[remote-controlled-man]

Summary

• Give a final answer rejected for tool-call markup one additional text-only turn after the normal iteration cap.
• Keep that turn out of tool execution and goal continuation, clear rejected streamed text, and retain the existing grounding, cancellation, token-budget and usage paths.
• Add eight regression cases through AgentLoop.run() and describe the bounded retry in the README.

Why

Refs #2122. The existing plain-text retry requires another iteration, but a forced-text answer normally arrives on the last permitted iteration. A valid re-emission was therefore unreachable after the cap.

The extra turn is only scheduled for a withheld tool-syntax answer. It is counted in reported iterations and usage. A second markup response or an unoffered structured tool call uses the existing degraded fallback; it cannot reopen research. This PR addresses the retry path; CLI/Web degraded-status presentation is unchanged.

Test Plan

• 56 relevant agent-loop, streaming/interruption, outcome, compaction, grounding and goal/context test modules: 1665 passed, no skips (9 dependency deprecation warnings).
• The two cap-recovery cases fail on the original base; all eight new cases pass with the fix. The complete DSML module has 11 passing cases.
• Changed-file Ruff, Black changed-line checks, ci_private_files_gate.py, all seven ci_grep_gates.sh gates, and git diff --check pass.
• Black checks cover the changed line ranges; unrelated whole-file formatting was not included.

The tests use offline scripted providers and synthetic tools. The full repository suite, frontend build, real providers and live/broker flows were not run.

Scope and risk

This changes the protected agent loop already discussed in #2122. One answer-only model turn can occur beyond the configured iteration cap; cancellation and an exhausted goal token budget still stop it. Broker, credential, MCP and network configuration are unchanged. Roll back by reverting the fix commit.

• DCO sign-off included.
• README updated.
• No secrets or hardcoded runtime paths.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota. Try requesting a review at a lower effort level.

An organization owner can choose the repository organization as the billing source for reviews requested by members with a paid Copilot license. Learn about Copilot code review billing.

[he-yufeng]

I checked the full diff and ran it in an isolated checkout of 1a7969a4484663ee68b35f380c97128e8b92b891.

The bounded retry is wired to the right failure mode. A withheld final answer gets one text-only turn only after the normal cap is exhausted; that turn cannot execute tools even if the provider returns structured calls, it cannot continue a goal, and it still stops on cancellation and the goal token budget. The new tests cover the cap case, the unoffered-call case, budget accounting, cancellation, grounding, and the correction path.

Verification: the focused loop tests pass 75/75, and the full non-E2E suite passes 21,273 with 111 skips. tools/ci_private_files_gate.py and git diff --check also pass. CI is green on all ten checks.

This is a good fix for #2122. Approving.