Tracking: work deferred from #388 (requestState rework)

[kingpanther13]

Follow-ups and deferred design work left open when #388 merged, so none of it gets lost. Items are grouped by who has the ball.

Deferred design work (from the #388 review, agreed as follow-ups)

• Worker-side slice budget / checkpointing. Detached workers (hub_set_rule, hub_set_native_app) run with no internal budget: the checkpoint/aggregation arms and _mrtrMaxContinuationSlices are unreachable for exactly the tools they were built for, and an extreme wizard drive runs until the platform kills it. Direction: give workers an internal slice budget (~2 min) that re-engages the existing checkpoint machinery.
• Client round-limit ceiling. Contention legs hold ~4.5s, so a worker running past ~45s needs more input_required rounds than the official SDK's default input_required_max_rounds of 10 — and rec.rounds only increments on a path detached tools never reach, so the server-side slice cap cannot hold the client limit it cites. Needs a design decision: document the ceiling, count rounds correctly, and/or surface a hint in the continuation envelope.

outputSchema removal

• Remove the outputSchema surface entirely: every tool's outputSchema definition, the publishOutputSchemas toggle, and the structuredContent emission it gates in _renderToolResult — all in the same PR, per the condition accepted in the feat: replace slow-write tokens with MCP request state, change to MRTR from new mcp spec  #388 review: no window may exist where a user can switch on an unexercised path. This also closes the accepted interim gap (zero validator coverage of structuredContent since feat: replace slow-write tokens with MCP request state, change to MRTR from new mcp spec  #388) and moots the known nested-schema blind spot in McpToolAnnotationsSpec's eliminated-fields scan.
• When the transition is old news, the loud opToken rejection in executeTool can be simplified or dropped.

level99's follow-up list (he offered PRs; recorded here in case they don't materialize)

• maxConcurrentWrites accepts a Long but _maxConcurrentWrites() does as Integer — a large value truncates and can silently disable the cap. Clamp/validate.
• MRTR_WORK_ITEMS has no sweep or cap, and _mrtrAbandon removes claimId before snapshotting cleanup, leaking a full args copy.
• _mrtrRecoverTerminalEvidenceLocked cannot fire in production (cold-cache/warm-evidence split only producible via a test-only seam).
• hub_call_rule's terminal aggregate double-counts retried failedRuleIds and is the only branch discarding aggregate.anyPartial.
• toolCheckRuleHealth guesses on a bare (Paused) suffix, contradicting the stopped comment above it.
• The new filter arg is missing from gateway summaries and searchHints, so it's undiscoverable in flat-catalog-constrained mode.
• ~280 lines of appCloner wizard logic live in the main file duplicated with mcp-app-cloner-lib.groovy, split by protocol era; the main copy dropped the "first click is silently swallowed" comment.

Open optimization

• E2e job time is ~70 min against the 60-min goal. The suite itself is ~56 min (from ~69); the remaining budget is fixed overhead: PR install ~5.9 min (load-bearing — proves the head landed), pre-run probe ~1.8, fixture cleanup ~3.4, restore ~1.1, conformance ~2.2. Further cuts need the install path or deeper suite work.

[coderabbitai]

🔗 Related PRs

#377 - feat: batch rule pause/run, modifyAction retarget, stageDisabled on clone/import [merged]
#378 - feat: deployment jobs argument, auto opTokens, and fail-loud multi-op guard [open]
#381 - chore: TEMP CI scratch — do not merge or review [closed]
#383 - feat: hub_call_device_commands — send commands to several devices in … [open]
#388 - feat: replace slow-write tokens with MCP request state, change to MRTR from new mcp spec [open]

---

📝 Issue Planner

_{Check the box below or use the @coderabbitai plan command to generate an implementation plan and prompts that you can use with your favorite coding assistant.}

• Create Plan

---

🧪 Issue enrichment is currently in open beta.

You can configure auto-planning by selecting labels in the issue_enrichment configuration.

To disable automatic issue enrichment, add the following to your .coderabbit.yaml:

issue_enrichment:
  auto_enrich:
    enabled: false

💬 Have feedback or questions? Drop into our discord!