analysis/summary_tables.md

Analysis summary (generated by scripts/analyze.py from evidence/)

Timeout (80408) events

• total: 33, all reasonForCode: 80408
• per day: 2026-09-07: 1, 2026-09-08: 4, 2026-09-09: 28
• densest 75-minute block: 20 events (the resume-button loop)

Quota/account events

• log lines: 50 → clustered (≤2 s apart): 18 events

• distinct codes seen: -32603, 100400, 112

• distinct message wordings: 3

  • Access denied, please make sure your account is in good standing. For details, see: https://help.ali
  • Allocated quota exceeded, please increase your quota limit. For details, see: https://www.alibabaclo
  • Your token-plan 1-week quota has been exhausted. The quota will reset at 09-08 01:45:00 UTC.

chat_finish census

• 245 lines total: 225 success:200, 20 error payloads, zero attributable to the 33 timeout events (no chat_finish line exists for them)

Three largest sessions (transcript entries)

session	entries	jsonl bytes
sess-11	626	379526
sess-14	280	154497
sess-15	208	253155

Task lifecycle (quest.log — the client's task = one conversation)

• 1252 log lines → 524 events after ≤2 s de-dup (102 events replicated into ≥2 IDE windows, 626 .skipNonQuestTask twins dropped)
• statuses: ActionRequired 33, Completed 208, Error 18, Running 258, Stopped 7
• Running spans: 257 closed, 1 still open at the extract horizon; median 126 s, 90th percentile 679 s, max 4131 s
• max simultaneous Running tasks: 2 (never 3+; 10 overlapping pairs of conversations), error.code.task.running.quota.exceeded fired 0 times, so that ceiling is only bounded ≥2
• ActionRequired (banner up, nothing generating): 33 events, 1:1 with the 33 timeout resume lines (largest offset 234 ms)
• conversation-level split of those timeouts, which supersedes the window-level correlation in report §4.2: sess-03 23, sess-04 4, sess-11 5, sess-16 1
• watchdog measured off the state machine: 61 s of silence before the banner (16 of 33 banners at exactly that span; the rest are mid-run stalls after some output), and 155 of 194 successful runs ran longer than that span — so it caps inactivity, not total duration

Per-interaction phase timing ([ACPProgressStateMachine])

• 885 transitions over 9 conversations: 252 prompts sent, 245 answered with a first streamed chunk, 4 rejected before any chunk, 208 generations finished
• wait for the first chunk (prompting → streaming, the UI's 'Thinking…'): median 9.6 s, 90th percentile 49 s, max 132 s, min 1.2 s
• generation time (streaming → chat_finish:success): median 105 s, 90th percentile 477 s, max 2310 s; 149 of 208 exceed the 61 s watchdog span
• what suspended a turn: permission_request 37, resume_tool_call 33 (resume_tool_call = the 80408 banner, a third stream confirming 33)
• operator hold: banners 32 resumed, median 30 s, total 90.5 min; permission dialogs 37, median 88 s, total 122.0 min

Request lifecycle (frozen log snapshot; scripts/latency_from_logs.py)

• 308 requests across 10 sessions — S 211, T 71 (timeout 33 / dialog 38), F/U 26
• first-attempt usable (0-retry S) 179/211 (84%)
• first-chunk wait median 11.3s (n=300); clean-success turn total median 120.0s (n=179); send-to-stall median 305.4s (n=71)

Incident 2 — collection round of 2026-09-10 (evidence/incident_20260910/)

• 62 request episodes (55 distinct submitted prompts) in 3 conversations on the observation day: 53 completed, 5 timeout parks, 2 dialog parks, 2 failed/cancelled
• the five timeouts all sit in one conversation at 490,789 tokens; the largest context that still completed that day is 452,447 tokens — a 452,447-versus-490,789 separation on the day (round 1 pooled data does NOT separate cleanly; see report §4.2)
• send-to-stall spanned 125s to 4861s (the upper value is a banner left parked for 81 minutes)
• paired experiment: the same conversation stalled five times at 494,193 tokens, then after the client's context-compaction control reported 117,153 tokens (Compression API call completed for session compact) a request carrying 85,115 tokens completed in 10.1 s — report §9
• three negative results from this round: tab visibility (a hidden conversation completed in 2.58 s), server-side cache residency across a 27.7-minute idle gap (17.11 s versus 17.58 s at the same size), and inflating an agent's context by asking it to read files (one such turn added 851 tokens)

Content hash sha256:2218b3e37e6e4307ff5b1451017d19d5f5c86ac217c803d160d0111a34f5ee77 - generated 2026-09-11 10:17:52 - rerun: python3 scripts/analyze.py