[kryptobaseddev]

Inside a local POSIX Orca session, cleo orchestrate lead now opens a Claude Code terminal using the native provider's production guard builder. The launch carries the denied/allowed tools, permission mode, empty setting sources, strict MCP file, settings hooks and allocated session unchanged. A private first-turn gate waits for the exact local runtime/workspace, visible terminal, fresh UI navigation and matching connected terminal, supervisor PID, mailbox admission and persisted PID. exec preserves that PID; the result supplies the exact terminal handle for supervision.

Orca refusals propagate their code, message, fix and nonzero CLI exit instead of becoming success:true with a failed status. Environment/workspace and startup refusals precede mailbox admission; failure after admission releases it. Known newly created terminals are closed. Lost create receipts never trigger a retry or a guessed close: deleting the private gate aborts that shell promptly. Partial Orca context hints without a terminal handle refuse loudly with a repair instruction; the native route remains available outside Orca. Inside Orca, --detached returns while --print runs in a visible terminal (focused to prove UI adoption), with mbx retained as its owner channel.

Refs: T14500 AC1,AC2

AC1 evidence:

• The actual production preparation in lead-mcp.test.ts inspects argv, real --settings, strict runner/user-mbx configuration, provider session id and CLEO identity for both modes. It also proves coded refusals survive the adapter and generic setup failures receive actionable nonzero refusals.
• The offline transport fixture in orca-execution.test.ts executes only an inert recording binary. It proves argument/environment quoting, awaited admission before first turn, early Windows/missing-handle/remote-workspace refusal, visibility loss after create, failed navigation, PID mismatch, disconnected terminals, changed runtime, lost receipt shutdown, exact-handle cleanup and reported unverified cleanup.
• Core fixtures prove preflight refusals take no mailbox claim, post-admission refusal releases it, and successful leads retain their own session/record. CLI fixtures prove each refusal is success:false, preserves code/message/fix and sets exit status 63.
• Inert local Orca probes confirmed that create returns surface:visible, show omits surface, switch returns an exact focus receipt with navigated:true, and close returns ptyKilled:true. The implementation checks fresh navigation and matching re-show identity, and rejects an explicitly hidden re-show. These probes launched no model or worker and their terminals were closed.
• The opt-in live harness check remains unrun, as dispatched. Use a CI-built CLI from this PR in a fresh isolated project registered in local Orca, with a disposable verification-only epic, no ready workers and orchestrate.leadWorker.v1 enabled. Run cleo orchestrate lead <disposable-epic> --adapter claude-code. Capture data.leadTerminal and data.lead; inspect the returned handle with orca terminal show --terminal <handle> --json and orca terminal read --terminal <handle> --screen --json, and its PID with ps -p <pid> -o command=. Verify the denied/allowed tools, Manual permission mode, empty setting sources, strict MCP, --settings, provider session id, actual tool/server inventory and an owner question in that terminal. The fixture preserves pre-turn settings/MCP contents; SessionStart removes the MCP file. End the disposable session and confirm its claim/mailbox release. Do not use or interrupt the dogfood lead. Attach this evidence before AC1 closure.

AC2 evidence:

• Unsupported Orca harnesses still refuse before querying the host or taking mailbox admission. Explicit backend worker preflight/start still refuses lead launches for every worker harness.
• Windows, missing/partial identity, remote or mismatched workspaces, missing terminal capabilities, hidden/disconnected owner channels, failed navigation/PID proof and unsupported account placement fail closed. The lead path never queries account inventory or uses worker-start.

Validation on head d9d4872d336279b3c9d7cb6c2a098752fb87e67c (single files through cleo run --wait --class test -- pnpm exec vitest run <file>):

• packages/adapters/src/__tests__/orca-execution.test.ts: 94 passed.
• packages/adapters/src/providers/claude-code/__tests__/lead-mcp.test.ts: 21 passed.
• packages/core/src/orchestrate/__tests__/lead.test.ts: 28 passed, with --config temp/vitest.lead-local.ts.
• packages/cleo/src/cli/commands/__tests__/orchestrate-orca-routing.test.ts: 30 passed, with --config temp/vitest.orca-cli-local.ts.
• The ignored local configs add exact bare-package source aliases for this unbuilt checkout; they do not mock package implementations. Required CI validates the built packages and full checks.
• Changed-file Biome, diff whitespace, skills manifest/commands/coverage, contracts purity/subpaths and sync-invariant gates passed. The canonical changeset parser accepts the native task metadata. Local forge-ts filters exposed existing repository-wide debt; CI documentation coverage passed on the current head.
• NEXUS impact is UNKNOWN because static coverage is partial; source callers were inspected. Imports preserve the contracts port, with no core-to-adapters runtime dependency or new package dependency. Process transport stays in adapters, orchestration in core and shared types in contracts per AGENTS.md.

Current-head CI: Linux/macOS builds and build verification, packed artifacts, full lint and typecheck, documentation coverage, Repo Lints, Post-build Checks and release readiness passed. All eight Ubuntu unit-test shards passed. macOS shards 1, 2 and 4 passed; shard 3 failed in the unchanged brain-search-readonly.test.ts (missing/stale FTS setup waits 1008–1430 ms under a held write lock, including the automatic rerun). The failure stack is confined to memory/brain-search and its existing test; this PR changes neither that path nor store code. Follow-up T14509 is filed under the existing epic T12956. CI failure log. The foreground CI watch finished with exit 1. The CI aggregate is red from that test dependency; cleo/merge-close is red because AC1 and AC2 have no independent verdict on this head.

Skipped additional harnesses and profile/plugin APIs. The unrelated BRAIN CI failure, live check and independent verdicts on the current head are unresolved; this partial-delivery PR intentionally does not close T14500.

CI check states captured after the foreground watch on this head:

Check	Result
Build & Verify (macos-latest)	success
Build & Verify (ubuntu-latest)	success
Build (macos-latest)	success
Build (ubuntu-latest)	success
CI	failure
Changeset Lint (T13281)	success
Contracts Dep Lint	success
Core Bundle Budget (T14257)	success
Detect Changes	success
Documentation Coverage (forge-ts)	success
Flaky Quarantine Budget (T14257)	success
Flaky Test Quarantine (T13145)	skipped
Lint & Format	success
PR names its tasks	success
Packed CLI and Studio Artifact	success
Post-build Checks	success
Release Pipeline Matrix	skipped
Release Readiness Preflight	success
Repo Lints	success
Scripts Tests	skipped
Skills Depth Check / progressive-disclosure-depth	success
Type Check	success
Unit Tests (macos-latest, shard 1)	success
Unit Tests (macos-latest, shard 2)	success
Unit Tests (macos-latest, shard 3)	failure
Unit Tests (macos-latest, shard 4)	success
Unit Tests (ubuntu-latest, shard 1)	success
Unit Tests (ubuntu-latest, shard 2)	success
Unit Tests (ubuntu-latest, shard 3)	success
Unit Tests (ubuntu-latest, shard 4)	success
Unit Tests (ubuntu-latest, shard 5)	success
Unit Tests (ubuntu-latest, shard 6)	success
Unit Tests (ubuntu-latest, shard 7)	success
Unit Tests (ubuntu-latest, shard 8)	success
Worktree NAPI Prebuild Gate (T11127) / gate	success
Worktree NAPI Prebuild Gate (T11127) / matrix.triple	skipped
Worktree NAPI Prebuild Gate (T11127) / native source hash	success
cant-napi build	skipped
cleo-supervisor smoke	skipped
cleo/merge-close	failure