Retain the two context-variant metadata additions by Michael Steuer. Keep the dedicated Astra reasoning contract already present in #103057 rather than replacing it with the GPT-5.6 vocabulary.
(cherry picked from commit add3a4fa6f31ea3f5fdad701a840d4f8530eb30e)
(cherry picked from commit 1672f12c260260ea38ed636528c02aee734d62b1)
Extend #104477 to the native thinking, vision, metadata, and local header paths identified by #87641. Materialize only at probe boundaries; leave the chat callable and cache ownership untouched. Local-wire A/B: thinking and vision show requests change from 403 to 200, while static credentials and callable chat retain success. Target suites queued.
Co-authored-by: liuhao1024 <sunsky.lau@gmail.com>
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).
Codex Responses reasoning and compaction items carry ciphertext the provider prices by its own
token count, never by bytes; a single native compaction checkpoint is ~5M chars, which the
bytes/4 estimator turned into ~1.29M "tokens" against a 204K threshold (#100611). #104192 deferred
that decision for one request; this removes the mis-pricing at the source so the preflight
estimator and the tail-budget walk agree (a mismatched size class protects blob-heavy rows as
"small" and compaction re-fires). Only real usage prices these items, and the usage anchor carries
that price forward.
evals/native_compaction/ab_checkpoint_preflight.py: preflight estimate after checkpoint
1,292,413 -> 58 rough tokens; the over-threshold negative arm still compresses.
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).
Six slugs land in the nous and openrouter curated lists, above the gpt-5.6 line:
openai/gpt-6-astra{,-fast,-flex} and openai/gpt-6-astra-pro{,-fast,-flex}.
Nous Portal serves the tiers as distinct slugs (verified live: each echoes its id, service_tier
default/priority/flex, cost 1x/2x/0.5x). OpenRouter serves them as ENDPOINTS of the base model
(tags openai/fast, openai/flex) and silently routes an unknown suffix to the standard tier at
standard price, so the OpenRouter profile rewrites a tier slug to its base wire model and pins
provider.only to that tier's endpoints (OPENROUTER_ENDPOINT_PINS). The base slug is pinned to
openai/azure/azure-us so default routing never lands on a flex or fast endpoint.
Provider-agnostic metadata: one DEFAULT_CONTEXT_LENGTHS entry (gpt-6-astra: 1,050,000, live on
OpenRouter for both models; substring-matches -pro and the tier suffixes). Pricing is skipped:
both routes bill via official_models_api. Reasoning floor not added (no evidence of long thinks).
No runtime consumer read the proxy (terminal_tool/environments call is_interrupted()/set_interrupt()
directly); its only users were tests patching tools.interrupt._interrupt_event, which had no effect on
the code under test. tools/terminal_tool.py's own re-export of the name is owned by another worker.
For each issue anchor present in BASE 63279301bc non-test .py and absent on HEAD, the BASE comment/docstring block was re-attached at the HEAD location of the code it explained (matched by the distinctive code line / enclosing def). Sentences already covered by an existing HEAD comment were deduped; the issue number always survives. Insert-only: no code lines changed.
commandcode (api.commandcode.ai) exposes authoritative
context_length via /models (muse-spark 1M, etc.) but as a
known provider it skipped the custom-endpoint probe at step 2
and has no models.dev entry, so every model fell through to the
256K DEFAULT_FALLBACK. Add a provider-aware branch mirroring
gmi/nous to resolve via _resolve_endpoint_context_length.
Fixes GOAT docs vs status-bar mismatch: muse-spark 1M was shown
as 256K.
Muse Spark 1.2 family (api.meta.ai) ships 1M context (models.dev
opencode/muse-spark-1.2 = 1048576, meta/muse-spark-1.2 = 1048576).
Zen/GO SG /v1/models only returns id (no limit.context), and
models.dev lookup via opencode was missing a hardcoded fallback, so
get_model_context_length fell back to DEFAULT_FALLBACK_CONTEXT=256k.
Banner showed Context: 256,000 for both zen and router-sg lanes.
Add longest-prefix entries 'muse-spark' and 'muse' = 1_048_576 so
all variants (1.1, 1.2, contributor, contributor-free) resolve to 1M
without network.
Follow-ups from review of the two salvaged #87490 commits:
- _pressure_with_real_floor now applies only on the rough fallback branch.
A valid usage anchor is provider-exact and wins as-is: on MoA turns the
anchor deliberately uses the pre-fold aggregator usage while
last_real_prompt_tokens holds the folded figure, so flooring the anchored
value would re-add fan-out tokens the anchor exists to exclude. Docstring
rewritten to describe the real path split (anchor since d3a1c46510).
- estimate_tokens_rough: encode with errors="replace". main's estimator
never raised; text.encode() on a lone surrogate (routine in tool output,
see message_sanitization) raised UnicodeEncodeError and would abort a
turn where main produced a slightly-off number.
- Record the cl100k/o200k/Qwen2.5 calibration for the bytes/4 rule.
- tests: accented Latin within +10% of the ASCII rule; mixed Cyrillic/ASCII
counts ASCII at one byte; lone surrogates don't raise; anchored pressure
is never floored (wiring shape).
The ~4 chars/token rule is calibrated for ASCII; Cyrillic, Greek, Arabic and
similar 2-byte scripts tokenize at ~2-3 chars/token, so chars/4 under-counts
them ~2x and the pre-flight pressure figure trails real usage by tens of
percent on non-English sessions. Counting UTF-8 BYTES at ~4/token uses the
encoding width itself as the corrective: ASCII is unchanged (1 byte/char),
2-byte scripts count at chars/2, and the CJK dense path keeps its explicit
~1 token/char rule with the sparse remainder byte-counted. The ASCII
isascii() O(1) fast path is preserved; the non-ASCII paths add a single
C-level encode over text that was already being regex-scanned.
Complements the last-real-prompt floor: the floor catches sessions that are
already at the ceiling, this keeps the estimate from lagging in the first
place.
_lmstudio_native_models and _apply_llamacpp_props; the router-mode child probe
shares one /v1/props -> /props fallback helper. HTTP call sequence verified
identical old vs new across 10 fake-endpoint scenarios.
_validate_cached_context_length (step 1), _resolve_bedrock_context_length (1b),
_resolve_custom_endpoint_context_length (2-3); resolution order and every log
message unchanged. Compact the per-step comment essays to their invariants.
- _server_root / _ollama_show_context / _longest_key_match / _probe_local_context_length /
_endpoint_model_entry replace 5 copies of the same local-probe and lookup bodies
- detect_local_server_type waterfall as an ordered (name, probe) table
- output-cap error classification via phrase-group tables shared by
is_output_cap_error and parse_available_output_tokens_from_error
- endpoint-scoped context overrides as a data table
- drop dead _fetch_codex_oauth_context_lengths, _resolve_codex_oauth_context_length,
_estimate_message_chars (zero refs repo-wide)
Verified with a differential harness (old vs new module, ~700 pure-function probes, 0 mismatches).
Every bot-to-bot DM is a fresh `hermes -p <bot> chat -Q` process, so it
pays agent startup on each hop. Profiling one hop showed the single
largest controllable cost was a live GET /models against the provider on
EVERY launch (0.3-0.6s normally, up to the 15s probe timeout on a slow
endpoint) — the in-memory endpoint-metadata cache is per process and the
Nous persistent context cache is bypassed by design so the portal stays
authoritative.
- model_metadata: memoize successful remote /models probes on disk
(cache/endpoint_model_metadata.json) with the SAME 300s TTL as the
in-memory cache, so authority semantics are unchanged (reconciliation
still lands within 5 minutes) but the answer is shared across
processes. Local endpoints are never memoized (LM Studio reloads).
- bot_relay: the cross-machine reply waiter polls the reply file every
250ms instead of every 2s — up to 2s of dead air on every relayed reply.
Nothing here changes turn ordering: DMs and group rounds stay serial.
Live (polis-hermes bot, spawn -> first API request, cold, 5-6 runs):
main median 1.23s (one 20.8s outlier = probe stall) -> 0.96s, no stalls.
On reasoning models a long tool loop replays the current turn's thinking +
scaffolding on every request, so the LAST request's prompt_tokens can exceed
the durable transcript by hundreds of K — all of which evaporates at the turn
boundary. The status bar and /context breakdown rendered that raw figure, so
users watched 'context' jump (e.g.) 850K -> 600K across a turn boundary and
read it as a broken compaction.
- conversation_loop: capture a turn-base usage anchor from the turn's FIRST
provider response (api_call_count == 1), where replay is minimal.
- anchored_context_tokens: new charge_stale_thinking kwarg forwarded to the
delta estimate (stale reasoning excluded on all but the newest assistant
message).
- cli status snapshot + context_breakdown: prefer the turn-base anchored
figure; fall back to last-response anchor / raw last_prompt_tokens.
- All _usage_anchor invalidation sites also clear _turn_base_usage_anchor.
Display-only: compression trigger math keeps using real last-request usage
(the inflated request is what actually risks the window mid-loop).
Run models locally as a first-class provider. The CLI grows a managed
llama.cpp runtime (engine install, model download, server supervision);
the desktop app grows the full setup and management story on top of it.
GUI surfaces ship behind the desktop --local launch flag (hermes desktop
--local, or the flag on the packaged app); backend routes and the CLI
are always live.
Runtime (hermes_cli/local_runtime/):
- curated GGUF catalog with per-machine variant selection: hardware
probe (VRAM/RAM/UMA), fit planning with spill accounting, quant choice
by context window
- derived recommendation: quality-ranked picks gated by a predicted
decode-speed floor, bandwidth-aware on unified memory; the decision
table is pinned as a test (pick AND reason per memory class), and the
Recommended badge explains its pick in a tooltip fed by the resolver's
actual branch
- engine install + model download with resumable split parts, cumulative
plan-level progress, and staged-model integrity (a split GGUF counts
only when every part is present)
- server supervision: spawn/adopt/stop, router mode with per-model load
progress relayed over SSE, abandoned-request cleanup
Desktop:
- Settings -> Providers -> Local models: one-click quickstart (install
engine, download the recommended model, boot) plus per-model download/
activate/eject, fit-ranked catalog with context pills
- model pickers (composer dropdown + Cmd+K) show staged local models,
in-flight downloads as live progress rows, and load-into-memory bars
- local-setup campaign tip for eligible hardware; System resources
statusbar widget (GPU/VRAM/RAM); in-chat load progress during sends
- friendly dead-server errors, and failed agent builds retry on the next
send instead of wedging the session
Co-developed with NVIDIA field feedback on RTX 5090 and DGX Spark.
Google Gemini/Gemma overflow errors read 'Unable to submit request because
the input token count is 32825 but model only supports up to 32768'.
parse_context_limit_from_error had no pattern for the 'supports up to N'
phrasing, so overflow recovery kept the wrong window and burned its retry
attempts instead of recalibrating to the provider-reported limit.
Add the anchored pattern (limit follows 'supports up to'; the larger input
count before it is never captured) plus regression tests covering the exact
message and the get_context_length_from_provider_error recalibration path.
Reported by @Artemonim in #57275 (residual claim 5).
Test seams and plugin engines monkeypatch estimate_messages_tokens_rough with (messages)-only signatures; route callers only pass the charge_stale_thinking kwarg on the False path.
The preflight trigger charged reasoning/reasoning_content on every assistant message while the tail-budget walks charged newest-turn-only (#73624), so reasoning-heavy codex_responses sessions fired compaction forever while the walk protected everything (middle_window_tokens=0, no_progress every turn, each attempt a full aux summarization).
Wire truth: the codex_responses input builder never ships the text thinking keys (encrypted codex_reasoning_items carry the chain and were already charged unconditionally by both sides), so the trigger overcounted reality; echo-back chat-completions families (DeepSeek/Kimi/MiMo thinking mode) replay stored reasoning_content on every turn, so there the walk undercounted. New single wire-truth predicate message_sanitization.stale_thinking_reaches_wire() now drives BOTH sides: trigger estimates exclude stale thinking on non-echo routes; tail/prune walks charge it on echo routes.
Also: reasoning/reasoning_content double-count fixed in both estimators (wire ships at most one; +53% overcount vs provider prompt_tokens per issue comment), and the commit-layer no_progress path now arms the structural no-op backoff so an unchanged-transcript compaction cannot re-fire every turn (defense in depth; overlaps the #96775 re-entry class).
Follow-up for salvaged PR #96939: main already added hy4-preview at
1_048_576 (f7c79efbac); the cherry-picked duplicate key later in the
dict silently overrode it with 1024000.
Every provider response carries usage.prompt_tokens — exact ground truth
for the full request (system prompt + tool schemas + history). Context-size
checks now anchor on the last main-loop response's usage and estimate only
the messages appended since, instead of re-estimating the whole history
with chars/4 heuristics and flat 1500-token image costs. The estimate error
window shrinks from the entire conversation to one turn and self-corrects
at every response.
- agent/model_metadata.py: capture_usage_anchor() / anchored_context_tokens()
with a structural base-message identity check that fails closed on any
transcript rewrite.
- agent/conversation_loop.py: anchor captured at the single main-loop usage
site (MoA uses pre-fold aggregator usage; advisor/aux calls never anchor);
pre-API pressure check prefers the anchor.
- agent/turn_context.py: preflight compression estimate prefers the anchor.
- agent/context_breakdown.py: /context display prefers the anchor.
- Invalidation: compaction rewrite (conversation_compression), codex native
compaction (codex_runtime), session reset/switch (run_agent), plus the
fail-closed structural check for splices/micro-compaction.
- Usage-less responses keep the previous anchor; no anchor -> pure
estimation fallback (first request of a session).
Live on both providers (verified 2026-08-28 against openrouter.ai/api/v1/models
and inference-api.nousresearch.com/v1/models) but absent from both curated
picker lists. Adds the entry directly below qwen3.8-max per newest-first
family ordering, an explicit 1M DEFAULT_CONTEXT_LENGTHS entry (new family
slug would otherwise fall through to the generic qwen 131072 catch-all —
same class as #69881), and regenerates model-catalog.json.
Scoped rollout: only the named providers touched. Pricing snapshot skipped
(both routes bill via official_models_api live pricing). Reasoning floor
already fires via the qwen3 prefix entry (180s, verified).
- move minimax/minimax-m3:free into the Free tier section (house
convention: :free SKUs group together, matching glm-5.2:free and the
nemotron :free entries) and regenerate model-catalog.json
- add Inkling family context length (1,048,576 — OpenRouter live
metadata, 2026-08-27) to DEFAULT_CONTEXT_LENGTHS; new family slug
otherwise fell through to no entry
- add Inkling to the reasoning stale-timeout floor table (300s tier,
same as Grok reasoning / Ox Alpha; OpenRouter marks the family as
reasoning-capable)
- widen the floor matcher's right-anchor separator class to include
':' so OpenRouter SKU suffixes (:free/:batch/:nitro) inherit the
family floor — inkling:free previously missed the inkling entry
- regression tests for the inkling floor + ':' separator
The two local-server context probes in _query_local_context_length read
data.get("max_tokens") as a context-window candidate. On an
OpenAI-compatible /v1/models passthrough max_tokens is the max OUTPUT
tokens, so a 1M-context model advertising a 128K output cap resolves to
128000 and auto-compaction fires ~7x early.
Route both branches through the module's own key vocabulary
(_CONTEXT_LENGTH_KEYS), which already classifies max_tokens as a
_MAX_COMPLETION_KEYS entry.
Local /v1/models probes treated Anthropic `max_tokens` (max output) as the
context window when `max_model_len`/`context_length` were absent. Anthropic
and Anthropic-compatible reverse proxies expose both:
max_input_tokens = context window (e.g. 1M for claude-fable-5)
max_tokens = max output (e.g. 128k)
That under-reported windows (1M → 128k), persisted the wrong value into
context_length_cache.yaml, and fired compression at ~96k (75% of 128k).
Route model objects through a shared helper that prefers input-window keys
via _extract_context_length, and only falls back to max_tokens when no
input-window field is present.
The image-routing vision path calls detect_local_server_type without
the provider's API key. Against a remote API-keyed endpoint (sglang /
vLLM with --api-key) every leg of the 5-request probe waterfall came
back 401 — and because a failed verdict was never written to the
in-memory cache (only positive verdicts were), the waterfall re-ran on
EVERY image-bearing turn (#89863: 51 detail-less busy-acks observed in
one Slack channel while the probe sprayed the user's own server).
Two changes:
- image_routing._should_probe_ollama_vision now takes the API key and
forwards it; a new _resolve_inference_api_key mirrors
_resolve_inference_base_url's resolution order (runtime value,
model.api_key, providers blocks) so the key always matches the URL
being probed.
- detect_local_server_type caches a None verdict in memory with a short
failure TTL (5 min, vs 1h for positives) so the next turn is served
from the negative entry instead of re-running the waterfall — while
a transient failure (server starting, key being fixed) recovers in
minutes. Negative verdicts are deliberately not written to the
cross-process disk cache.
Review findings on #92797 (@100yenadmin):
- is_codex_900k_base() is now the single source of truth used by picker
synthesis, context resolution, /model validation, and wire stripping.
Eligibility is an exact table (sol/terra/luna, gpt-5.4, daybreak alias)
plus date-shaped 5.6 snapshots — family-prefix matching removed, so
non-routable -pro slugs and unknown descendants never gain variants.
- strip_codex_context_variant_suffix() strips conditionally: ineligible
aliases (gpt-5.5-900k) are returned unchanged and fail honestly at the
API instead of silently running as the base model at 272K.
- validate_requested_model() rejects ineligible *-900k aliases before the
hidden-slug soft-accept, and accepts valid variants missing from a
stale catalog without letting the typo auto-corrector eat the suffix.
- Codex context resolver drops vendor/ namespaces, so
openai/gpt-5.6-sol-900k resolves to 900K like the bare id.
- Table-driven regression covering eligible bases/snapshots/namespaced
ids and rejected -pro/-mini/5.5/unknown aliases, asserting context AND
wire model.
The Aug 16 change that auto-raised gpt-5.4/5.6 Codex OAuth context to the
live-verified 900K burned through subscription usage for users who never
asked for the larger window (bigger window = more input tokens per request).
- Base Codex slugs (gpt-5.6-sol/terra/luna, gpt-5.4) now resolve to the
advertised 272K again — the cheaper limit is the default.
- The model picker synthesizes explicit <slug>-900k variants (e.g.
gpt-5.6-sol-900k) for every live-verified slug; selecting one opts into
the 900K window. Slugs that genuinely enforce 272K (gpt-5.5,
gpt-5.4-mini) get no variant.
- The -900k suffix is Hermes-side only: stripped before the model id hits
the wire (main transport + auxiliary Responses adapter), and pricing
aliases the variants onto the base entries.
- Docs: new opt-in section in context-compression-and-caching.md.
GLM-5.3 is live on api.z.ai (coding plan endpoint) but had no entries in
Hermes, so it silently fell back to the generic 202K GLM context —
triggering premature context compression on a 1M-window model.
- model_metadata: 'glm-5.3': 1_048_576 (same base model as 5.2; 1M
context / 128K max output per docs.z.ai/guides/llm/glm-5.3, verified
2026-08-14)
- auth: add glm-5.3 to coding-plan probe lists (global + CN)
- models: add glm-5.3 to picker/model lists (6 sites)
- zai provider: reasoning_effort mapping covers glm-5.3 (accepted live
by the endpoint, HTTP 200)
Michael Steuer