test(agents): verify nested reasoning through public SDK

[decode2]

Linked issue

Refs #1702 (approved); SDK acceptance of public reasoning/revocation. Both issues remain open.

Stack

After #1737, base feat/1702-reasoning-tool. Test/documentation-only delta:245 changed lines.

PR type

• New feature (type:feature)

Summary

• Extend the permanent actual SDK fixture using public AgentSession.bindExtensions(ExtensionBindings) with mode: rpc and typed ExtensionUIContext. No private SDK assignment or direct executeTool shortcut.
• Verify five nested local registry requests independently of two main driver turns per tool prompt across four actual managers/runtimes with unique provider/API IDs.
• Assert explicit deny/unknown/headless zero-call, once re-prompt, session-target reuse, restrictive revoke and controlled in-flight public-source-change rejection.

Changes

Path	Change
tests/orchestrator-consultation-sdk.test.ts	Supported SDK UI binding, deterministic provider capture and real tool routing
docs/gentle-agents-activity.md	Evidence and precise limitations
odd/tasks/agent-coordination.md	Verification and incomplete issue status

Test plan

• Focused actual SDK fixture three consecutive final passes (combined execution included).
• Twelve-file SDK/helper/consent/consultation/state/catalog/discovery/presence/agents/prompt ownership/budget/routes suite:271pass,0fail.
• Types186baseline diagnostics,no regressions,12pairs improved;8runtime modulesmatch;diff whitespace clean.
• Native assessment245lines,medium,large runtime writer,underbudget: self-verification stands,no separate verifier required. Native outcome unknown,not approved/closed.
• Actual helper context contains static read-only system, one question/public JSON and empty tools; owner/caller private history,parent instructions,cursor/privatecapabilities excluded.
• Limits/abort signal/finiteusage/model IDs/non-authority fields match. Zero owner calls except explicit publication prompts; zero caller Git probes during consultations.
• Initial exact two presence records, per-manager actual socket readiness,nine→ten pagination,withdrawal/restoration/replacement and isolated cleanup remain covered.

Proof limits

Selector choices are explicitly simulated TEST choices, not real human consent, interactive TUI or RPC wire-client proof. Providers are local deterministic/no paid/outbound calls. Ignored-abort/runtime-replacement races remain controlled unit evidence, not new SDK claims. Windows fixture skips; no Windows integration proof.

Isolation

Only owned0700temporary root and exact canonical profile-derived socket leaf are created. Guards require absent-before/owner/mode/no-symlink and real actual socket paths. Public shutdown callbacks+abort/dispose precede bounded withdrawal and empty-leaf nonrecursive cleanup; UID parent/historical profiles remain untouched. Real user profile/settings/auth are never read or mutated.

Contributor checklist

• Approved issue linked nonclosing; exactly one selected type.
• Tests/docs/conventional commit,no attribution trailers.
• No production changes; shellcheck/skill checks not applicable.
• Remote checks pending after publication.

No fabricated strict-TDD RED or real human/native grant. Owner decision protocol remains pending. Parent #1737 functional remote checks passed; #1736's separate Windows candidate-marker failure is still not diagnosed. No main merge,auto-merge or reload.

Summary by CodeRabbit

• Tests
  • Expanded SDK integration coverage for nested helper consultations, including simulated consent choices, session-scoped permission reuse and revocation, bounded advice, and handling of source changes during execution.
  • Verified that caller history remains available to the main model but is excluded from helper advice, and that Git probes are distinguished by host.
• Documentation
  • Updated acceptance-test documentation with nested helper execution results and clarified that interactive human consent and Windows execution remain unverified.

[coderabbitai]

Navigate logical layers of code changes, visualize relationships, and explore their blast radius.

Note

Currently processing new changes in this PR. This may take a few minutes, please wait...

⚙️ Run configuration

• Configuration used: Repository UI
• Review profile: ASSERTIVE
• Plan: Advanced
• Run ID: 61d337c9-0673-4f5a-85f9-741918834235

📥 Commits

Reviewing files that changed from the base of the PR and between 40fe896 and 433ff31.

📒 Files selected for processing (3)

• docs/gentle-agents-activity.md
• odd/tasks/agent-coordination.md
• tests/orchestrator-consultation-sdk.test.ts

 ________________________________________________________________________________________________________________________________
< Design using services. Design in terms of services-independent, concurrent objects behind well-defined, consistent interfaces. >
 --------------------------------------------------------------------------------------------------------------------------------
  \
   \   \
        \ /\
        ( )
      .( o ).

✨ Finishing Touches

📝 Generate docstrings

• Commit to this branch
• Create a new PR

🧪 Generate unit tests (beta)

• Commit to this branch
• Create a new PR

• Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}