#2957: Add cost-protection pause for suspicious long-running sessions

Problem

A real WebUI session on the stable local runtime showed how a long-running agent turn can keep consuming model calls and time after it has likely entered a low-value recovery path. The observed turn was manually cancelled after about 6 minutes and had already made repeated model calls, context-compression attempts, and tool retries. The task had extracted the target article body but had not converged to writing the file or returning a final answer.

This is not a user-error problem. Users should not need to diagnose live agent logs to know whether a turn is productively working, retrying, compressing, or stuck in a high-cost recovery path.

Desired behavior

WebUI should not automatically decide that the task failed. Instead, when objective high-risk signals accumulate, WebUI should pause before the next expensive step and ask the user what to do.

Example copy:

This run may be stuck in a high-cost recovery path: 13 model calls, 11 context compressions, 2 tool errors, and no final assistant output yet. Continue?

Actions:

• Continue
• Stop
• Summarize and stop

Candidate risk signals

• Long active run age with no final assistant output
• Repeated context compression in one turn
• Compression timeout or repeated compression failure
• Repeated API retry or provider connection failures
• Repeated tool errors for the same target
• High model-call or tool-call count for one user turn

Scope for first PR

A first slice should be deliberately narrow:

• Add a backend cost-protection/risk gate at safe run-loop boundaries, before issuing the next model/tool call where feasible.
• Emit a structured event to WebUI when the gate pauses a run.
• Render a confirmation card in the existing chat flow with Continue and Stop actions.
• Keep auto-stop out of scope unless the user has configured a hard budget.
• Preserve existing Stop/cancel behavior.

Non-goals

• Do not let WebUI silently judge a task as failed.
• Do not pause inside an already-blocking model/compression call unless that call already supports cancellation.
• Do not introduce a new runner process or large runtime-adapter migration in this slice.

Evidence from local incident

Observed on a local 8787 stable runtime session while clipping a WeChat article:

• Active for about 6 minutes before manual cancellation
• 13 model API calls
• 11 context-compression attempts
• 2 tool errors
• At least one 120s auxiliary compression timeout
• Article body had been extracted, but no final saved Markdown was produced

Contract routing

Task type: runtime / streaming / user-facing safety UX Touched areas: run lifecycle, SSE events, chat UI, cancellation/continue controls Relevant docs:

• AGENTS.md
• CONTRIBUTING.md
• docs/CONTRACTS.md
• docs/rfcs/webui-run-state-consistency-contract.md
• docs/rfcs/hermes-run-adapter-contract.md
• docs/UIUX-GUIDE.md
• DESIGN.md