for the later real macOS journey.

Narrow native initialize-metadata acceptance result: Blocked / Changes Requested.

OMP launched with the task server connected; /status rendered the native dashboard. Bounded PTY navigation did not reach the MCP inspector, and a distinct documented native TUI /mcp test attempt produced no metadata-bearing output. The exact unresolved boundary is PTY input → native command dispatch/inspector navigation. This does not prove that the TUI lacks the fields and does not indicate a Runtime change.

Historical wire metadata remains appium-mcp-server / 1.30.0; separately verified AutoGenesis package identity remains autogenesis-appium-mcp 0.5.0. They were not relabeled as new native TUI proof. Existing 31-schema, canonical prompt-get and matching Contract identity receipts remain valid and were not repeated.

Stopped without further retries. No credentials, business tools, applications, repository or GitHub source changes were used; task processes and disposable TUI state were cleaned. Real Mac2/Calculator acceptance remains gated. #21 returns to Blocked pending an approved OMP client-side observability route or an explicit acceptance-contract decision.

Product Owner approved protected-effect plan issue21-macos-mac2-01 to unblock the already-authorized Calculator smoke.

Authorized effect: in the existing task-owned environment only, run appium driver install mac2@4.5.1 --json with HOME bound to /tmp/autogenesis-issue65-macos-acceptance/task/omp-home. This may write only task-owned .appium and npm state and run package lifecycle scripts. No sudo, global Appium mutation, user HOME/config copy, OMP/AutoGenesis reinstall or source change.

Candidate metadata checked before approval: appium-mac2-driver 4.5.1; registry integrity sha512-tlvVAIMNzDsx1KFRgACg2X6Xx72D0ISyLVNNKhHP89k5yIzwA7AcnNokcYbEhFdGSmwyjJLVaHjNPod8sZi5sw==; current public Microsoft npm feed; peer Appium ^3.0.0-rc.2; engines Node ^20.19 || ^22.12 || >=24, npm >=10; current Node 26.12.1, npm 11.20.0, Appium 3.7.0 satisfy them.

After install: verify exact installed inventory, run appium driver doctor mac2, require zero required fixes. If doctor, privacy permission, server smoke or session prerequisite fails, stop before repair and request a new effect. If Ready, continue exactly one Mac2 Calculator 2+3=5 session and cleanup. Rollback is appium driver uninstall mac2 under the same task HOME; do not apply rollback unless requested or needed after failed installation.

Plan issue21-macos-mac2-01: installation passed; Runtime target configuration now Action Required.

Official mac2@4.5.1 installed at 2026-11-09 15:21:18 UTC in the approved task-owned HOME. Registry, installed-lock and cached-tarball SHA-512 match the approved integrity. Exact installed inventory contains only mac2@4.5.1. appium driver doctor mac2 reports 0 required fixes; its one optional Full Disk Access warning concerns native video retrieval, which this smoke does not use.

The unchanged OMP MCP command still omits --platform mac, so AutoGenesis Runtime selects its default ios profile. The packaged fallback Mac profile also has empty bundleId. Retargeting exceeds the approved .appium/npm write scope, so the acceptor stopped before Appium/MCP start or session creation. Calculator was not touched and result 5 remains unobserved. Final task process count: 0 at 15:24:15 UTC. Successful Mac2 installation is retained.

Read-only Interface verification: Runtime supports --platform mac and --config <explicit-json>; ConfigManager accepts a task-owned config containing a mac profile; the installed Calculator bundle ID is com.apple.calculator. A separate approval is required to create that task-owned profile and atomically update only the task-owned OMP MCP entry/args, with rollback to the current config bytes.

Product Owner approved protected-effect plan issue21-macos-config-01.

Authorized writes are limited to task-owned acceptance state: create one explicit JSON Mac profile outside the business project/repository, with existing Runtime conventions platformName: Mac, automationName: Mac2, verified bundleId: com.apple.calculator, reset disabled, existing timeout/quiescence defaults, and server_url: http://127.0.0.1:4723; atomically update only /tmp/autogenesis-issue65-macos-acceptance/task/omp/mcp.json args to add --platform mac --config <absolute-task-profile> while preserving the approved executable and project root. Record exact before/after bytes and SHA-256; validate JSON and task-root scope before replacement; restore original bytes on write/validation/reload failure.

No business-project, user HOME/config, privacy, repository, package, OMP/Runtime version or shared state change is authorized. After OMP native reload and matching config evidence, start only task-owned loopback Appium/MCP processes, create exactly one Mac2 Calculator session, perform visible 2+3, observe 5, then close session/processes. Any permission, port, server, launch or session prerequisite failure requires stop-before-repair and a new decision.

Plan issue21-macos-config-01: stopped before writes due to external OMP version drift.

The native launcher now reports OMP 18.18.3; the previously qualified/approved 18.0.5 location is unavailable. The acceptor did not upgrade OMP. Because the compensating initialize-identity exception and existing client receipts were explicitly version-scoped, they were not transferred silently.

No profile creation, MCP config update or reload occurred. Exact original task-owned OMP MCP bytes were backed up and remain unchanged; before/after SHA-256 6fbc8300d997f9ff9ac416ac2eea8490a59f54e1d10f27cb76097f130aca2667. Appium/MCP starts: 0; sessions: 0; Calculator untouched; result unobserved. Final task process count 0 at 2026-11-09 15:28:36 UTC. Mac2 4.5.1 installation remains retained; repository/GitHub source unchanged.

Required Product Owner decision: either restore exact OMP 18.0.5 or explicitly change the qualification baseline to 18.18.3 and rerun the minimum client compatibility receipts before applying the already-approved task-owned Mac configuration.

Product Owner changed the native OMP qualification baseline to current stable 18.18.3; plan issue21-omp-18.18.3-01 supersedes only the client version baseline. Do not reuse 18.0.5 compatibility conclusions as passes.

Ordered thin requalification, before any Mac profile/client write: record exact native binary/version; isolated fixed-token inference with MCP disabled; current task server connected; 31 live schemas using the existing task-only flat-exposure setting; successful canonical prompt-get using actual metadata arguments; actual native resource read matching Contract 2.1.0 / 3768d673273fc82609a94a218a355622233e10892a9f2392688711aa4572d32c. No business tool, Appium server, Mac2 session or Calculator interaction during these gates. Any failed gate stops without substitution.

Only if all compensating receipts pass, the accepted initialize-serverInfo observability exception is extended specifically to OMP 18.18.3. Then apply the already-approved issue21-macos-config-01 task-owned profile/MCP-args change and continue the already-approved single Mac2 Calculator 2+3=5 smoke. Existing Mac2 4.5.1 installation remains valid. No OMP downgrade, package reinstall, user config/privacy change, repository edit, commit, push or merge is authorized.

Calculator smoke: task-only approval-controller correction blocked by independent security review; no launch occurred.

OMP 18.18.3 gates 4 and 5 passed before this boundary: canonical prompt expansion 4,992 bytes with abort acknowledged and zero tools; actual native resource read matched Contract 2.1.0 and required manifest. The documented initialize-serverInfo exception is now compensated specifically for OMP 18.18.3. Native Mac config reload/effective config, Appium 3.7.0, Mac2 4.5.1 and loopback readiness also passed.

The first controller revision was not executed. Independent review found:

High confidence/high severity: after policy failure/abort is latched, queued in-phase requests can remain approval-eligible while awaiting acknowledgment, violating failure-stop/no-retry.
High confidence/medium severity: the one-launch reservation file and replacement do not fsync the containing directory, so a system crash can lose the marker preventing another launch grant.

Stopped per plan before further repair. Live approvals 0; MCP launch dispatches 0; Mac2 sessions 0; Calculator untouched; authorized launch remains unconsumed. Cleanup verified zero task processes and closed ports 4723/10100 at 2026-11-09 16:03:30 UTC. Repository, GitHub source, user config/privacy and accepted task Mac configuration unchanged.

A fresh Product Owner decision is required to repair both task-only harness findings, rerun independent security review, and only then consume the already-approved single launch.

Product Owner approved task-only harness plan issue21-controller-fix-02.

Authorized changes are limited to /tmp acceptance-controller artifacts: make approval eligibility fail closed immediately after any policy failure or pending abort; separate the immutable one-launch reservation from mutable response status; durably commit reservation and every atomic replacement by fsyncing the file and containing directory. Preserve the one-launch cap and reject any approval when reservation state is missing, ambiguous or stale.

Required order: deterministic event-order simulations including queued-after-failure; crash-point/restart simulations around reservation create/replace/response; focused independent security review of exact before/after controller artifacts. Only if every check and review passes may the controller consume the existing unspent approval for one native Calculator launch. Any finding stops without another repair. No product/repository/user config change, commit, push or GitHub mutation by the acceptor.

Native macOS Calculator smoke remains Blocked: stop after two task-controller security-review iterations.

Plan issue21-controller-fix-02 addressed the original post-failure approval and reservation durability findings. Deterministic validation passed 12 event-order cases, 11 invalid/stale-state cases and 26 crash/restart checkpoints. Fresh independent review nevertheless found one remaining medium-severity/high-confidence path: a business-turn timeout raises before fail_closed() latches failure; cleanup then calls state/drain while task-phase approval eligibility can remain active, so a queued business approval could be granted after timeout.

Per the explicit stop condition and no-repeat-method rule, no third one-off repair is attempted. This now requires a separately designed, independently reviewed acceptance harness rather than another temporary controller patch.

No real approval, MCP launch dispatch or Mac2 session occurred across either revision; the authorized launch remains unspent. Calculator was untouched. Zero task processes and closed ports 4723/10100 were verified at 2026-11-09 16:14:39 UTC. Configuration, repository, GitHub source and user privacy/config remain unchanged from the accepted task-owned Mac configuration.

Preserved passed evidence: OMP 18.18.3 fixed-token inference; connected task MCP; 31 live schemas; native TUI prompt argument metadata; canonical prompt expansion 4,992 bytes with zero tools; actual matching Contract resource identity; compensated serverInfo exception for 18.18.3; task Mac profile/native reload; Appium 3.7.0; Mac2 4.5.1 with zero required doctor fixes; loopback readiness. Calculator 2+3=5 remains unexecuted/unproved.

Product Owner scope correction for the current delivery: acceptance is the direct user journey only—installed canonical autogenesis Skill guides the Coding Agent to load the existing AutoGenesis MCP entry; the client discovers callable AutoGenesis tools; the user requests Calculator 2+3; the Coding Agent uses its native approval UI; the real result 5 is observed and task-owned resources are cleaned. OMP serverInfo, MCP prompt/resource exposure and a custom acceptance approval controller are diagnostics/non-goals, not gates. Do not build or repair a parallel approval state machine; use OMP's native lifecycle and UI. Pi remains unsupported while stock Pi has no MCP. This decision supersedes the current-phase gate language in prior compatibility comments without relabeling unexecuted behavior as passed.

OMP 18.18.3 direct macOS journey: Changes Requested — Mac2 input crashes WebDriverAgent.

Executed the Product Owner's corrected direct journey only, using the existing accepted task-owned OMP MCP entry/profile and OMP native approval UI. No custom approval controller, prompt/resource/serverInfo gate, reinstall, alternate client, or source modification.

Observed:

Existing OMP entry loaded; 31 callable AutoGenesis tools remained available.
Native approval UI granted app_launch; a real Mac2 session launched Calculator and returned its native accessibility summary.
An unnecessary command+1 attempt first produced macos: keys → socket hang up → WebDriverAgent host exit code 65.
One fresh Mac2 session then launched successfully. The normal unmodified payload {"keys":["2","+","3","="]} independently produced the same socket hangup and WebDriverAgent code-65 exit after 71 ms.
Input therefore did not complete and visible result 5 was not observed. This localizes the current blocker below OMP Skill routing, MCP discovery, native approval UI, and application launch: Mac2/WebDriverAgent fails on its standard key operation in the tested native macOS/Xcode environment.
No further retry, WDA rebuild/signing/privacy change, alternate UI route, or substitute acceptance was attempted.

Cleanup passed through the same MCP: app_close success and session_close success. Managed Appium exited 0; isolated OMP exited after its cleanup response. Repository, user config/privacy, packages, and normal OMP state were unchanged.

Full direct-journey evidence is recorded on #65: #65 (comment). #21 remains incomplete; Calculator 2+3=5 is not passed.

Navigation Menu

Repository navigation

[Compatibility]: Qualify OMP through canonical autogenesis Skill + stdio MCP #21

Client and support state

User outcome

Observable conformance journey

Common acceptance

Accepted native macOS evidence

Remaining acceptance and dependency

Explicit non-goals

6 remaining items

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

[Compatibility]: Qualify OMP through canonical autogenesis Skill + stdio MCP #21

Description

Client and support state

User outcome

Observable conformance journey

Common acceptance

Accepted native macOS evidence

Remaining acceptance and dependency

Explicit non-goals

Activity

houlianpi commented on Oct 29, 2026

houlianpi commented on Nov 9, 2026

houlianpi commented on Nov 9, 2026

houlianpi commented on Nov 9, 2026

houlianpi commented on Nov 9, 2026

6 remaining items

houlianpi commented on Nov 9, 2026

houlianpi commented on Nov 9, 2026

houlianpi commented on Nov 9, 2026

houlianpi commented on Nov 9, 2026

houlianpi commented on Nov 9, 2026

houlianpi commented on Nov 9, 2026

houlianpi commented on Nov 9, 2026

houlianpi commented on Nov 9, 2026

houlianpi commented on Nov 10, 2026

houlianpi commented on Nov 10, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions