← Back to Blog Aug 30, 2026 PR #1286

PR #1286 — Eliminating Agy Headless 5-Minute Timeout, Enabling Plan Mode, and Capturing Prompt Cache Telemetry

The Broader Goal at the End of the Previous PR

After establishing full headless parity for OpenAI Codex (PR #1275) and xAI Grok (PR #1279), the objective shifted to evaluating harness parity for the remaining core fleet members: Google Agy (Antigravity CLI) and Anthropic Claude.

Strategic Shifts in This PR

For months, Agy dispatches that lasted longer than 5 minutes failed intermittently with Error: timeout waiting for response (HARNESS_INTERNAL_TIMEOUT), often leaving zombie jobs in daemon_jobs. Simultaneously, Agy could not be dispatched for read-only audits or code reviews without write permissions due to an unnecessary PermissionEnforcementError fail-closed check. In addition, prompt cache hits in Gemini were discarded as 0 cache reads, distorting cost tracking.

PR #1286 resolves all three structural defects natively.

What This PR Shipped

  1. Pass --print-timeout 30m0s on All Headless Agy Dispatches:
    • Declared --print-timeout and --mode in synlynk/_constants.py under HARNESS_CAPABILITY_BASELINES["agy"]["dispatch_flags"]["valid_flags"].
    • In synlynk/dispatch.py:dispatch_agent(), dynamically appends ["--print-timeout", "30m0s"] for agy dispatches, preventing the default 5-minute client-side timeout.
  2. Native Read-Only Planning Mode:
    • In synlynk/dispatch.py:_permissions_to_flags(), replaced the PermissionEnforcementError for permissions <= {"read:*"} with ["--mode", "plan"], allowing Agy to be safely dispatched for read-only audits and reviews.
  3. Capture Gemini Prompt Cache Telemetry:
    • In synlynk/costs.py:_extract_agy_structured(), extracted cache_read_tokens = int(usage.get("cache_read_tokens", 0)) instead of hardcoding 0. During our dogfooding run, 7.14M prompt cache tokens were accurately captured!
  4. Comprehensive Test Coverage:
    • Added unit tests verifying --print-timeout 30m0s injection, --mode plan translation, and cache token extraction across tests/test_dispatch.py, tests/test_agent_cli.py, tests/test_constants.py, tests/test_cost_ledger.py, and tests/test_costs.py.

What This Achieved on the Path to Autonomy

Agy can now reliably execute deep, multi-minute implementation tasks and read-only reviews without premature aborts or distorted billing telemetry.

← Back to all posts