[MCP] Adopt MCP Tasks so long run_agent can outlive the host session #102

Why

#83 (merged) keeps a long run_agent inside the current host request: progress notifications, a bounded lock wait, cancellation, and documented host timeout settings. That still assumes the tool call is one open request the host waits on.

In late October 2026 hosts moved past that model. Claude Code v2.1.304+ (public beta, protocol revision 2026-07-28) can turn a long MCP tool into a server-side task: the tools/call returns as soon as the server accepts the job, the job keeps running on the server, and the client polls status (including across /resume). Coverage and the six host-side config decisions: MCP tasks landed in Claude Code (2026-10-26). Related industry notes on progress resetting client timeouts (still useful alongside tasks): MCP “Request timed out” (-32001) on long tools (tested 2026-10-09).

run_agent is exactly the long call (Allrecipes budgets 300 s; the README chat-model Allrecipes run was 138.7 s). Today s1a-mcp is pinned to mcp>=1.26,<2 because openjiuwen / fastmcp still need mcp 1.x (#74). MCP Tasks need the 2026-07-28 revision and a server that answers with a task when the client declares io.modelcontextprotocol/tasks.

Proposed scope (one PR, opt-in)

1. Spike whether Tasks can land while staying on mcp 1.x, or whether the mcp 2 / fastmcp 4 move from [Deps] Bound mcp below 2 and record what blocks the move to the 2026-07-28 SDK #74 is a hard prerequisite; record the blocker in the issue if so.
2. If feasible: make run_agent (and only that tool) return an MCP task when the client advertises task support; keep today's synchronous path when it does not.
3. Progress / status updates on the task; cancel maps to the existing cancellation path from [MCP] Keep long run_agent calls inside host tool timeouts: progress, bounded lock wait, cancellation, documented settings #83.
4. Document Claude Code / Codex settings for Tasks vs plain tool timeouts in docs/skills.md (link the host docs; do not invent host flags).
5. Contract tests with a fake client capability blob: task path vs sync path; no live browser required.

Acceptance criteria

• With a client that does not advertise tasks, behaviour matches today (existing MCP tests pass).
• With a client that does, run_agent accepts and reports a task id / status without holding the original request open for the full browser budget.
• Cancel stops the agent within the same bounds as [MCP] Keep long run_agent calls inside host tool timeouts: progress, bounded lock wait, cancellation, documented settings #83.
• Skills / README name which hosts were verified and which still need a raised tool_timeout only.

Out of scope

• Reworking decide / list_agents into tasks.
• Lifting the mcp pin for its own sake (only if Tasks require it; otherwise leave [Deps] Bound mcp below 2 and record what blocks the move to the 2026-07-28 SDK #74's follow-up alone).
• Guaranteeing every host (Codex, Cursor, Claude desktop) supports Tasks — document what does.

References

• Claude Code MCP tasks write-up: https://dev.to/piekwerk/mcp-tasks-landed-in-claude-code-a-tool-call-can-now-outlive-your-session-and-6-config-decisions-2l5k
• Progress / timeout reset measurements: https://nandanpriyadarshi.medium.com/mcp-request-timed-out-32001-on-long-tools-when-progress-notifications-help-6ebbdb9234ab
• Related: [MCP] Keep long run_agent calls inside host tool timeouts: progress, bounded lock wait, cancellation, documented settings #83 (in-session progress), [Deps] Bound mcp below 2 and record what blocks the move to the 2026-07-28 SDK #74 (mcp<2 pin), [MCP] Give s1a-mcp tools titles, behavior annotations and typed output schemas #82 (tool titles / schemas)