Diagnosed from a live voice session (e1b9fccb90ef): three turns failed with
raw '**Error:** HTTP 502 ... hermes-{claude,codex}-broker' because the agent
pod hosting the model brokers rolled mid-conversation. The voice client
correctly refused to speak the error envelope, but it then dropped the user's
utterance and forced them to repeat it three times.
Voice: on a TRANSIENT provider error (5xx/502/'error sending request'/timeout),
conversation mode now re-runs the errored turn in place through the app's own
regenerate action (which truncates the errored turn — no duplicate user
message) and stays in Thinking so its cues cover the reconnect gap. Bounded to
MAX_TRANSIENT_RETRIES (2); a non-transient error or an exhausted budget still
drops cleanly to 'let's try that again — listening'. The raw error is never
spoken. New probe scenarios cover retry-then-recover and the bounded-then-drop
path; source contract updated.
STT: the same session mis-transcribed 'CUI' as 'cue'. Prime the default
initial_prompt with the domain acronyms the user uses (CUI, FOUO, DoD, NIST,
CMMC, FIPS, RMF, POA&M, ATO, SBU) so they bias to uppercase forms.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BvMSXH8VH2tMWXanb8SJdf
Default final beam 5 -> 2: the accuracy/speed knee - most of beam
search's benefit at ~2x greedy instead of ~5x, protecting commit
latency on the Jetson (still env-overridable via
HERMES_STT_FINAL_BEAM_SIZE). Prompt now also primes common names (Amy,
Claude, Hermes) so 'Amy' stops transcribing as 'aiming'.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BvMSXH8VH2tMWXanb8SJdf
Proper nouns (Córdoba, Cancún) and dropped words came from decoding the
committed transcript with the small model. The image already ships
large-v3-turbo, so the final decode now uses it with beam_size=5, a
temperature fallback ladder, and a proper-noun/accents initial_prompt
that fixes first-pass capitalization and diacritics across EN/ES/RU;
the rolling previews stay on tiny at greedy so the on-the-fly feel is
unchanged. The accurate decode runs in the speculative predecode during
the end-of-speech silence and is cache-reused at commit, so perceived
latency stays low. All decode knobs are env-overridable for on-device
tuning (beam/temperature/prompt), with small as the guaranteed-present
rollback if turbo underperforms on the Jetson. 206 STT tests pass.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BvMSXH8VH2tMWXanb8SJdf
