Voice replies had to come from the session's main model, so a heavy model made every spoken
answer slow (#83894). A new auxiliary slot, voice_chat, picks the model for voice turns:
- CLI/TUI voice mode, the Desktop voice conversation (prompt.submit voice_turn) and messaging
voice notes mark the turn; agent/voice_turn_route.py binds the route after the prompt, row and
preflight compaction were settled against the main model and undoes it before memory sync and
background review, so the next typed turn and the review run on the main model.
- A turn larger than the voice model's compression threshold stays on the main model with a
one-time notice: switching to voice never compacts the conversation.
- Usage lands in session_model_usage under task voice_chat; it never rewrites the session's
recorded model/route, so resume keeps the model the user picked.
- Slot appears in Desktop Settings -> Models, the dashboard, and `hermes model`; GPT-Live
delegations keep using the session model.
* fix(agent): free-tier rate limits end an attended turn at once
A free-tier (no sign-in) session that hit the service-wide rate limit slept
up to 600 s, retried into the same refusal, slept again, and then sent its
third attempt on a guest JWT that had expired during the wait. The welcome
host answered 403 "invalid jwt", which read as the tier being switched off.
The user watched a spinner for about 18 minutes and then saw the wrong card.
- An attended session (desktop, TUI, the interactive CLI, ACP) on the free
model ends the attempt cycle at once when the cooldown is longer than
60 s: fallback first, else the reset time and the sign-in / model doors.
"Attended" is keyed on the session source, so `hermes chat -q`, cron and
messaging sessions keep the long wait. The source vocabulary moves from
run_agent.py into agent/session_source.py, which adds is_attended().
- on_free_model() (welcome host + anonymous credential) is the one free-tier
check, used by the cutoff, the "free model is busy" wait copy and the
outage copy.
- A welcome-host 403 "invalid jwt" is its own route kind, session_expired:
the existing one-shot route heal re-exchanges the credential and retries;
if that fails the card is the retryable outage card.
- interruptible_backoff_sleep adopts a fresh Nous key after each wait.
_adopt_nous_key_before_expiry owns its refresh failure, so both call
sites drop their try/except.
- A retry wait over 60 s is announced in the transcript when it starts, for
every provider; shorter waits stay buffered. Every wait line names the
attempt that runs after it.
- max_retries_exhausted_result's max_retries becomes attempts (what it
carries); a one-attempt rate-limit end gets its own lead.
- One Retry-After reader, retry_utils.provider_retry_after_seconds, shared
by compute_error_backoff and the recovery ladder.
- _session_default_route replaces _session_default_model everywhere in
tui_gateway, so a not-yet-built free-tier session reports nous/welcome on
create, reuse, resume, branch and row persist, not the silent default.
* fix(agent): free-tier cutoff no longer ends subagent or library turns, paid waits look like main (review)
A delegated child of a TUI/Desktop session and a library caller (no platform) are unattended, so they keep waiting.
Rate-limit wait lines keep main's attempt numbers and stay buffered for every provider; only a non-rate-limit Retry-After over 60 s is announced, as on main.
The unattended free-tier wait line keeps the reset time beside "The free model is busy".
The free-tier cutoff names the reset time for any cooldown it fires on (over 60 s), not only from 120 s.
session.info for an env-seeded model off the free tier carries the model alone again, as on main.
Behaviour tests for the cutoff, the subagent and unattended waits, and a paid long wait.
OpenAI's Ultrafast service tier (service_tier="ultrafast") is broadly
available for GPT-6 Astra in the API and Codex (Pro 500 / Enterprise),
billed at 6x Standard. GPT-6.1 Sol Ultrafast is announced as coming soon.
- One tier table: agent.fast_mode.parse_service_tier / STATIC_TIERS now
back the CLI, gateway and TUI config loaders (three hand-copied parsers).
- resolve_fast_mode_overrides(tier="ultrafast") sends the tier only for
Ultrafast models on api.openai.com / chatgpt.com; anything else gets
nothing rather than a silent swap to another paid tier.
- /fast ultrafast on CLI, gateway (typed + picker when supported) and TUI;
static-tier attach sites pass the tier through.
- Pricing: the response's SERVED service_tier is folded into usage (Codex
stream assembler keeps it); served ultrafast prices from a published
6x row (272K whole-request tier included). A request asking for
Ultrafast but served at default bills at Standard.
- i18n: 4 new keys in all 17 locales; option lists mention ultrafast.
When every row create of a turn loses to the SQLite lock, the queued token delta's
"ensure the row exists" guard becomes the session's first writer and minted the row as
source='unknown'. That placeholder was permanent on the real path even with the upsert
repair from #112045: the turn lease (turn_facade_lease.admit_durable_turn) treats an existing
row as proof the create already happened and sets _session_db_created, so the creator never
returns to repair it. Live probe: a platform="desktop" AIAgent whose create_session raised
"database is locked" for the whole first turn ended with a source='unknown' row on base AND
on the contributor head; with this change the row is minted 'desktop' by the guard itself.
Producer fix: update_token_counts gains an optional source= that the two agent call sites
(agent/turn_usage.py, agent/codex_runtime.py) fill from _session_source_for_agent(platform),
the same value _ensure_db_session would stamp. record_auxiliary_usage has no surface and
keeps the placeholder, which the creator's upsert now repairs.
Salvage trims: the contributor's SimpleNamespace dispatch test is replaced by a real-AIAgent
invariant test under tests/agent/ (the dispatch hunk in _run_prompt_submit is kept; the
INSERT-OR-IGNORE is idempotent under prompt.submit's own persist); narration comments cut
to the WHY; docs list 'unknown' among the startup-sweep sources.
Refs #111999
A flat per-image constant (1500 in the trigger estimator, 1600 in the tail-budget walk) is wrong in
both directions: a screenshot costs ~1,100 tokens on one provider and 4,000+ on a local mmproj
model. In a GUI loop on a 64K window the estimate sat at ~20K while the real prompt passed 80K,
so compaction never fired and the provider rejected every request (#70328).
The provider prices every image exactly on the request that carries it, so the cost is
observable from usage alone, with no vendor formula: with a fresh usage anchor, the residual
between the next real prompt_tokens and anchor + text-only delta is the price of the N images
that delta introduced.
- agent/image_token_cost.py: calibrate_from_usage() runs in record_response_usage before the new
anchor is captured; the learned value (EMA, plausibility-banded) is kept per model@host in
~/.hermes/cache/image_token_costs.json and bound per turn through a ContextVar.
- estimate_messages_tokens_rough, _content_length_for_budget (tail walk) and gateway hygiene all
read the same bound value, so trigger and walk agree; the per-message memo now caches text
tokens and image COUNT so a recalibration re-prices cached rows.
- One flat default (1500) remains only until the first vision turn; the duplicate 1600 is gone.
evals/token_accounting/ab_image_cost_calibration.py (real AIAgent, fake provider pricing images
at 4,000, one screenshot per turn, 64K window): main learns nothing (1500) and the tail walk
under-prices its own protected tail by 56.5%; this branch learns 4,374 after one vision turn
and the walk's error is +8.5%.
Reporter and first-fix credit: @JonthanaHanh (#70328, #70463).
Diagnosing the 1,393-agent run's cache misses took a state.db join and a live
probe because none of the three were on the one line we log per call:
write=<n> cache_creation tokens; a write costs 50x a read, so this is the
money, and "read stuck, write large" on consecutive lines is a
routing miss visible without a probe
id=<id> the provider's response id (Anthropic msg_..., OpenRouter/Nous
gen-...): what a provider needs to look a request up
upstream=<n> who actually served, when the route reports it (OpenRouter's
`provider`); how we learned GMI was not serving
Fields are appended to the existing line and omitted when absent, so every
existing parser (evals/postmortem, the two cache_prefix probes) keeps
matching; the forensics parser reads them when present.
The streamed chat path fabricated `id="stream-<uuid>"` and dropped the
chunks' id/provider on the floor; it now keeps the first chunk's id and
provider, falling back to the fabricated id only when the stream never sent
one. Nothing depended on the prefix.
Live (Nous, Fable 5.1, both wires): chat -> `write=3759
id=gen-1788727882-... upstream=Anthropic`; native -> `id=gen-1788727891-...`
(the Anthropic Message object's id was already real). Tests: fields present
and omitted, prefix unchanged; forensics parser reads new and old lines.
Two parallel "real usage" mechanisms fought each other: the usage anchor (real + delta) and the
compressor's rough/real projection (should_defer_preflight_to_real_usage with
last_rough_tokens_when_real_prompt_fit / _pending_request_rough_tokens / note_request_rough_estimate
baselines). The projection stored an anchored, real-scale figure as its "rough" baseline, so a
rewind that invalidated the anchor produced phantom growth and a spurious compaction (#103391).
Now there is one authority:
- Post-tool gate (turn_preflight.compress_after_tool_results): anchored figure first (the raw
last_prompt_tokens ignored the tool results just appended), then real, then rough.
- Gateway hygiene (run_turn._hmwa_hygiene_plan): real session count, else the anchor persisted on
the session row, else rough.
- Preflight / pre-API gates: an anchored figure is never deferred. A whole-context rough estimate
over threshold waits ONE request for the provider's real count instead of compressing on a guess
(first request, rewind/edit-resend, reloaded history without a persisted anchor).
- The wait is one request, never a disable: a provider that omits usage
(note_usage_less_response, #2153 class), a real reading already over threshold, a rough figure
past the whole window, and provider-proven overflow all compress immediately; the post-compaction
latch (#36718 / #104192) is unchanged.
- Projection baselines and their bookkeeping deleted (-101 LOC in context_compressor); the fixtures
that scripted whole-history estimates now state the fact they relied on (provider omits usage).
Fixes#103391 (closes#103397 by construction — the baseline it repaired no longer exists).
The usage anchor (real usage.prompt_tokens + delta estimate of what was appended since)
identified the priced transcript by id() of the last message, so it was None on EVERY
gateway turn (history is re-read from the DB each turn) and in every fresh process
(--resume, desktop per-turn serve). Those are exactly the surfaces where the bytes/4
estimate then fired local compression against payloads the provider priced far under
threshold (#99421, #104462).
- agent/usage_anchor.py owns the anchor: content fingerprint instead of id(), persisted on
the session row (model_config._usage_anchor) via set_usage_anchor(), restored on the first
resumed turn while the durable transcript still matches, cleared with the row on
compaction / codex-native rewrite / session reset.
- Callers repointed from model_metadata (the compat table follows).
Design and persistence slot from #99585 by @686f6c61; re-authored against the Sep 2026
layout (the branch predates the model_metadata / agent_init split).
fix(delegate): nested orchestrators get their workers' results back — delegate_task exempt from the 420 s tool deadline; summary budget uses current prompt, not the session sum
Portal will serve anthropic/* from more than one upstream (OpenRouter
passthrough today; GMI/Vertex once it is back online). The native Messages
wire is the better transport but is only safe where the upstream keeps
prompt-cache routing sticky: measured false on the OpenRouter path (14-20% of
consecutive calls re-write the previous turn; #104284 moved the default to
chat), untested on GMI. Hermes cannot see the upstream in the request, only
in the response: OpenRouter stamps `provider` (chat wire) and mints
`gen-<unix>-<rand>` ids; GMI/Vertex returns Anthropic-native `msg_...` ids
and no provider.
`auto` therefore starts every session on chat (correct on both upstreams),
classifies the first response, and switches that session to native only
when the upstream is GMI AND `agent/nous_wire.py::GMI_NATIVE_WIRE_CLEARED`
is True. The switch is scheduled at response time and applied at the start
of the next iteration (turn_iteration_prep), so nothing is rebuilt while a
response is being consumed; it goes through switch_model so the client,
cache policy and _primary_runtime stay consistent. One decision per session,
call 1 only; unknown upstream never switches; a failed switch logs and stays.
GMI_NATIVE_WIRE_CLEARED is False: until the 20x6 concurrency probe
(evals/postmortem/live_ab) is clean on a GMI-served anthropic/* id on the
native wire, `auto` behaves exactly like `chat`. Flipping it is the whole
rollout once GMI is measured. Default stays `chat`.
Tests (17): classifier on real Portal response shapes from both wires and
both upstreams; chat for openrouter/unknown, GMI gated on the flag; one
decision per session, call 1 only, explicit chat/native never auto-switch,
other providers/models untouched, switch failure swallowed and final;
record_response_usage on a real AIAgent invokes the hook once.
Live (auto, real Portal, Fable 5.1): arm A, real classification
(OpenRouter today) - stays on chat through a tool loop and a second turn,
cache 97-99%. Arm B, classifier forced to gmi with the flag on - call 1 on
chat, switch applied before call 2, calls 2-3 on the native wire in the
same session, tool result and both turns correct, cache 97-99%. An earlier
shape that switched inside the response path broke call 1 (SimpleNamespace
has no .content); the scheduled apply is why.
Independent review found two holes in the first fix. A parent with no usage
row yet was treated as 0 tokens used, so a 190K/200K prompt received a
384K-char dynamic summary budget instead of ~4K; the budget now returns None
(static ceiling only) when nothing is known. And under MoA the folded usage
includes advisor prompts that are not in the parent's context, over-stating
the prompt size and wrongly truncating summaries; turn_usage now records the
aggregator's pre-fold prompt_tokens as _last_prompt_size_tokens and the
budget reads that first.
Tests (2 new): unknown usage -> None; MoA-folded and unfolded parents with
the same real prompt get the same budget.
