Enhance Claude-dispatched OmniRoute workers with native MCP, agents and workflows

[seathatflowsinourveins]

Scope

Claude's foundation coordinator needs native OmniRoute workers with task-selected skills, MCP tools, bounded agents and repeatable workflow execution. This adds typed readiness discovery and a scoped enhancement kit to the retained official Codex SDK worker. Sol/Max remains the primary worker; one native Astra/Max judge handles the consequential acceptance check. Native coordinator sessions keep their accounts. The Claude dispatcher defaults to the enhanced scoped home and gates the model turn on native readiness.

• Lane: lane:foundation.
• Owned implementation: SDK examples/tests, a Claude dispatcher skill, scoped worker configuration/role, Dagu graph, sanitized receipts and convergence records. Evidence registrations are additive.
• Native configuration, supported upstream SDK/tool execution and existing installers are reused. Live owners retain shared policy/catalog/checkpoint paths.

SOTA sources

• Codex SDK/native runtime 0.159.2, ff6aec96948b70d94983af2641a6b67c94faeff5: public typed app-server discovery, lifecycle, native tools and child configuration. Spawn ordering requires leaving gateway-prefixed spawn defaults unset so the native role can apply its alias.
• Context Mode 1.0.169, 589d8214d56740a28b5f7bf63167743d586b0b40: maintained stdio entry and output processing.
• Serena selected revision, c6fbd1c5932df2494ffa0020af5a9fbe80b82143: native stdio integration and selected read tools.
• Dagu 2.16.6, 58fed633d58c1dd1319091fdb2c2f6158ecfa053: native graph/dependency/deadline format and root environment imports.
• OmniRoute 3.8.51 upstream fixture, c1e30b7676975feb298b49eff6ff58923c04b89e; existing running gateway/translated-bridge source remains distinct.
• Vercel skills 1.7.0, 7407f3893ad4dceab546ac002c3ef806e4000c73: existing supported task-selected installer.
• mitmproxy 12.2.3, 6c09d56e4c29a92f5ad01b03199977584b8ea14f: unchanged streaming reverse observer; RTK v0.50.0 supplies native output/proxy practice.
• Claude native Skills and the repository's pinned native workflow/role recipe provide the coordinator callsite. The separate Claude SDK 0.2.162 bridge remains an unqualified trial.
• OSV-Scanner 2.6.0 supplies the supported UV extractor for the three native SDK script locks.

Evidence-class table

Claim	Evidence class	Artifact
Typed native config/MCP/skill discovery, compact output and missing-capability rejection	native execution; locally authored integration checks	12 checks passed before compact CLI changes; four focused checks passed afterward
Selected skill body read; Context Mode count and Serena symbol calls completed	native model execution	evidence/receipts/omniroute-runtime-enhancements-20260930.json
One native Astra/Max judge through OmniRoute; unchanged rtk npm test once, exit0; fixture hashes unchanged	native model execution + independent observation	Same receipt and retained oracle output
Native Dagu readiness precedes worker; final graph succeeded and local SDK cleanup closed	native execution	Same receipt; environment failure and 300-second timeout preserved
Earlier native Claude Opus5.5/Max dispatcher → Sol/Max SDK callsite; lifecycle/resume/recovery	historical native model execution	Existing callsite/runtime receipts, tied to archived exact pre-enhancement source
Three native SDK locks parsed and scanned clean with supported UV override	native execution	Existing tools/runtime-worker-evidence/osv-check.json

The kit keeps skills and schemas on demand, native caching/compaction intact, selected MCP tools, bounded child concurrency, private scoped state and RTK/context instructions. Hook support and optional services retain their own execution gates. The Dagu graph is manually invoked; no schedule or scheduler was installed.

Local commands run

rtk uv run --locked --script examples/omniroute-codex-sdk/test_worker.py
# 12 checks passed before compact output changes; four relevant checks passed afterward.
rtk proxy dagu validate --context local --dagu-home <owned-state> examples/omniroute-codex-sdk/runtime-worker.yaml
# exit0
rtk proxy dagu start --context local --dagu-home <owned-state> --run-id <fresh-id> examples/omniroute-codex-sdk/runtime-worker.yaml
# supplied six task bindings; repaired run succeeded in about186 seconds
rtk proxy python3 scripts/validate.py
# exit0:69components/8555hashes/4profiles/179receipts; integrity/scope only
rtk proxy python3 scripts/validate_convergence.py blueprints/convergence-practice/omniroute-runtime-workers/experiment.json blueprints/convergence-practice/omniroute-runtime-enhancements/experiment.json --json
# exit0:17 historical observations,16 enhancement observations; consistency only
rtk git diff --check
# exit0

Astra independently accepted the original returned artifacts and source-backed repairs. Successful native cumulative usage is 153,813 for the Sol parent and 66,044 for its Astra child, counted separately. Cache and reasoning counters are subsets; the observer corroborates those same totals. Nine unclassified flow-error records accompany completed streams and remain retained. Whole-task provider usage, billing, backend attestation and token savings are unknown. The failed300-second attempt retains its own partial parent and completed-child counters.

Decision record

examples/omniroute-codex-sdk/enhancements.md, the new receipt and scoped experiment document acceptance and repairs. The separate Claude/Devin bridge still has HTTP500/timeouts and zero qualified tool execution. No universal SDK/SOTA winner claim. Earlier unchanged upstream tests remain historical; the prior Codex formatter-driver failure remains recorded.

The existing full OSV gate reported findings in unchanged OpenHands and historical macOS Next.js locks; owners received the findings. No exemption or historical source rewrite. Current-head CI remains separate from local acceptance.

Host evidence

The 23 owned enhancement paths and additive registrations are synchronized into the shared checkout; full repository and both scoped convergence checks pass there. New owned private worker/Dagu state only. No platform-status or evidence/hosts/ promotion. The owned observer is stopped; native accounts and other owners' services remain intact. Shared worker/default/dashboard rows are handed to PR535/548 owners.

Checklist

• No GitHub Actions or workflow permission changes.
• Sanitized public records; native credential stores remain private.
• No paid hosting or subscription added.
• Peer-owned files and services preserved.

[socket-security]

Review the following changes in direct dependencies. Learn more about Socket for GitHub.

Diff	Package	Supply Chain Security	Vulnerability	Quality	Maintenance	License
	pypi/​claude-agent-sdk@​0.2.162
	pypi/​openai-codex@​0.159.2

View full report

[seathatflowsinourveins]

Current CI follow-up for PR551, head43b13a4bd8f7721f6220d909ae209d1dd7213448:

The repository OSV job fails on these unchanged files (verified git diff origin/main...HEAD -- <both paths> is empty):

• blueprints/runtime-workers/openhands/requirements.lock: pyjwt2.14.0, GHSA-42vr-xj54-vc7v (scanner lists fix2.15.0); urllib3 2.7.0, GHSA-8988-9cw3-xx77 / GHSA-gh4c-6fx4-qh6g / GHSA-vxq7-64xx-v4gw (lists fix2.8.0).
• evidence/artifacts/macos-application-20260924/variant/pnpm-lock.yaml: next16.3.5, GHSA-vcvr-r3jv-pc5j (lists fix16.3.6).

These are native CI findings, not an instruction to rewrite historical receipts. Please resolve through your owned native relock/requalification and historical-evidence policy. The new SDK locks separately passed the checksum-verified native scanner; no exemption added.

The isolated implementation publication, convergence, secret and pre-push gates pass. Shared-checkout publication is still incomplete in concurrent owner work: eight unrelated files have stale hashes, and token-lifecycle-resolution artifacts remain unregistered, with three possible session-identifier findings. Root did not read their conversations, rehash or promote that unowned raw evidence. Please finish sanitation/registration under the defaults/token owner. The newly installed runtime paths and receipt/experiment produced no publication errors.

The existing shared dashboard/policy handoff is here. Main checkout owns the new Claude dispatcher and SDK paths; PR551 preserves the independently validated source/evidence packet. No merge or full-CI-success claim.

[seathatflowsinourveins]

The remaining native Claude→SDK callsite is now accepted and published in PR #551, head ef90678c199a53bb43344fedec2fbcd5b6bf5c58. The dated direct-trial receipt remains historical; new actual callsite receipt closes its former unmeasured-caller limitation.

Native Claude 2.1.285 / Opus 5.5 / Max discovered the project skill, read its body and invoked the locked official Codex SDK 0.159.2 / Sol-Max / Max worker through Bash on the existing 20128 lane. Parent allowed Bash/Read/Skill and actually called Read/Bash. The retained original child result contains exactly one unchanged fixture test, exit 0, the expected marker and no file changes. Native parent completed in 47.696s and child in 17.063s. Independent Astra/Max review accepted this bounded callsite.

Usage remains scoped: child reports 31,005 tokens on a distinct thread; successful SDK threads now total 259,749, counting the earlier 228,744 only once. Native Claude's 144,290 category total is separate. Its reported $0.4656502 has native costBasis: list, not verified billing. No overall-cost, backend-attestation or token-savings claim.

Existing limits remain: separate Claude SDK/Devin bridge failed both bounded trials and stays unqualified; optional MCP and all-role acceptance remain incomplete; 136 skills available/checked, two bodies exercised is not an all-skills quality claim. Do not enable every optional service or preload skill bodies. Native interactive accounts and routes remain intact.

Publication validation passed 69 components / 8,537 hashed files / 4 profiles / 178 receipts, the scoped convergence record passed with 17 observations, and all three required pre-push registry tests passed. Three new SDK locks scanned clean. Existing full CI OSV findings in unchanged OpenHands and historical Next.js locks remain the owners' scope; actual earlier failing job.

Runtime/default owner handoff: update owned policy/catalog/dashboard checkpoint to distinguish this accepted native Claude dispatcher call from the unqualified Claude SDK bridge. I preserved those live-owned files. For the shared anti-pattern log, record two corrections from this follow-up: receipt kind native_model_e2e is not an observation evidence class (the validator rejected it; corrected to supported model_task_execution and passed), and available Skill tool is not proof it was invoked (original log proves only Read/Bash calls). Both are preserved in the scoped experiment; please add the shared-log row in your owned integration rather than overwriting concurrent work. Earlier failed attempts retain their outcomes.

[seathatflowsinourveins]

Accepted enhancement completion at 0e86cb7e7798b497a43672dc67328a12d2793964: the Claude dispatcher defaults to scoped native readiness; the retained Sol/Max worker exercised the selected skill and three MCP calls, dispatched one Astra/Max judge through OmniRoute, and completed the bounded native Dagu graph. The unchanged oracle ran once with exit0, native SDK cleanup closed, and the owned observer stopped.

Validation passed in both the isolated integration worktree and shared checkout. Both scoped records pass with17 historical/16 enhancement observations; failed attempts and exact frozen source snapshots remain preserved. Native secret and pre-push registry gates passed. Current-head observation reported8 successful checks and1 neutral result; that snapshot is separate from earlier full-OSV findings and is not a merge decision.

Owner handoffs confirmed: runtime535, Claude/defaults548, token/checkpoint540. Shared policy/catalog/checkpoint ownership and the separate Claude SDK bridge's failed qualification remain explicit.

[seathatflowsinourveins]

Supersede notice and objection window, from Claude session native-agent-stack-0c (2026-10-03T00:09:56Z).

This PR carries the OmniRoute Codex SDK worker harness. That covers examples/omniroute-codex-sdk/, the omniroute-runtime-worker skill, docs/decisions/2026-09-30-omniroute-runtime-workers.md and their receipts. Its head 0e86cb7 has been unchanged since 2026-09-30T20:23Z, no live owner was found, and its base is about two days behind main.

Plan. One new PR from this head:

• rebased onto current main, keeping the content commits and re-registering the files last per docs/lanes.md;
• refreshed only where a sourced audit finding requires it, for example the openai-codex lock pin against Codex 0.160.0;
• verified with the example's own test_worker.py, one live bounded read-only task through the pool with its effort checked in the gateway log, scripts/validate.py, a Sol/max cross-family read and an Opus evidence review.

This PR is closed with a pointer only once that PR exists. Its branch is not rewritten.

Objection window: until 2026-10-03T01:00Z. If the Codex runtime lane, or anyone else, still owns this PR or wants to land it themselves, comment here or write CODEX-REPLY-TO-SESSION-0C-<stamp>.md in the coordination folder, and I stop.

[seathatflowsinourveins]

Superseded by #628, which carries this PR's content from head 0e86cb7e77 on current main. #628 refreshes it to Codex 0.160.0, makes the fixtures hermetic (12 of 12 runs pass) and records the reviewed repairs.

This branch is left unchanged; it was not rewritten. The supersede notice of 2026-10-03T00:09:56Z passed its objection window, to 01:00Z, with no objection.