[POLAARK]

Summary

Clicking Looks correct on a checkpoint gate (confirm_summary / request_plan_approval) in the UI never authorizes anything. The agent re-raises the checkpoint after every click, up to CHECKPOINT_MAX_ROUNDS.

Cause. The UI renders checkpoint gates with the generic QuestionEditor, which stores option indices:

{ "answers": [{ "selectedOptions": [0] }] }

chosenLabel (lambda/agentcore/mcp/process-bridge.js) only read perQuestion[0].answer, freeText, decision or a bare string, so it returned "" for every UI answer:

• The resume message (formatResumeAnswer) told the agent The human answered the summary-confirmation checkpoint: "" — nothing is authorized yet … call confirm_summary again.
• On the resume leg, rehydrateAuthorizations never minted the receipt.
• A UI Request changes never recorded the withdrawal of an earlier approval.
• The inline path returned re-ask.

Observed on a deployed 2.11 stack: four answered rounds, each stored as {answers:[{selectedOptions:[0]}]} with option 0 = "Looks correct", and the agent kept saying the confirmation "came back empty".

Fix.

• chosenLabel(answer, questions) resolves a single well-formed index against the gate's own questions[0].options, which raiseCheckpoint writes, so the labels stay owned by the platform. All five call sites pass the gate's questions.
• Only that exact label authorizes. A multi-select, out-of-range, negative or non-integer index, or a typed note with no option, names no label and gets a re-ask.
• answerFreeText passes the reviewer's note (answers[0].freeText) to the agent as the requested-changes feedback.
• Existing answer shapes behave as before.

Tests

New tests use the real UI shape:

• inline approve and reject with a note
• resume-leg receipt minting
• resume-leg withdrawal: the existing test now runs for both the perQuestion and UI shapes
• formatResumeAnswer approve and reject
• malformed UI answers create no authority

The UI-shape tests fail on main and pass with this change.

• vitest run --project=agentcore: everything passes except spawn-shutdown.test.js (5 s timeout under full-suite load). It passes 9/9 when run alone, repeatedly, and does not touch this code.
• npm run lint: 0 errors. The 3 warnings in the touched files are pre-existing.
• npm run format:check and npm run secretlint: clean.

Not in this PR

changeControlChoice (run-stage.js) parses the strict change-control gate the same way, so it probably has the same blind spot for UI answers. I left it out to keep this PR focused; it can be a follow-up that shares the index-to-label helper.

[JWThewes]

That's a good finding. Thanks for this one. Looks good to me.

[POLAARK]

Sorry I had a review agent run a little cleaning. Should be good right now @JWThewes

[eipasteur]

TEST PURPOSES ONLY — DO NOT TAKE INTO ACCOUNT.

Automatically computed PR review signals (deterministic)

Cognitive complexity and branch coverage help focus review on functions and their tests; these signals are advisory.
2 changed JS/TS files analyzed · 1 flagged function · 1 improvement.
Coverage: PR measurements from incomplete tests; no comparable base, so no coverage delta.

Complexity and coverage — 2 functions

	File / function	Complexity	Δ	Branch coverage	Δ	Uncovered branches	Depending files
🟠	lambda/agentcore/mcp/process-bridge.js:506 — createProcessBridge / rehydrateAuthorizations	41 → 43	+2	— → 78.0%*	—	— → 11	2
🟢	lambda/agentcore/commands/run-stage.js:1308 — formatResumeAnswer	26 → 24	-2	— → 68.0%*	—	— → 16	1

🟢 Improving · 🟠 Needs attention · 🔴 High risk · 🟣 Highest risk

— = unavailable · n/a = not applicable · * = incomplete tests

---

Isobar map

Fixed repository map: red marks changed files; green marks static name matches or import links across components. These links indicate possible reach, not proven runtime impact or a risk grade.

Isobar by r3al1tym.

Signals

• A: agentcore/mcp/process-bridge.js:506 — function CC 41→43 (+2); 2 direct production importers; 17 file-level import-link targets.
• B: agentcore/commands/run-stage.js — 1 direct production importer; 13 file-level import-link targets.
• Total static reach: 2 unchanged production files; 15 unchanged test files.
• Revision: fac3645; updated 2026-10-27 17:28 UTC.

Legend

• ○ modified · ◇ added · × deleted. Blue markers: tests; gray markers: documentation.
• Solid green: import-linked name matches. Dotted green: broader file-level import fallback. Both are static evidence, not proven execution.