3.1 KiB
3.1 KiB
Competitor profile protocol diagnosis (no paid retry)
Base: 471243390; runtime inspected read-only as user hermes.
Observed evidence
- Installed
/opt/hermes/bin/hermes chat --helpexits 0 and accepts every argument used by runner, including--query-file -,--oneshot,--format stream-json,--toolsets,--max-turns,--run-budget, and--ignore-rules. stream-json implies quiet. Invalid flags/banner/model unavailability are not supported by the evidence. - Canonical competitor config, .env, logs, and private cache are readable/writable by hermes. No credential values were returned.
- Read-only SQLite metadata for session
20261024_111941_64c954shows 8 API calls, 8 tool calls, noended_at, and no final assistant answer. Calls were three searches followed by five extracts; the last persisted entry is the fifth extract request without its result. - Agent logs show startup at 11:19:36, first search at 11:20:03, and last extract start at 11:21:23. This is consistent with runner's 120-second hard wall killing an active retrieval, including CLI startup. It is NOT evidence of invalid CLI flags or an unavailable model. Exact original exception cannot be proven: runner deleted stdout/stderr and collapsed the cause to RUNNER_ERROR.
- Installed web_extract schema requires char_limit >=2000 and its implementation clamps to 2000. Runner requested 1000 and falsely promised that would fit the emitter's 5000-character cap. Several persisted tool outputs exceeded 5000; such event output is truncated and cannot be decoded as retrieval proof. This remains fail-closed; truncated text is never accepted as source evidence.
Changes
- Preclaim
chat --helpcapability check for each canonical profile environment. No model/credential resolution occurs in this help check; it cannot establish upstream model availability. - Preserve constant safe categories for wall-clock timeout, process/init failure, malformed stream, and forbidden tool. Parse nonzero terminal envelopes before classifying; never persist or reflect provider text, prompts, stderr, or raw output.
- Correct char_limit to 2000 and bound requested search/retrieval exploration so the model has turns left to finish. Prompt guidance is not an enforcement guarantee and does not make unreliable retrieval available.
- Actual installed
StreamJsonEmitteris exercised directly without constructing an agent or invoking a model. Synthetic process failures are explicitly test-only and not research evidence.
Verification and limits
python3 -m unittest discover -s infra/hermes-editorial/tests -v discovers all editorial Python tests. The installed-emitter test runs when the installed module exists, or set HERMES_STREAM_JSON_SOURCE to a copied installed source. Never substitute a made-up envelope fixture for this check.
No task/lease/recovery mutations, runner execution, model invocation, production cron, deploy, or push were performed. No increase in cost reservation or wall-clock limit. Lost live output means timeout is strongly indicated, not conclusively reconstructed. Web availability and successful paid end-to-end completion remain unverified. Reservations remain unknown/retained.