A rejected OAuth token (HTTP 401 'User not found' from Nous Portal, Codex,
xAI…) reached the desktop error card as 'Provider error' with Retry as the
first action, which just replays the same dead credential.
Backend: nonretryable_client_error_result dropped failure_reason /
failure_retryable, so error_surface classified every non-retryable 4xx as a
retryable provider failure. It now stamps the classifier verdict like the
max-retries path, and auth-layer descriptors carry auth_kind
(oauth|api_key, derived from the provider catalog tab) + provider_label.
Desktop: an auth/oauth surface renders 'Authentication error', explains that
the <provider> sign-in expired/was revoked, and offers 'Sign in to <provider>
again' which launches that provider's existing onboarding OAuth flow scoped
to the failed session's gateway profile. Retry stays as the follow-up click.
Re-login to the provider already in use keeps the current model instead of
swapping in the recommended default.
Follow-up to #106866 from the post-merge audits (@ehz0ah, @victor-kyriazakos):
- _recoverable_pool_provider: the "endpoint mismatch, not a dead key" shield now (a) compares the full
origin via base_url_origin() (scheme+host+port — a different port or an HTTPS->HTTP downgrade is a
different trust boundary) and (b) applies only when the rejected client carried the SESSION's key, so
an independently owned auxiliary pool keeps rotating at its own configured origin.
- _resolve_named_custom_branch: explicit/per-task base_url and api_key compose OVER the providers.<name>
entry field-by-field instead of the entry silently replacing them (URL-only, key-only, both, neither).
- _acreate_with_progress: plain-create fallback only for a rejected stream NEGOTIATION (mirrors the sync
wrapper); a mid-stream failure propagates to the classified recovery ladder instead of re-sending the
whole prompt non-streaming.
Tests (all red on main): foreign origin x {host, port, scheme} shields the session key; same origin and
independent aux pool still rotate; named-provider override matrix through resolve_provider_client on a
real temp HERMES_HOME; async content-then-error is not retried non-streaming.
`auxiliary.<task>.provider: openai` was rewritten to custom + api.openai.com/v1 unconditionally, and the
api-key discovery rung used the registry default endpoint, so proxy/gateway users (OPENAI_BASE_URL or
providers.openai) had compression hop to the public endpoint with a proxy-issued key -> 401 -> the
key/pool was quarantined and the session sat over the compression threshold.
- _expand_direct_api_alias: a providers.openai entry keeps its name (named-custom branch applies its
base_url/key); otherwise OPENAI_BASE_URL wins over the public default.
- _resolve_api_key_provider: when the bound session runtime is that provider, use its endpoint + key.
- _recoverable_pool_provider: a rejection at a host other than the session's configured endpoint for
the same provider is an endpoint mismatch, not a dead key -> no rotation / unhealthy mark on the pool.
Complete the seam default from the salvaged commit: _relay_async_completion needs an async twin of
_create_with_progress, and the async primary attempt previously streamed only for stream-only
providers, so a hooked async compression call ticked the watchdog zero times.
Adds one invariant test: with a hook installed, BOTH relay defaults stream and tick per chunk;
without a hook they are byte-identical plain creates. Red on origin/main.
* feat: add session-scoped connector access for onboarding
* fix(connectors): availability is the config flag AND the portal entitlement — no free-tier leg
The port carried a third availability leg from hermes-magic: a stored guest
(free-tier) identity short-circuits the managed-tool entitlement check. That
leg reads hermes_cli.anon_auth, which does not exist on hermes-agent main, so
connectors_available() raised ImportError inside its fail-closed try and the
whole connector surface was silently dark on a plain upstream checkout.
On this tree availability is the two-leg AND the design started with:
tools.connectors.enabled AND managed_nous_tools_enabled(). The free-tier leg
is a hermes-magic concern and belongs in hermes-magic's own delta over this
branch, next to the identity it depends on. Its integration test goes with it.
* docs(tool-search): connectors section — remote tools through the bridge
The squashed port carried the code but not the user-facing docs. Restores the
Connectors section of the Tool Search page and the connector-gateway host /
CONNECTOR_GATEWAY_URL override on the Tool Gateway page, updated for the
manage_connections tool and the pure-connector batch rule.
* fix(tool-search): connector tools rank with local tools in one pass instead of taking leftover slots
dispatch_tool_search ran BM25 over the local catalog, filled `limit` slots,
then appended connector hits only into slots left empty. On a 300-tool
catalog no slot was ever empty, so with Gmail and Google Calendar connected
"send gmail email" returned five betterstack tools and zero connector tools.
The gateway's hits for a query now become catalog entries (connector name,
slug words, description as the search text) and join the local catalog for
that query's BM25 pass. One ranking, one rarest-token admission rule for both
sources, `limit` as the total per query. The merge loop and the separate
record builder for connector hits are gone; `_shared_tool_record` serves both
sources.
The gateway search timeout rises from 8 s to 30 s. One request with six
use_cases measured 7 s, so 8 s sat on the edge and cut real answers off; the
failure path is unchanged (local-only results, no error to the model).
Live, 311 local tools + gateway, before -> after:
"send gmail email": 5 betterstack tools -> gmail SEND_EMAIL, CREATE_EMAIL_DRAFT
"read google calendar events": 5 betterstack tools -> googlecalendar EVENTS_LIST_ALL_CALENDARS
"linear create issue", "betterstack incident": unchanged
Benchmark (25 labelled queries): connector recall 0.09 -> 0.82, precision@5
0.18 -> 0.59, false positives on absent intents 17 -> 2.
* refactor(tool-search): connector leg into tools/connector_search.py
tools/tool_search.py is a facade. The connector leg (gateway hits as catalog
entries for tool_search, remote schemas for tool_describe, the
connections_in_scope gate) was appended to it by the port. It now lives in
its own sibling, tools/connector_search.py, and the facade imports the three
entry points: connections_in_scope, connector_entries_by_group,
remote_schemas_for.
No behaviour change. The tool_describe remote block became
remote_schemas_for(names, current_tool_defs, connector_describe) with the
same inputs, the same silent-degradation contract and the same injection
seam the tests already use.
* fix(tool-search): at most 7 queries per call, the gateway's search limit
One tool_search call sends all its queries to the connector gateway as one
search request. The gateway answers 7 use_cases per request and returns
HTTP 502 for 8 or more (measured 2026-09-09, re-measured with one-word
use_cases: it is a count limit, not a size limit). With the client cap at
10, a model sending 8 to 10 queries lost every connector hit for that call
and saw local-only results with no error.
The shared constant splits: _MAX_QUERIES_PER_CALL = 7 for search,
_MAX_DESCRIBE_NAMES_PER_CALL = 10 for describe, which has no remote count
limit. Eight or more queries now get the existing "too many queries" retry
hint before any request is made. No chunking: one call, one request.
* fix(tool-search): the model is told that connectors__ names are manage_connections accounts
tool_search results carry names like connectors__gmail__CREATE_EMAIL_DRAFT and
manage_connections is the tool that checks and connects those accounts, but
nothing told the model the two are the same thing. A model that hit
CONNECTION_REQUIRED had to infer the fix on its own.
The tool_search description gains one sentence making the link, added at
assembly only when manage_connections is in the session's tools. Signed out
or with connectors off the tool is absent and the description is unchanged,
so it never names a tool the model cannot call. This follows the existing
rule for cross-tool references (tools/AGENTS.md): they are added dynamically
from the session's actual tool set, never hardcoded in a schema.
Tool defs are fixed for the life of a conversation, so the description is
byte-stable per conversation; this is a one-time prefix change.
Live, real get_tool_definitions() against a signed-in home: sentence present.
Same home with auth.json removed: manage_connections absent, sentence absent.
* fix(connectors): /stop halts a connector batch before the next remote call
dispatch_connector_batch runs every remote entry of a tool_call batch in
sequence. The executor only checks the interrupt flag between tools, and
the whole batch is one tool to it, so a /stop landing during entry 1 of
20 still sent the other 19 to the gateway.
The loop now reads tools.interrupt.is_interrupted before each dispatch.
Once set, it stops calling handle_function_call and fills every unstarted
slot with the loop's existing error-slot shape, code INTERRUPTED and the
message "Stopped by the user before this call was made.", so the result
envelope stays valid and the counts stay honest. Entries already
dispatched keep their real results.
Test: three connector calls where the fake client sets the interrupt on
the first execute. The client sees exactly one call and slots 2 and 3
carry INTERRUPTED. Red on the base branch, green with the fix.
* test(connections): schema assertions become dispatch contracts
test_schema_documents_wait_and_its_timeout froze description fragments
("REQUIRED", "can NOT disconnect", "Nous Portal"). A wording edit fails
it while a real regression (a disconnect that reaches the gateway) does
not. That is a snapshot of prose, not a behaviour contract.
Delete it. The requirement that wait needs connectors is already covered
by test_wait_requires_connectors. The user-only disconnect boundary is
now asserted as behaviour: action disconnect with a connector returns an
error and the fake client records no call. That replaces the earlier
de-authenticate test, which only checked that the word "dashboard"
appeared in the error text.
Test count in the file goes from 26 to 25.
* docs(tool-search): connector batches are one gateway request per entry
The user guide said a connector batch travels as one gateway request. It
does not: model_tools_connectors.dispatch_connector_batch re-enters core
dispatch per entry, and each entry becomes its own execute request in
bridge._run_remote (plus at most one literal-slug retry when the gateway
reports TOOL_NOT_FOUND under the conventional slug). The docstrings in
tools/tool_gateway/bridge.py and tools/tool_gateway/__init__.py still
described the abandoned V1 plan and claimed nothing outside the package
imports it.
Rewrite those sentences to match the code: one request per entry, in
input order, dispatched from model_tools_connectors.py, with the per-entry
approval and interrupt behaviour that motivated the split. The guide also
still showed the single-call shape tool_call(name, arguments); both
places now show the `calls: [{name, arguments}]` array the schema
advertises and note that a single local call is an array of one.
Docs only, no test.
* fix(tools): the between-turns refresh never rewrites the bridge tools
The per-turn MCP refresh folds a fresh tool snapshot into the live array
with preserve_prefix: order and membership stay, but a name present in both
takes the fresh schema. That is right for ordinary tools, whose schema is a
constant. tool_search is the one tool whose description is derived from the
session: the deferred-tool count, the embedded listing, and, on this branch,
whether manage_connections was present. A late MCP server or one failed
portal lookup (manage_connections' check_fn fails closed) changed those bytes
on the next turn, and every byte after tool_search in the cached prefix was
re-prefilled. The array also contradicted itself in that case: the flapping
manage_connections was carried forward while the description lost its hint.
The bridge entries now keep the bytes they were built with for the life of
the conversation. Nothing is lost: tool_search reads the live catalog at
dispatch, so tools that arrived late are still found; connector availability
is checked at dispatch too. The compaction-boundary rebuild (content_aware,
the one sanctioned cache break) still refreshes the description.
Consequence: connector exposure in the prompt is decided once, at agent
build, by whether the user was signed in then. That is the intended
contract.
* refactor(tool-search): normalize_tool_call_entries lives with the other argument validation
The port appended the tool_call argument parser to the tool_search facade.
The family already has tools/tool_search_validation.py for exactly this
work (schema validation of deferred call arguments), so the parser moves
there and the facade imports it. No behaviour change; the one test that
imported it now imports from the defining module.
* refactor(connectors): delete the unused batch dispatcher; _run_remote becomes run_remote
bridge.dispatch_calls and its helpers (_dispatch_calls_inner, _run_pre_dispatch,
_run_local, _error_slot, _maybe_parse_json) and the LocalDispatch / PreDispatch
seams had no production caller. Connector dispatch runs through
model_tools_connectors: dispatch_connector_batch re-enters handle_function_call
once per entry, so scope, hook, approval and middleware policy fire against each
composed name inside core dispatch, and dispatch_connector_call hands the single
planned entry to the bridge's transport function. Only tests called the batch
dispatcher, and they exercised policy seams that production never wires.
The transport function is the module's real entry point, so it drops the
underscore: _run_remote becomes run_remote, body unchanged. The module
docstring now describes the two legs that exist (availability with D32 silent
degradation, and run_remote) instead of the injected seams. Imports that only
the deleted code used are gone; merge.py is untouched because every export
still has a caller.
Tests that drove dispatch_calls are deleted where they covered the removed
seams (pre_dispatch blocks and rewrites, local_dispatch classification, mixed
batches). The literal-slug fallback, the per-entry transport failure, and the
hook rewrite reaching the gateway request body are re-targeted at
handle_function_call('tool_call', ...) with the fake client swapped in at
bridge._default_client_factory, the same seam test_connector_dispatch_policy
uses. Each re-targeted test fails when the retry is disabled in run_remote.
* fix(connectors): search keeps the twin a colliding name reaches, and says so
format_connector_name strips the toolkit prefix, so GMAIL_FETCH_PROFILE and a
literal FETCH_PROFILE on gmail both compose to connectors__gmail__FETCH_PROFILE.
describe and execute decode that name to the prefixed slug first, so the
literal twin is unreachable under it. If a vendor ever shipped both, search
could describe the literal under a name that runs the prefixed tool.
Search is the one place that sees both twins in one response. It now keeps
the twin the name reaches and drops the other with a WARNING that names both
slugs, whichever the gateway listed first. Short names stay; no marker, no
per-process map, no change to describe or execute. No such pair exists in the
live catalog today; the guard turns a silent alias into a logged one.
_is_azure_foundry_responses matches the azure-foundry provider or the
project-scoped services.ai.azure.com host. #101243 proposes narrowing it
to that host alone. The multi-item rejection this branch handles happens
on resource-level *.openai.azure.com hosts too, so the pruning gate now
uses _is_azure_responses: the provider, or either Azure host.
Refs #105369
The first request of every turn after the first replays the encrypted
reasoning item of every earlier response. Azure Foundry accepts one such
item and rejects two or more with HTTP 400 "Conflicting authenticated
continuation identities". In-turn requests never hit this because
_is_post_tool_replay already suppresses replay on Azure when the request
ends on a tool result.
On Azure Foundry, build_kwargs now keeps codex_reasoning_items only on the
newest assistant row that has any. include still requests
reasoning.encrypted_content, compaction checkpoints stay on every row, and
the canonical messages are not mutated. Other Responses endpoints are
unchanged.
Rebuilding a failed session from state.db through the transport and
sending it live: unpatched, 5 reasoning items, 400; patched, 1 item, 200.
Fixes#105369
Azure Foundry (gpt-6-astra) answers a request that replays encrypted
reasoning from more than one prior response with HTTP 400 "Conflicting
authenticated continuation identities", code invalid_value. The classifier
mapped that to a non-retryable client error, so the turn aborted and every
later turn in the session failed the same way.
Map the message to invalid_encrypted_content. The existing recovery in
turn_recovery then strips the replay state and retries once.
Refs #105369
The classifier fix routes Codex patch-budget 400s into the shrink
recovery, but _image_error_max_dimension still returned None for the
Codex wording, so the recovery fell back to the 8000 px default cap
and skipped images between ~5542 and 8000 px that already exceed the
30000-tile budget — burning the single shrink retry without shrinking
anything. Parse the reported patch limit and convert it to a per-side
pixel cap of isqrt(limit)*32 (5536 px for 30000 patches), which keeps
a square image under the budget.
OpenAI Codex Responses rejects an image whose tile-patch budget exceeds
its 30000-patch ceiling with wording ("requires N patches after
processing, exceeding the limit") that contains none of the existing
image-size vocabulary, so the 400 fell through to format_error
(non-retryable). The reactive image-shrink recovery in turn_recovery was
therefore bypassed and the session kept failing — or failover re-sent
the identical oversized image to another model.
Route the patch-budget wording to image_too_large (retryable) so the
existing shrink pass re-encodes the image under the ceiling and retries.
Fixes#106337
finish_reason='length' has two causes: the answer was long (max_tokens reached),
or the prompt itself left no room to generate. _continue_text treated both the
same: append the fragment + a continuation nudge and retry, up to 4 times. In the
second case every retry sends a strictly longer prompt, so each attempt is worse
(Ollama n_ctx=32768: 32,638 -> 32,685 -> 32,732 prompt tokens, all truncated), the
user is told "model hit max output tokens", and max_tokens is not the lever.
The response's usage already carries prompt_tokens and the compressor already
resolves the model's context window; compare them once per truncation. Under
_MIN_CONTINUATION_HEADROOM (512) free tokens the turn ends on the first
truncation, keeps the partial text, names the context window as the cause and
points at /compress or a larger window. Unknown usage or window keeps today's
behaviour. max_tokens semantics untouched.
Co-authored-by: gaoanze888 <214786078+gaoanze888@users.noreply.github.com>
Co-authored-by: KoNit-K <124019182+KoNit-K@users.noreply.github.com>
The two facades post-#102117 (chat_completion_helpers, agent_runtime_helpers) must not grow
behaviour; the marker/predicate belong with the sibling that already owns primary-cooldown
state shared by the fallback walk and restore_primary_runtime. Also drops the two
try/except-return-False wrappers around pool.entries() and normalize_model_for_provider
(neither raises for a live pool / the never-raising normalizer).
Tests trimmed to the two invariants (#106475): the marker fires only on the single-credential
Codex entitlement 400 and the walk skips the rejected slug in either form; restore is gated on
a rejected primary and still restores an unrelated one.
A Codex ChatGPT-account 400 ('The X model is not supported when using Codex
with a ChatGPT account.') names the model, so with a single credential the
slug is dead for that account. The fallback walk still re-selected it and
restore_primary_runtime switched back to the primary at the start of every
turn, announcing an unverified 'Primary model restored' — the two warnings
alternated forever with zero delivered answers (#106475).
Record the rejected (provider, model) pair on the non-retryable client-error
path (only when no multi-credential pool exists — rotation covers that case,
#71970), skip rejected entries during the fallback walk, and gate
restore_primary_runtime on the primary's slug so the session fails closed
with the terminal entitlement error instead of oscillating. Fixes#106475.
(cherry picked from commit 471435b10288f15387b2549a8b02551d82c0f670)
Hermes Console ran each command on a ThreadPoolExecutor and cancelled only
the asyncio waiter. A command that forks an AIAgent (`curator run
--consolidate`) kept its worker thread and its in-flight provider request
alive after the prompt said "cancelled" — a llama.cpp generation kept
decoding for 30+ minutes and held the inference slot (#106179).
Root cause: the host owning the thread never knew about the agent created
deep inside the synchronous command, so it could not call the existing
cooperative `interrupt()` path that closes the request sockets.
Fix: `agent/interrupt_scope.py` gives the host an `InterruptScope`; the
console binds it around the worker (ContextVar), and every
`AIAgent.run_conversation()` registers itself with the bound scope for the
turn. On cancel, timeout and disconnect the console calls `scope.cancel()`
(hard-interrupts every registered agent; an agent registering after the
cancel is interrupted on entry so the cancel cannot lose the race with a
turn that has not started) and awaits the worker Future with a 10s bound
before reporting cancelled/timeout. Queued-but-unstarted work is dropped via
`Future.cancel()` alone.
Live repro (fake OpenAI-compatible provider blocking like llama.cpp, real
/api/console, real curator dispatch, real AIAgent + direct request path):
origin/main at the "cancelled" frame -> request_exited=false,
worker_exited=false; with this change -> both true, provider saw the peer
close, prompt reported cancelled 0.27s after the frame.
Salvage of #106197 by @kyssta-exe (executor-future handle, cancel/timeout/
disconnect propagation) and #106320 by @Xixiartemis (deterministic
lifecycle regression: terminal(cancelled) => no owned request or worker
remains live; interrupt-on-late-registration). Both rebuilt slimmer: #106197 keyed its fallback on Future.cancel()
returning False, but the handle it held was run_in_executor's asyncio
wrapper, whose cancel() returns True while the thread keeps running, so its
thread-name abort registry was never consulted; #106320's command-scoped
ownership model is folded into one small module hooked at the turn facade
instead of a per-caller `bind_agent`.
Co-authored-by: kyssta-exe <218078013+kyssta-exe@users.noreply.github.com>
Co-authored-by: Xixiartemis <182932319+Xixiartemis@users.noreply.github.com>
Novita returns HTTP 429 with message 'server overload, please try again
later' and error type 'server_overload' when its server is genuinely busy.
Neither phrase was in _OVERLOADED_PATTERNS, so the 429 fell through to
_V_RATE_LIMIT and set should_fallback=True + should_rotate_credential=True —
rotating the credential / falling back early instead of retrying the same
key. Add 'server overload' and 'server_overload' to the overload tuple so
this reaches the existing _V_OVERLOADED verdict (retryable, no rotation).
Closes#106205
An aux task pinned to a private endpoint via `base_url` (a home Ollama
box at `byron.local`, a LAN IP, localhost) is the intended per-task
endpoint feature and can never bill a provider. The Settings → Model
banner still counted it as "still run on openai" forever and offered
"Reset all to main", which would wipe the working local setup; the
post-switch `stale_aux` report had the same blind spot; and the aux row
never showed the `base_url` the backend already sends, so the pin was
indistinguishable from a paid-provider pin.
- `GET /api/model/auxiliary` now stamps each task with `local_endpoint`,
the verdict of the one canonical classifier
(`agent/model_metadata.py::is_local_endpoint`) — no TS mirror of the
private-range rules, so frontend and runtime cannot drift.
- Desktop: the persistent banner filter is the pure
`staleAuxAssignments()` and skips `local_endpoint` pins; the pinned row
appends ` · <base_url>` when one is set.
- `_stale_aux_pins` (post-switch report) skips local pins the same way.
- `is_local_endpoint`: `*.local` (RFC 6762 mDNS) now counts as local, and
IPv6 literals no longer ride the "no dots ⇒ unqualified host" rule, so
a global-scope address (`2607:f8b0::1`) is not local while `::1`,
ULA and link-local still are via the `ipaddress` scope checks.
Slim redo of #106236 (@webtecnica) and #106234 (@huklaa), which fixed the
same symptom with a client-side classifier copy; the bug class, row
display and mDNS/IPv6 classifier corrections are theirs.
Refs #106228
Co-authored-by: Hukla <129692708+huklaa@users.noreply.github.com>
Co-authored-by: webtecnica <webtecnica@gmail.com>
When the redirect cap trips, the correction that cancelled the final attempt is
still sitting in _pending_redirect; finalize_turn's clear_interrupt() would drop
it silently. Drain it into the steer slot so it rides result["pending_steer"] and
becomes the next user turn on every surface that already honours that key.
Both new exit reasons get a turn-completion explanation so the user sees why the
turn stopped instead of an empty reply.
The redirect and rebuilt-for-fallback restart paths in apply_retry_restarts
refund the iteration budget and re-issue the iteration with no per-turn
bound. A redirect/interrupt that keeps re-arming the flag refunds forever,
so the turn loop never exits and the durable session turn lease is held
indefinitely (concurrent processes block up to LEASE_WAIT_SECONDS).
Add a per-turn restart_count accumulator (threaded through _run_phase like
the other loop locals) and break out once it exceeds max_retries, matching
the bound the compression path already has.
_sum_clarify only extracted the top-level ``user_response`` key, so batch clarify
results (questions=[...] -> responses[].user_response) fell through to the generic
placeholder and the summarizer never saw the user's answer/permission decision.
Closes#106077.
The overflow-terminal path ends the turn without reaching finalize_turn, so
a transcript that overflowed right after a tool batch ended on a raw tool
result; strict providers reject the next user turn (tool -> user). Close it
with the same final text, mirroring the truncated-tool-call terminal above.
Also: classify once before either log so an overflow no longer emits a
"so the loop can continue" WARNING followed by the contradicting "NOT
seeding" one; reset the stale-streak breaker once for both branches; drop
the "compression could not recover it" wording (this path never reached
compression); trim the test file to the three tests that bind behaviour
(stream -> terminal stub; 413 stays non-terminal; the terminal ends the
turn, closes the tool tail, carries compression_exhausted).
Address review P1s (andrexibiza) on #106266:
1. The overflow-terminal exit in recover_from_truncation now forwards the
#98722 typed compression_exhausted bit (partial_result/end_turn gained the
flag) so the gateway resets/moves future input to a clean session instead
of leaving the bloated durable session authoritative for the next turn.
2. _overflow_terminal is scoped to FailoverReason.context_overflow ONLY.
payload_too_large (413) has its own byte-scored recovery owner
(turn_overflow._recover_payload_too_large, #88960/#47339) that must not be
bypassed; a post-delta 413 keeps its normal continuation stub. Regression
covers both lanes.
Tests: unit asserts result compression_exhausted=True on the marker; a
real streamed partial hitting a 413 payload-too-large error keeps content and
is not terminal. 50 streaming/continuation/gateway regressions pass.
When a stream delivered text and then died on a context-overflow /
payload-too-large error, the partial content (often tens of KB) was seeded as
a length-continuation stub, growing the transcript monotonically. In a
session whose transcript cannot be compressed back under budget
(protect_last_n covers everything -> no_progress, or the summary would
itself be larger -> would_grow), every later request is larger than the one
that just failed — an unrecoverable loop where the user sees a 30+ minute
fake hang and the only remedy is killing the session (#106260).
classify_api_error already labels these errors context_overflow /
payload_too_large (should_compress=True). _partial_stream_stub now returns
an EMPTY stub marked _overflow_terminal for that class instead of seeding
the recovered text, and recover_from_truncation treats the marker as
terminal: the turn ends via the recovery contract with a clear message
(start /new) and the transcript is not polluted with the partial.
Normal partials (network stall, output-cap truncation, tool-call drops) are
unchanged — only the overflow error class changes behavior.
Tests: stub marker + empty content; a real streamed partial hitting a
'maximum context length' error returns the terminal stub; recover_from_
truncation ends the turn (no fragment/nudge appended) on the marker while a
normal stub still runs the continuation path. 67 streaming/continuation
regressions pass.
_maybe_inject_run_budget_wrapup() appends its wrap-up notice to the newest
role:"tool" message in place, with no _DB_PERSISTED_MARKER check. Its sibling,
_maybe_inject_iteration_budget_warning(), got exactly this guard added in the
same recent saga (turn_iteration_prep.py), with the comment "an older turn may
already be cached."
The reachability is structural, not an edge case: _maybe_inject_run_budget_wrapup
is only ever called from prepare_iteration(), at the START of the next iteration
-- strictly after tool_executor.py's _flush_session_db_after_tool_progress has
already flushed and marked the previous iteration's tool row persisted. So every
successful injection was mutating an already-persisted row: the wire request for
that turn carried the notice, but the durable transcript never did, diverging
replay from the live bytes and invalidating the provider's prompt-cache prefix
from that row onward.
Fix:
- Add the same _DB_PERSISTED_MARKER guard to _maybe_inject_run_budget_wrapup,
scoped to the specific tool row the reversed scan lands on (not just
messages[-1], since this function -- unlike its sibling -- scans backward for
the newest tool row rather than only checking the tail).
- Wire _maybe_inject_run_budget_wrapup into _flush_session_db_after_tool_progress
(pre-flush), mirroring exactly how _maybe_inject_iteration_budget_warning is
wired in both places. Without this, the guard alone would make the notice stop
firing in the common case, since prepare_iteration's call site almost always
hits an already-persisted row -- the pre-flush call site is what actually lets
it land in durable bytes.
Verified empirically: read the real call graph (tool_executor.py's three
_flush_session_db_after_tool_progress call sites cover every tool-completion
path) to confirm the guard's premise, then added an end-to-end test using a real
AIAgent + SessionDB that flushes and checks the persisted row for the notice
text. Mutation-verified: reverting the two production files drops exactly the 2
new/updated assertions (28 pass, 2 fail); reapplying restores green (30 passed).
Also ran the sibling iteration-budget-warning and /steer suites (71 passed) to
check for interaction regressions -- none.
Drop the bare "tool.content" pattern (any 400 mentioning tool.content in a
non-list context would be sent through the image-strip path) and the profile
flag snapshot test; the behaviour tests (classifier verdict + proactive
downgrade) already pin the contract.
- Declare `supports_vision_tool_messages=False` and `supports_vision=True` on `opencode_go` provider profile in `plugins/model-providers/opencode-zen/__init__.py`
- Route HTTP 422 errors through `_IMAGE_TOOL_RULES` and add `tool.content.str`, `tool.content`, and `input should be a valid string` patterns to `_MULTIMODAL_TOOL_CONTENT_PATTERNS` in `agent/error_classifier.py`
- Add unit tests for OpenCode Go proactive tool result downgrade, HTTP 422 Console Go classification, and profile capability contract in `tests/run_agent/test_multimodal_tool_content_recovery.py` and `tests/plugins/model_providers/test_opencode_go_profile.py`
The ledger in verification_evidence.db exists only to feed the verify-on-stop
guard, but the recorder kept running on every foreground terminal command and
every file edit after #53552 turned the guard off by default. Users who never
opted in still accumulated a multi-MB database (7 MB / 4.6k rows on one install).
Every ledger entry point (record_terminal_result, record_verify_run,
mark_workspace_edited, verification_status) now checks verify_on_stop_enabled()
first and returns without opening or creating the database when the guard is
off. verification_status reports {"status": "disabled"} in that case; no client
consumes the verification.status RPC yet, so nothing downstream changes.
Existing ledger tests pin HERMES_VERIFY_ON_STOP=1 since they exercise the ledger
itself; the new test proves the off path never creates the file (red on base).
Follow-up to #106317. Typing the steer row (display_kind="steer") for the renderer and the
alternation-repair guard collided with the convention that any display_kind on a user row means
scaffolding: is_user_originated_turn / _is_actionable_user_turn / split_user_originated_turn
returned False for it (tail anchoring, auto-focus, dispatcher views, resume counts) while
_is_real_user_message returned True (anchor restoration) — the two predicate families disagreed
on the same row, and list_recent_user_messages (/undo, /rewind) skipped it in SQL. A steer
carries full user authority; the steer kind is now whitelisted in all four.
Also: the pre-API drain's requeue tail reuses _requeue_pending_steer instead of a copy; the TUI
history projection compares against STEER_DISPLAY_KIND; the steer() docstring describes the row.
Follow-up to #106312. _preflight_timeout_result carries the prior history without this turn's
user row (#7100); with a repeated prompt ("continue") the verbatim scan resolved to the
historical copy and exported it as this turn's proven boundary — the exact relabeling the export
exists to prevent. Nothing is exported for that envelope now.
The trailing `agent._persist_user_message_idx = idx` ran after finalize_turn had already flushed
the transcript, so it never influenced a persist and the next turn reset it: dead state, removed.
Both steer sites now build the row through one helper, prompt_builder.steer_user_row:
a role:user row with display_kind="steer" and no leading blank lines. The alternation
repair (_merge_consecutive_users) skips a steer-typed prev row, so a run that ended
right after a steered batch (Ctrl-C, interrupt) does not get the next real prompt
merged INTO the already-persisted steer row — which would have rewritten it in place
and re-broken live≠replay parity, the exact class this PR fixes.
TUI/desktop history projects the steer row as the user's own words instead of the
model-facing marker wrapper; 'steer' joins the display_kind union. The compression
anchor scan keeps its tool-row branch for transcripts persisted before this change and
its docstring says so.
Second site of the same bug class #104444 fixes in apply_pending_steer_to_tool_results:
_inject_steer_into_newest_tool_result (the drain that runs when a /steer lands during an
API call) mutated the newest role:tool row in place. That row was already flushed
append-only, so the replayed history diverged from the live request bytes at the
injection point and broke the prompt cache exactly like the post-batch path.
Deliver it the same way: a standalone user row inserted right after the newest tool
result (not yet persisted, so the next flush writes it to the transcript). Restash when
there is no tool row yet, unchanged. Stale comments claiming steer lands "in the newest
tool result" and agent/AGENTS.md's alternation rule now describe the real shape.
`apply_pending_steer_to_tool_results` used to smear the steer text onto
the last `role:tool` message's content. That tool row had already been
flushed to the session store and carries `_DB_PERSISTED_MARKER`; the
append-only persistence never rewrites it, so the replayable transcript
diverged from the live request bytes at the injection point — resumed
sessions (surface switch / process restart / background-review close)
missed the provider prompt cache (75-85% hit) and the user's mid-run
instructions were never part of the durable history.
The steer is now emitted as a standalone `role:user` message (marker
text preserved):
- role alternation stays legal: assistant(tool_calls) -> tool -> user is
the documented 'user jumped in mid-run' pattern that
`repair_message_sequence` deliberately keeps;
- the appended dict carries no `_DB_PERSISTED_MARKER`, so the next
`_flush_messages_to_session_db` writes it to the session store —
transcript bytes and replayed history finally agree, and the steer
becomes searchable/retrievable like any other user message;
- the no-tool-result fallback (interrupt) still requeues the steer, which
the caller then delivers as a normal next-turn user message.
Tests: TestSteerInjection updated for the new shape plus a persistability
assertion (no marker => flushable); tool-batch-segmentation malformed
scenario updated. steer + segmentation suites: 67 passed, 1 skipped.
The salvaged regression test exercised only repair_message_sequence and
reanchor_current_turn_user_idx — pre-existing helpers — so reverting the fix left it
green. It now drives prepare_iteration on a real AIAgent with adjacent user rows and
asserts the returned index addresses this turn's row and mirrors into
_persist_user_message_idx (red without the re-anchor: IndexError).
Both re-anchor sites (repair and compression restart) call
turn_context_compaction._reanchor instead of inlining "reanchor + mirror", so they
cannot drift. The export tests fold into one parametrized invariant plus the
run_conversation envelope test; the WHAT-restating comment shrinks to the WHY.
Hosts that settle their own transcript by index (hermes-webui) cannot prove which
row of result["messages"] is the current user turn once this loop rewrote history
(alternation repair, compaction, post-turn micro-compaction): the instance-side
_persist_user_message_idx predates those rewrites, and a text match relabels an
identical historical prompt and claims its old answer. Only the producer can
assert the coordinate against the exact list it returns.
run_conversation now wraps the turn (_run_conversation_turn) and stamps the pair
through export_current_turn_boundary on every envelope that leaves the loop
(success, partial/error, interrupt, retry-exhausted, tool-limit, preflight
timeout, codex runtime), computed on the final messages after finalize_turn and
micro-compaction. The pair is exported only when the addressed row is this turn's
user message verbatim (reanchor's last-match rule); a rewritten row exports
nothing so hosts fail closed. The final index is mirrored into
_persist_user_message_idx for the persist override.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013gp366ijf39n4UUtJhZuMh
prepare_iteration() runs repair_message_sequence_with_cursor() before each API
call; the repair merges adjacent user rows in place (after a compaction, the
role=user summary sits next to the protected first user message). The loop's
current_turn_user_idx was recorded at turn start, so after a merge it points
past the current user row: the per-turn context injection (prefetch/plugin
context) silently misses it, and hosts that settle the transcript by this index
(hermes-webui) write the current user turn to the FRONT of the context —
rewriting the prompt's leading messages every turn (0% prefix-cache hits at
200K+ tokens, ~100 s re-prefill per turn) and duplicating the user's question.
The in-loop compression restart path already re-anchors; do the same after a
repair that changed the list: reanchor_current_turn_user_idx (last user row
carrying this turn's text), return the index through the IterationPrep verdict
so the loop state picks it up, and mirror it into agent._persist_user_message_idx,
which hosts read when the result carries no index. The new phase parameters
default to None so direct callers keep their signature.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013gp366ijf39n4UUtJhZuMh
The salvaged commit forked an explicit /refine under a new "refine_review" origin so
the memory delete gate would not treat it as unattended. But is_background_review()
is the key for every other review guard — skill_manager_guards (curator-owned-only,
read-before-write), skill_manager_tool (archive instead of rmtree), skill_ledger
actor, write_approval staging, the [auto] tag — so a /refine fork silently escaped
all of them.
Carry attendedness separately: the fork keeps origin "background_review" and sets
_review_attended; turn_context binds it beside the origin ContextVar; the memory
gate keys on the new is_unattended_review(). Also run the gate AFTER
_validate_single_op / the operations list check, as memory_tool's own docstring
requires, so an invalid replace is rejected now rather than staged and failed at
approve time.
Review follow-up on #105944 (#105921):
- explicit /refine forks now run under the refine_review write origin
(explicit flows from the CLI/gateway handlers through
_spawn_background_review_now and spawn_background_review_thread down
to build_cache_parity_fork), so a user-requested review keeps the
full memory operation set; only automatic reviews stay behind the
unattended delete gate.
- the unattended delete gate now stages the denied replace/remove (or
whole batch) into the pending store instead of dropping it: the
fork's own review summary is never published, so a plain denial lost
the consolidation request with no surfacing path. The staged proposal
carries a proposal_staged marker that summarize surfaces as an action
line, and a staging failure still fails closed to a plain denial.
- regression tests: explicit-path origin pass-through, refine_review
keeping replace working, near-limit denial end to end (add rejected
by budget -> replace staged -> proposal surfaces, store unchanged).
The review fork's tool whitelist granted the whole memory toolset
whenever the profile had memory enabled, regardless of which nudge
fired, so a skill-nudge fork held remove/replace on MEMORY.md it was
never asked to use; combined with the memory tool's near-limit
'consolidate now' hint, an unattended fork deleted standing rules with
no user in the loop.
- Pass review_memory from spawn_background_review_thread through
_run_review_in_thread/_run_review_fork into _review_tool_whitelist;
a skill-only review no longer gets the memory tool at all.
- Fail-closed operation gate in memory_tool: a background-review fork
may add, never replace/remove (single or in a batch) — consolidation
decisions reach a human via the review summary instead.
- Keep the deny/prompt wording in sync with the whitelist so a
memory-less review doesn't advertise memory.
Fold the identical heappush(...) into PeriodicScheduler._requeue; notify() instead of
notify_all() now that the scheduler thread is the only condition waiter; drop the
PR-history paragraph from the module docstring (the commit carries it).
Tests: the blocked-sibling test asserted `sibling_ran.wait(0.30)` — a wall-clock bound
under the repo's ≥2 s flake floor; it now asserts the sibling fired while the blocker
still held its worker. The worker-start-failure fake keys on this scheduler's own
_run_callback rather than the global thread-name prefix so a leaked handle on _DEFAULT
cannot consume the single injected failure. The base-green no-overlap test is dropped
(it does not prove the fix).
oMLX/MLX prefill memory-guard rejections name an allocation peak in BYTES but
close with "Reduce context length", so _CONTEXT_OVERFLOW_PATTERNS claims them
and the turn enters the compress-and-shrink loop. Compression cannot lower a
prefill peak — the prompt is usually far below the window — so it burns the
compression budget, re-hits the wedged server on every attempt and ends in
"Cannot compress further" plus a destructive session reset.
Classify them as `overloaded` instead: retry with backoff, no compression, no
session reset (mirrors 503/529 recovery).
The guard runs before the overflow check AND before the usage-limit
disambiguation. The second ordering matters more than it looks: "memory limit
exceeded" contains "limit exceeded", so a status-less rejection — a proxy that
flattened the body — is currently classified as `billing` and rotates a
healthy credential.
Sites covered:
- _OVERFLOW_AS_5XX_RULES, which _400_TAIL_RULES extends → 400, 500, 502,
503, 529
- _MESSAGE_HEAD_RULES for the status-less path (ahead of usage-limit)
- _ERROR_CODE_VERDICTS for the structured oMLX codes
- _classify_400, because _by_status runs before _by_error_code, so a body
whose wording a proxy stripped would otherwise fall through to
format_error
Every pattern names memory/allocation in bytes, never a token or window count,
so the list stays disjoint from _CONTEXT_OVERFLOW_PATTERNS. Both oMLX wordings
are kept: 0.5.6 says "Prefill would require ~13.87 GB peak", 0.5.7 reworded it
to "predicted peak would require/exceed" and both are in the field. A test
pins that a genuine window overflow still compresses.
Refs #52261. Supersedes #52289, which predates the classifier rewrite and can
no longer be merged.
agent.tools is always a list (agent_init assigns it from
get_tool_definitions) and every entry is a well-formed function schema,
so the isinstance ladder over parent/entry/function/name guarded shapes
that cannot reach this helper. Use the same two lines agent_init uses;
`or []` keeps the empty-surface contract from the previous commit.
Docstring cut to the WHY (the between-turn refresh note described the
other guard). Tests trimmed to the two invariants: inherited tools
survive the compaction-boundary refresh (deep-copy isolation folded in),
and an empty parent surface is copied and frozen. Literal sentinel
asserts replaced with the constant.
Copy and freeze review_agent._tool_snapshot_generation even when parent.tools is an empty list ([]). Previously, the truthiness check (\
ot parent_tools\) caused an empty parent tool surface to be skipped, allowing newly available late MCP or plugin tools to be retained on the review fork and leaving its snapshot generation unfrozen. This broke the byte-parity contract when no tools were active on the parent.
Returning early only when parent_tools is not an instance of list or tuple guarantees that an empty tool snapshot is faithfully inherited and frozen. Adds dedicated regression test test_unrouted_review_fork_inherits_empty_tool_surface.
Freezes review_agent._tool_snapshot_generation to _FROZEN_TOOL_SNAPSHOT_GENERATION
(2_147_483_647) when inheriting the parent tool surface for same-model cache parity.
When in-place compaction boundaries trigger refresh_agent_mcp_tools(content_aware=True),
the staleness guard in _publish_tool_snapshot refuses the rebuild (snapshot_generation < published_gen),
preventing agent.tools from being reconstructed from the raw registry and preserving
inherited memory-provider and late tools across compaction boundaries (#103579).
Adds unit regression test verifying tool preservation across content_aware refresh.
Co-authored-by: joaomarcos <joaomarcosdias444@gmail.com>
Ensure unrouted background_review forks inherit the parent's full advertised
tools[] surface. Without this, skip_memory=True caused memory-provider tools
(e.g. fact_store/fact_feedback) and dynamically injected plugin/late MCP tools
to be omitted from the fork's tools array, breaking byte-exact prefix-cache parity
and incurring full cold-read costs on providers where tools are part of the cache key.
Inheriting the full parent tools array preserves complete prefix cache parity
while execution dispatch remains strictly bounded by the thread tool whitelist.
_session_db is always a SessionDB (agent_init / delegate_tool), so the
"fail closed on a store wrapper" hasattr was defense around code that
cannot fail; the store's own guard binds the value into SQL, so the
prologue only needs the sibling idiom isinstance(_row_id, int) that
session_persistence and transcript_repair already use. The positional
hazard is explained once, on set_latest_user_api_content. The in-place
compaction test duplicated test_api_content_sidecar's
test_inplace_compaction_backfills_sidecar_into_db verbatim (its row_id
parameter was never varied); dropped, as was the positional-helper tail
of test_older_identical_row_is_untouched already covered there.
The turn-start stamp had grown its own copy of the "what does the current
user row hold" rule (persist override = clean transcript, live content =
wire bytes = sidecar when they differ) that _db_flush_row already
implements. Two copies drift; extract durable_user_row_content() in
session_persistence and call it from both.
Also: reuse _persist_lock() instead of a third open-coded lock/nullcontext
ladder; drop the hasattr guard on set_latest_user_api_content (it predates
this fix and exists on every SessionDB); cut the comment to the WHY;
trim the new test file from 18 cases to the 7 invariants (real close
flush E2E, repeated-"ok" positional protection, API-only pre-flushed
turn, normal path writes nothing, compaction keeps positional, store
guards). Still 3 red / 4 green when agent/turn_context.py is swapped
for main's copy.
_stamp_api_content_sidecar read _row_id without holding
_session_persist_lock. A close/early flush holds that lock while it
commits the row and only afterwards writes _row_id back onto the live
dict; a stamp that ran in between saw no id and skipped the backfill,
the flush finished with api_content = NULL and marked the message
persisted, and the turn-start persist skipped it — the row kept the
wrong bytes with no writer left to fix it.
Run the _row_id read and the DB backfill under the (re-entrant) lock,
re-checking _row_id after acquiring it. Race reported by @ehz0ah on
Co-authored-by: sal <141555468+salch-cred@users.noreply.github.com>
#102411; same fix shape as @salch-cred's follow-up on #103721.
The api_content sidecar ('persist what you send') preserves prompt-cache
stability across turn boundaries by persisting the exact API-bound bytes
(including memory-manager prefetch, plugin injections, and API-only notes)
and substituting them on replay.
When a user turn was already materialized in the database before the
sidecar could be composed (in-place preflight compaction or a close/early
flush racing the prologue on the CLI path), the turn-start crash persist
marker-skips that message. Previously, the backfill was gated strictly on
in-place compaction (_preflight_compressed and _last_compaction_in_place),
so racing CLI flushes left api_content = NULL in SQLite and broke prompt
caching on subsequent turns (#102194).
Positional approaches (such as #102239 and #102286) using LIMIT 1 on the
newest active user row are unsafe: repeated common inputs ('ok', 'yes',
'continue') cause the backfill to match and overwrite the PREVIOUS turn's
row with the new turn's sidecar, corrupting history and breaking cache parity.
Resolve all landing blockers and review feedback from #102411:
1. Bounded state owner (Sahilvishnaliya):
Add SessionDB.set_message_api_content(session_id, row_id, content, api_content)
to SessionMessagesMixin in hermes_state_messages.py instead of growing
hermes_state.py. Update set_latest_user_api_content docstring with durable
warning on the positional hazard.
2. API-only turns & durable content selection (ehz0ah):
When a pre-flushed clean input has an API-only difference (e.g. voice
prefix or model-switch note):
- Retain the differing API-facing bytes as api_content even when no
new memory or plugin context was injected.
- Derive the durable content guard using _override_replaces_content so
the SQL 'content IS ?' guard matches the clean override text stored
in the DB row rather than the restored wire text.
3. Turn prologue gating (_row_id) & fail-closed store duck-typing (ehz0ah):
In agent/turn_context.py::_stamp_api_content_sidecar: check _row_id on
the live user dict (stamped by _insert_message_rows and synced by
sync_flushed_message_markers). If valid (positive int, not bool), address
by exact ID. Do NOT fall back to positional matching when a row ID is
present: if an external or custom wrapper lacks set_message_api_content,
fail closed and skip rather than corrupting a neighbouring row. If absent
but in-place compacted, fall back to positional update. On normal turns,
skip the backfill entirely (single atomic INSERT).
4. Real lifecycle test coverage (salch-cred, ehz0ah):
Comprehensive tests in tests/agent/test_api_content_row_addressed_backfill.py
covering store guards, surrogate scrubbing, gate non-arming, older identical
row protection, real close-flush row_id synchronization, API-only clean
override preservation with exact wire replay, and duck-typed store fail-closed
verification when set_message_api_content is absent.
Fixes#102194.
Closes#102411.
- _transcript_row_texts re-implemented agent.message_content.flatten_message_text
and the api_content sidecar rule; the note can only land on a user row,
so the transcript scan now skips assistant/tool rows (the bulk of the bytes).
- Three sites computed "names of agent.tools"; tools.mcp_tool_agent gains
agent_tool_names() used by the switch note and conversation_loop, which
also stops importing the private _def_name across modules. The name list
is only captured when a switch was announced.
- split_runtime_boundary() is the single owner of the runtime-block
rpartition/END check for both identity_line_value and
_stored_prompt_matches_runtime.
- platform_surface_hint was a public alias of _platform_hint; the function is
now platform_hint (its docstring pointed at the pre-move module).
- consume_gateway_turn_context_notes and consume_surface_switch_note share
_pop_turn_note so the two one-shot channels have identical semantics.
- platform check hoisted above the transcript scan.
Teknium