fix: notify users via Slack when agent hits model call step limit

[langsmith-forge]

Problem

The agent silently terminates after hitting a 1000-step recursion limit without sending any Slack notification to users. Users start a task, the agent runs for 1000 steps and then stops, leaving users with no completion message, no PR, and no indication of what happened.

Traces:

• Agent silently stops at 1000 runs, 0 Slack replies
• Agent sends early Slack msg at run 22, then runs 978 more steps with no completion
• Agent silently stops at 1000 runs, 0 Slack replies
• Agent sends early Slack msg at run 15, then runs 985 more steps with no completion
• Agent silently stops at 1000 runs
• Agent silently stops at 1000 runs

Root cause

When DEFAULT_RECURSION_LIMIT = 1_000 is hit, LangGraph raises GraphRecursionError which bypasses all @after_agent middleware (including open_pr_if_needed). Users get no Slack notification that the agent stopped. In a typical 1000-run trace: ~72 LLM calls + 141 tool calls + 787 chain runs = 1000 total — so ~72 model calls hits the hard limit.

Fix

Two changes in agent/server.py and a new middleware file:

1. Added ModelCallLimitMiddleware(run_limit=60, exit_behavior="end") to the middleware stack. This intercepts gracefully at 60 model calls — before the 1000-step recursion limit fires — injecting an AI message with "Model call limits exceeded: ..." and routing to end (which DOES run @after_agent middleware).

2. Added notify_step_limit_reached — a new @after_agent middleware in agent/middleware/notify_step_limit.py — that detects the limit marker in the last AI message and posts a Slack thread reply:

   "I've reached my maximum step limit and had to stop. The task may be incomplete. You can retry with a more focused request, or ask me to continue from where I left off."

The notify_step_limit_reached middleware is placed last in the list so it runs after open_pr_if_needed (after_agent hooks run in reverse list order), ensuring any partial work is committed before the user is notified.

Evidence

No new tests written — this is a behavioral infrastructure change (adding middleware + a config parameter). Tests asserting exact middleware configuration or prompt content would be brittle. The production traces are the evidence.

• CI checks pass locally
• Existing tests pass, no regressions (107 tests pass with uv sync --locked --extra dev)
• Change is minimal and scoped — no drive-by refactors