iii-harness-53-bugs.md

53 things from ~4 months of running an agent harness in production.

iii, the engine that powers the harness, is really cool too: https://iii.dev/docs/install

• crash mid tool call left a dangling call with no result that poisoned every later request workers#507 workers#630
• resumed turn raced the worker booting and called a function that wasn't registered yet workers#507
• no durable queue for turns so wedged ones sat in running forever workers#464
• cut stream parsed as null args and got reported as a successful empty turn workers#878
• seven providers treating any connection close as done workers#878
• cached tokens double billed on six providers workers#740
• file reads counted twice so we saw phantom context overflow workers#552
• retry budget of 1 killing turns one step from done workers#552
• compaction returning an empty context workers#552
• child agent death silent by construction workers#552
• lock order deadlock between parent retasking a child and the child finishing workers#973
• ci only dodged the deadlock by scheduling luck workers#973
• stop button clobbered by a stale write from the step it was stopping workers#313
• one 18mb directory listing wedging the whole agent runtime through restarts workers#1175
• poison message reflushed on every reconnect workers#1175
• queue faithfully redelivering the poison workers#1175
• wake notifications cutting json at 600 chars, mid array workers#1167
• 10mb session log replayed synchronously under a mutex workers#1119
• approval hook registered 171 times because the instance count lagged workers#797
• resuming past only the first duplicate hook, held forever workers#797
• failure cleanup timing out and logging skipped, same as a no op workers#1108
• a span emitted per token plus a persisted state write per event workers#205
• 41% more tokens, 95% of it repeated context cache reads workers#1123
• large tool calls ending with incomplete args, then re emitted workers#1123
• registrations lost when the runtime reloaded a different worker workers#666
• agent uninstallable because its dependency graph had 65 edges and the limit was 64 iii#2006
• websocket reset loop every 2.8 seconds
• stop command vanished because registrations weren't replayed after reconnect
• startup sweep rerunning the same doomed call after every boot
• parent parked forever waiting on a child that died out of band
• two parallel approvals both waking one turn, shell command executed twice
• provider stream stalled mid delta, keepalive pings counted as activity
• 120s idle guard that never fired, 300s hard kill with an opaque error
• no read timeout on upstream clients
• agent restarting the very process hosting its own turn
• "confirm before restarting" in the prompt, not a guarantee
• generated code polling state 140 times a second, 29.7k spans in one trace
• trace cache eviction went quadratic under a lock every thread needed, kill -9
• null serialized as "no result", sdk saw undefined, infinite loop
• 16 compactions, then a 5m char hook result, "compaction succeeded"
• five sub agents hitting the turn limit, all reporting completed
• two ids that differed by one character colliding, second message silently dropped
• malformed deny rule failing open
• unknown model fell to an 8k window, sub agents had about 6k usable
• first turn system prompt 14 tokens different, second cache write every session
• 79% of tool schema tokens were prose
• any result under 2000 chars immortal in context, 63k to 76k token floor
• no completion trigger so agents used sleep 60 as a wait primitive
• 48% of root turns were status polls
• sub agents not inheriting reasoning effort, 32 min to 18 min, $1.16 to $0.50, same score
• two components each persisting a registration token, desync no restart could fix
• two runtimes on one port, function ownership flapping, oom sigkill
• 761mb of traces at 20mb a minute in a quiet session

the harness is open source: iii-hq/workers/harness.