Agent Flight Recorder

Done. The sample log (270 events, 30 cycles, one run id) was written through the skill's own TelemetryLogger, and demo_recorder.py dispatches the two tools exactly as a harness would. Per-million rates in it are illustrative placeholders, not list prices.

What the agent sees from get_execution_metrics(last_cycles=10) (per-cycle rows compacted to one line each):

{"cycles": {"from": 21, "to": 30, "count": 10},
 "per_cycle": [
  {"cycle": 21, "model": "claude-sonnet-5-5", "action": "WAIT",     "input_tokens": 267420, "output_tokens": 4033, "cached_tokens": 180428, "llm_calls": 4, "est_cost_usd": 0.3586, "tool_calls": 3, "tool_errors": 1, "duration_s": 41, "failures": 0},
  {"cycle": 22, "model": "claude-sonnet-5-5", "action": "WAIT",     "input_tokens": 261557, "output_tokens": 4330, "cached_tokens": 174327, "llm_calls": 4, "est_cost_usd": 0.3612, "tool_calls": 1, "tool_errors": 0, "duration_s": 47, "failures": 0},
  {"cycle": 23, "model": "claude-sonnet-5-5", "action": "POST",     "input_tokens": 360693, "output_tokens": 5360, "cached_tokens": 253420, "llm_calls": 5, "est_cost_usd": 0.4586, "tool_calls": 3, "tool_errors": 0, "duration_s": 49, "failures": 1},
  {"cycle": 24, "model": "claude-sonnet-5-5", "action": "COMMENT",  "input_tokens": 190394, "output_tokens": 3220, "cached_tokens": 148904, "llm_calls": 3, "est_cost_usd": 0.1991, "tool_calls": 3, "tool_errors": 0, "duration_s": 24, "failures": 0},
  {"cycle": 25, "model": "claude-haiku-5-5",  "action": "WAIT",     "input_tokens": 99191,  "output_tokens": 2212, "cached_tokens": 70113,  "llm_calls": 2, "est_cost_usd": 0.0471, "tool_calls": 1, "tool_errors": 0, "duration_s": 18, "failures": 0},
  {"cycle": 26, "model": "claude-sonnet-5-5", "action": "RESEARCH", "input_tokens": 509318, "output_tokens": 7296, "cached_tokens": 20044,  "llm_calls": 5, "est_cost_usd": 1.5653, "tool_calls": 1, "tool_errors": 0, "duration_s": 57, "failures": 0},
  {"cycle": 27, "model": "claude-sonnet-5-5", "action": "WAIT",     "input_tokens": 384184, "output_tokens": 5870, "cached_tokens": 255861, "llm_calls": 5, "est_cost_usd": 0.5333, "tool_calls": 3, "tool_errors": 0, "duration_s": 63, "failures": 0},
  {"cycle": 28, "model": "claude-sonnet-5-5", "action": "WAIT",     "input_tokens": 391884, "output_tokens": 6346, "cached_tokens": 303208, "llm_calls": 5, "est_cost_usd": 0.4325, "tool_calls": 3, "tool_errors": 1, "duration_s": 89, "failures": 0},
  {"cycle": 29, "model": "claude-sonnet-5-5", "action": "COMMENT",  "input_tokens": 208681, "output_tokens": 3055, "cached_tokens": 158197, "llm_calls": 3, "est_cost_usd": 0.2259, "tool_calls": 3, "tool_errors": 0, "duration_s": 25, "failures": 0},
  {"cycle": 30, "model": "claude-haiku-5-5",  "action": "POST",     "input_tokens": 106292, "output_tokens": 2405, "cached_tokens": 78918,  "llm_calls": 2, "est_cost_usd": 0.0473, "tool_calls": 2, "tool_errors": 0, "duration_s": 20, "failures": 0}],
 "spend": {"total_usd": 4.2289,
           "by_model": {"claude-sonnet-5-5": {"usd": 4.0616, "calls": 26}, "claude-haiku-5-5": {"usd": 0.1673, "calls": 12}},
           "note": "estimated from logged tokens; the budget ledger is authoritative"},
 "tools": {"calls": {"recall": 10, "web_search": 6, "publish": 4, "read_file": 3}, "errors": {"web_search": 2}, "error_rate": 0.087},
 "actions": {"WAIT": 5, "POST": 2, "COMMENT": 2, "RESEARCH": 1},
 "failures": [{"cycle": 23, "type": "action_failed", "detail": "publish returned 422: duplicate of post published in cycle 20"}]}

That result is about 3,000 characters, well under the 7,000-character cap.

What the agent sees from query_telemetry(event_type="tool_call", errors_only=true, fields="tool,error,duration_ms"):

{"event_type": "tool_call", "returned": 4, "truncated_for_size": false,
 "rows": [
  {"ts": "2026-10-16T14:51:42", "cycle": 28, "tool": "web_search", "error": "TimeoutError: read timed out after 30s", "duration_ms": 30322},
  {"ts": "2026-10-16T14:21:47", "cycle": 21, "tool": "web_search", "error": "HTTP 503 Service Unavailable from upstream search provider", "duration_ms": 720},
  {"ts": "2026-10-16T13:47:17", "cycle": 12, "tool": "web_search", "error": "HTTP 429 Too Many Requests: rate limit exceeded, retry-after=18", "duration_ms": 2211},
  {"ts": "2026-10-16T13:27:12", "cycle": 7,  "tool": "web_search", "error": "TimeoutError: read timed out after 30s", "duration_ms": 30270}]}

Each failure was followed by a successful retry except the 503 in cycle 21.

What the data says about the budget blowup in this sample:

• Cycle 26 cost about four times a typical sonnet cycle. Its cached share dropped to 4 percent while input tokens hit 509k. That is the signature of a context rewrite that cold-started the prompt cache.
• Normal sonnet cycles run 65 to 80 percent cached. Haiku-only cycles cost under a tenth of a sonnet cycle.
• Over all 30 cycles the estimated spend is about 10 USD at the placeholder rates, with sonnet at 95 percent of it.

The first check I'd have the real agent run on its own log is cached_tokens / input_tokens per cycle. If the ratio collapses in the cycles where you blew the budget, the fix is in whatever rewrites the context, not in the model choice.