A background memory/skill review (agent/background_review.py) forks a
second, complete AIAgent in a daemon thread that deliberately shares the
live agent's own session_id for prompt-cache warmth. Nothing previously
stopped a user's next live turn from starting while that fork was still
mid-conversation, letting both stream against the same session_id and
credentials concurrently. That produced two observable failures:
- Doubled prompt-token accounting on the live turn's own calls (the two
concurrent request/response streams under one session_id confuse the
token-usage bookkeeping), triggering premature context compression.
- A lockup that a normal interrupt could not clear: the review fork is a
fully independent AIAgent with its own _interrupt_requested flag, and
was never added to the parent's _active_children list -- the only list
AIAgent.interrupt() actually walks for cross-agent cancellation -- so a
live-turn Ctrl+C had no propagation path to it at all.
Fix, three files:
1. agent/agent_init.py -- add _background_review_agent /
_background_review_lock tracking state to every AIAgent, mirroring the
existing _active_children pattern.
2. agent/background_review.py -- the review fork now registers itself on
the parent's _active_children right after construction (reusing the
same list/lock interrupt() already fans out to for real subagent
delegation), and unregisters on every exit path (success, the
tool-whitelist finally, and the outer exception safety-net). All
registration is defensive (getattr/try-except) so an AIAgent built
without going through agent_init.py's setup degrades to "no
cross-turn cancellation" instead of aborting the whole review.
3. agent/conversation_loop.py -- at the very start of every
run_conversation() turn, if a prior background review is still
in-flight, it is now proactively cancelled via interrupt() before the
live turn proceeds -- fire-and-forget, non-blocking, adds no latency.
Adds 3 regression tests to tests/run_agent/test_background_review.py,
confirmed to fail against the pre-fix code via a scripted revert.
Verified: ruff clean on all touched files; 66/66 background-review and
interrupt-propagation tests pass; 256/256 across turn_finalizer +
run_agent regression suites; no fork-only symbols in the diff.
Follow-ups from review of #82049:
- extract append_user_instruction() into agent/skill_commands so the
stable-prefix construction cannot drift between the skill and cron
builders (the registered prefix must stay a byte-prefix of the built
message); cron no longer imports the private _SINGLE_SKILL_INSTRUCTION
- add the startswith guard to the skill builder registration site,
matching the stronger cron guard
- rename _MAX_BYTES to _MAX_CHARS (sum(map(len, ...)) counts characters,
not bytes) and correct the comment
- collapse find_stable_prefix's two-lock dance into a single critical
section (scan is <=32 short-circuiting startswith calls, measured
2-4us; drops the snapshot copy and the TOCTOU re-check)
- document the split-shape lifetime (marked-endpoint window) in the
module docstring
- add a contract test for the helper's byte-prefix invariant
(mutation-checked)
Follow-up review of the builder-declared cache boundary (#81867) found three
ways the split could silently stop paying off, or keep paying more than it
should, on a long-lived gateway process.
Flattening no longer consults the registry. `strip_anthropic_cache_control`
matched the decorated split by looking the first block up in the prefix
registry, so a mid-turn failover that re-decorates a request built many
messages earlier (#72626) would fail to flatten once _MAX_ENTRIES newer
scaffolds had been registered in between, and would hand the next provider
the two-part shape instead of the canonical string. The split is now matched
by its shape: a marker on the *first* part of a user message is something no
other decoration produces (list content otherwise gets its marker on the last
part, and the two-part [static, volatile] split is role-gated to system), so
the ""-join stays provably byte-exact without any process state. This drops
`is_registered_stable_prefix` and one lock acquisition per stripped message.
Lookups now refresh LRU position. A scaffold fired every minute by cron could
be evicted by a burst of one-off skill invocations while still being the
hottest prefix in the process, silently reverting it to whole-message caching.
Registration now also evicts by total retained bytes (4 MiB). Entries hold
whole expanded skill bodies, so a 32-entry cap alone does not bound memory.
The newest entry is always kept, so a single oversized scaffold still gets a
boundary instead of disabling the split.
Tests: eviction-then-failover round-trip, LRU refresh on hit, byte-cap
eviction, and oversized-single-entry survival.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Webhook/cron skill invocations concatenate a large static scaffold
(activation note + expanded skill body) with a small volatile tail
(ticket payload, timestamps) into one user string, and the Anthropic
cache planner marked that whole string as a single atomic block — so a
few changed tail bytes forced a full cache rewrite on every invocation.
Instead of re-parsing scaffold marker strings out of the message at
request time (fragile when a payload or skill body quotes the marker),
the builders now register the exact stable-prefix bytes in a small
process-local LRU registry at construction time. The cache planner
splits a registered user string into [marked stable prefix, unmarked
volatile tail] request-locally; canonical session history stays a plain
string, and the failover stripper flattens the split back byte-exactly
via an O(1) registry lookup. Unregistered messages keep the existing
whole-message policy.
Covers the single-skill builder (webhook + slash command + TUI) and the
cron job prompt assembler (multi-skill, bundles, skipped-skill notice),
with registration guarded against injection-scanner sanitization.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Folds the model-switch fix in with the untitled retry. They answer
different halves and each is wrong alone: counting alone left a session
that merely opened with machinery nameless forever, because nothing
reconsidered it, and the stored title alone would never title at all on a
store too old to report one. Skip only when both agree — past the opening
turn, and already named.
Counting a turn now judges a multimodal one on its text, so "here's a
screenshot, fix the login" counts as the question it is rather than
reading as machinery and undercounting the conversation.
Co-authored-by: yy28 <yy28@vip.sina.com>
Switching models before sending the first real message titled the session
"[System: The active model for this chat has…" instead of the user's actual
question.
`_append_model_switch_marker` persists its notice with `role="user"` because
strict OpenAI-compatible providers reject a system message that is not first
(#48338). Titling had no way to tell that apart from a genuine opening turn,
which caused two distinct failures:
1. `_MACHINE_PREFIXES` did not cover the marker. Its `[System: ` prefix
matches none of `[CONTEXT COMPACTION`, `[Runtime note:`, or `[SYSTEM]`
(different case, no closing bracket), so `is_titleable_user_message()`
returned True and the marker was formatted into the title.
2. `maybe_auto_title()` counted the marker as a user message. With the marker
present, the first real question arrived at `user_msg_count == 2` and the
`> 1` guard returned early, so the session was never titled at all and its
`title` column stayed NULL. Fixing only (1) would therefore have traded a
wrong title for a permanently missing one.
Add the marker prefix to `_MACHINE_PREFIXES` (kept in sync with
`tui_gateway.server._MODEL_SWITCH_MARKER_PREFIX`) and count only titleable
user messages when detecting the opening turn.
The guard stays narrow: ordinary user text that happens to start with
"[System:" still titles normally.
Adds 6 regression tests, verified to fail without the fix.
Two lookalike gaps found auditing the titler.
_MACHINE_PREFIXES missed the compressor's legacy summary opener and the
"[System note:" injections, so a compacted or resumed session could be
named after the note that carried it. Take the summary prefix from the
compressor that emits it rather than keeping a fourth local copy.
The fast-model exclude list covered embedders but not the other non-chat
siblings a provider names after its chat model — "gpt-4o-mini-tts"
satisfies the "-mini" rung and cannot answer a prompt.
An opener is not always titleable — an image with no caption, a compaction
handoff, a bare slash command — and those sessions stayed unnamed for
life, because the guard that stops re-titling a named session also stopped
the nameless one from ever asking again. Let a later turn name a session
that still has no title.
The derived title also ran the collision dedupe inline on the turn.
It is a slice of the user's own words, so it collides constantly — people
open sessions with "hi" — and resolving "hi #47" is a widening scan on the
critical path for a name the model replaces a second later. Decline it
there and let the background stage, which can afford the scan, pick it up.
The fast-model picker reads /v1/models to find the small model a provider
currently serves, and it asked anonymously. Most of those endpoints need a
key, so the fetch 401'd and the empty result read as "this provider has no
small model" — the picker fell back to its curated list and never noticed.
Worse, a failed fetch cached its empty result forever, so one bad moment
during startup disabled live model discovery for the life of the process,
and the processes that read this run for weeks. Give the failure an expiry
and pass the provider's credentials.
The bare family rungs (-mini, -flash, haiku) also picked whichever id
sorted first, which is the oldest generation a provider still serves:
gpt-3.5-mini over gpt-5.4-mini, claude-3-haiku over claude-haiku-4.5.
Compare the digit runs as numbers so the rung meant to keep us current
does.
Titling is two-stage — a slice of the user's own words lands inline, the
model's version replaces it a second later — and the platform rename lanes
fired on both. That is two rate-limited calls to reach one name, and
Discord allows two channel renames per ten minutes, so the throwaway could
be the one that survived. The callback now carries which stage it is, and
the lanes take the model's.
The relay lane also asked where the reply landed at title time, which is
before the model has answered: it polled the send-result cache for ten
seconds and read the timeout as "never auto-threaded", so any turn with
tool calls in it silently kept its raw thread name. Wait on the send
itself instead — the adapter already owns that cache, so it can say when a
reply arrives and, just as usefully, that one arrived carrying nothing.
The turn prologue titles every session, and it is shared by every agent —
including the ones no person is reading. A cron job already names its own
session after the job in its finally block, so the titler spent a side-LLM
call per fire to write the delivery scaffolding over it for the length of
the run. A delegated child's session is hidden from every picker, so a
batch at max_concurrent_children paid N title calls for N names nobody
opens.
Both are the same class of run that already sets skip_memory to stay off
the auxiliary path, so keep the titler off it too.
_PROVIDER_PREFIXES was a hand-maintained frozenset, so providers that ship
as plugins (bundled like fireworks, or user plugins under
$HERMES_HOME/plugins/model-providers/) were never recognised as
provider: prefixes in model strings, and metadata/context-window lookups
received the unstripped string. Mirror the _URL_TO_PROVIDER auto-extend
that already sits below it: add each registered profile's name and
aliases after discovery. The _OLLAMA_TAG_PATTERN guard keeps model:tag
strings intact.
Fixes#66106
The #70716 regression fix changes popen_start_new_session from False to
True in the systemd-scope branch. Update the assertion in
test_wraps_in_systemd_scope_when_supervisor_and_available and the
docstring in test_systemd_post_spawn_failure_never_kills_gateway_process_group.
/home and /Users hold home directories; neither is a workspace. A session
whose cwd was one of them got promoted to its own auto project, so the
sidebar listed a lowercase "home" row beside the synthetic Home bucket.
Both POSIX spellings are excluded on every host — macOS ships an empty
/home autofs stub, and container/remote shells hand back Linux paths — as
are the filesystem root and the parent of $HOME.
The existing test asserted supports_draft_streaming returns True with
rich_messages=True and rich_drafts=False, but the PR's gate now makes
it return False. The test already force-sets _use_draft_streaming=True,
so the assertion was redundant — updated to reflect the new behavior.
Also removed redundant manual attribute overrides in test 3 where
_make_adapter(extra={'rich_messages': False}) already sets the flag.
When rich_messages is on and rich_drafts is off, transport=auto used
sendMessageDraft (MarkdownV2 tables→bullets) then finalized via
sendRichMessage. Users saw a crooked first bubble and a second wiki-style
final. Decline drafts in that config so auto uses edit-in-place + rich
finalize on one message.
Fixes#78524
Agent Plugins v1 packages with 'streamable-http' mcp.json entries now load
through Hermes' existing URL-based MCP client instead of being reported and
skipped. The stdio-only limitation was the agreed follow-up slice from
PR #81196.
Boundary rules from the v1 spec (§7.2.1) are enforced:
- URL must be absolute http(s), no user information, no fragment; plain
HTTP only for localhost/loopback hosts.
- Configured package headers are never forwarded across a cross-origin
redirect: translation marks entries strict_redirect_headers, and the
redirect hook in the native runtime strips those headers (plus
Authorization) whenever a redirect leaves the original origin. On mcp <
1.24.0, where the client cannot hook redirects, such servers fail closed
with an actionable upgrade message.
- Legacy 'sse' entries remain reported and skipped.
The redirect hook is extracted into a testable module-level factory
(_make_redirect_header_stripper); default behavior for native config
servers is unchanged (Authorization-only stripping).
The native Relay pipeline binds its Futures to the event loop that
entered run_in_session_async. While a managed tool callback executes,
that loop is blocked until the callback returns — so a nested managed
relay call made from inside the callback (vision_analyze's auxiliary
LLM call on a worker-thread loop) awaits a Future that can never
complete: 'RuntimeError: Future attached to a different loop', or a
deadlock, plus 'Event loop is closed' at shutdown when the orphaned
future completes late. (#77244)
Fix: managed_callback_guard, a ContextVar depth marker set around every
Hermes callback the relay adapters hand to the native pipeline
(relay_tools.execute invoke, relay_llm execute/execute_async invoke,
ManagedLlmStream run_callback). resolve_execution_context returns the
no-relay triple while the marker is set, so nested calls run unmanaged.
The marker propagates through contextvars.copy_context() into the
worker threads tools use for their internal async work.
Top-level turn LLM calls and tool wraps stay fully managed — verified
live: vision_analyze works under active shared metrics while the main
turn still records managed llm.execute events.
Alternative fixes considered and rejected: removing
retain_managed_execution (kills the shared-metrics managed pipeline)
and gating on main-thread identity (managed tool wraps legitimately
run on the run_agent thread, so that gate disables relay everywhere).
The Anthropic SDK's streaming accumulator builds ParsedMessage snapshots
whose ParsedTextBlock content doesn't match the generic union pydantic
expects, so model_dump() on stream events (message_stop) emits
PydanticSerializationUnexpectedValue UserWarnings straight into the
user's CLI output mid-response.
Pass warnings=False at every helper that dumps arbitrary SDK models
(relay_llm/_jsonable, relay_tools/_jsonable, anthropic_adapter
_to_plain_data, run_agent _hook_jsonable, chat_completion_helpers
extra_content/reasoning_details sites, chat_completions transport),
with a TypeError fallback for duck-typed model_dump implementations.
Adds regression tests including a precondition test that proves the
fixture still trips the warning without suppression.
Two cache-key correctness follow-ups to #82229 (review feedback):
1. Abbreviated commit pins are immutable too. The installer's
is_valid_commit() accepts 7-40 hex chars, but the Python refresh
exempted only exactly-40-hex names — an abbreviated pin like
install-4ce1994.ps1 could be overwritten with a branch script. The
predicate now mirrors the Rust rule (7-40 hex = immutable, never
rewritten), applied to the sanitized target ref.
2. Refresh only the update-target ref's cache key. The helper rewrote
EVERY mutable-ref entry with the active checkout's script: with
install-main.ps1 and install-bb_gui.ps1 coexisting, updating main
replaced both with main's script — cross-branch cache poisoning in
the other direction. It now computes the single cache key for the
branch being updated, using the installer's own ref sanitization
(sanitize_ref: non [A-Za-z0-9._-] -> '_', so bb/gui ->
install-bb_gui.ps1), and touches nothing else. Entries the
bootstrapper never wrote are not created.
The branch is threaded from the existing `branch =
_resolve_update_branch(args)` in both _cmd_update_impl call sites and
_update_via_zip (main-only by its own guard).
Regression tests lock down both invariants: abbreviated-SHA pin
untouched (including when passed as the branch), coexisting mutable
refs (main refresh leaves install-bb_gui.ps1 byte-identical),
sanitize_ref parity, and uncached-ref no-op.
E2E on the incident machine's real bootstrap-cache: planted a stale
install-main.ps1 + sibling install-bb_gui.ps1 + abbreviated pin
install-4ce1994.ps1; refresh("main") healed main byte-exact and left
both others untouched; refresh("4ce1994") was a no-op. The pre-existing
40-hex pin entry in the real cache was also untouched.
Pre-#67193 hermes-setup binaries (June 2026 and earlier, including the
newest published build) resolve bootstrap-cache/install-<branch>.ps1 by
"exists -> reuse forever": a branch-ref cache entry written at install
time is never re-downloaded, so every GUI update/repair executes a
months-stale install script. The binary has no self-update path, so no
amount of `hermes update` fixes it.
Live incident (2026-08-09, ryanc): install-main.ps1 cached June 4 lacked
the #81327 venv process-tree sweep; the bootstrap venv stage died with
"Cannot remove item venv\Scripts\python.exe: Access denied" on a
straggler backend pair, twice, despite every relevant fix already being
merged on main - the installer simply never ran that code.
Fix: `_refresh_bootstrap_cache_scripts()` runs at the end of every
update path (git, zip, already-up-to-date repair), overwriting mutable
branch-ref cache entries with the freshly pulled scripts/install.ps1 /
install.sh. The stale binary's unconditional reuse becomes a feature: it
"reuses" a file the update keeps permanently current. Post-#67193
installers re-download on every run anyway, so this is a harmless
pre-seed of identical bytes for them.
Scope guards: 40-hex commit-SHA entries are immutable pins and are never
touched; .ps1 gets the UTF-8 BOM to match the installer's cache format
(#67193); best-effort - a failed refresh never fails the update.
E2E on the incident machine: poisoned the real
bootstrap-cache/install-main.ps1 with a stub, ran the real function -
healed byte-exact to the checkout's script (BOM intact, #81327 tree-kill
sweep present).
Follow-up to #82179 addressing helix4u's review comment
(#82179 issuecomment-5229441571). Three parts:
1. Desktop teardown (salvaged from #77436, @4adwentures): the update
hand-off's releaseBackendLock() sent SIGTERM to the primary backend
BEFORE taskkill /T. If the launcher exits first, Windows can no longer
enumerate its descendants and they survive holding the venv — the
Electron path that creates the orphan #82179 then has to repair.
New stopBackendTreesForUpdate() tree-kills the live root first, with
the behavioral vitest from #77436. The scanner half of #77436 is
deliberately NOT taken (superseded by #82158's full-cmdline scan).
2. Tree-aware orphan classification: _orphaned_desktop_backend_pids()
previously refused the whole holder set when any holder had a live
parent. But the scanner legitimately returns an orphaned serve root
AND its descendants (the venv trampoline's uv-managed interpreter
worker — which carries the same backend argv — plus .hermes-runtime
children). Those have a live parent: the orphan root itself. Now
holders inside an accepted orphan root's tree fold into that root
(only roots are returned; taskkill /T reaps descendants), and
live-parent backends defer to the ancestry check instead of refusing
outright. Anything outside an orphan tree still refuses.
3. Tests for the mixed shapes: root+managed-runtime child,
grandchild depth, non-descendant stray alongside an orphan root
(still refuses), descendant exited mid-classify.
E2E on a real Windows box: spawned a detached backend-shaped orphan
that itself spawned children (3 python descendants); the scanner-shaped
mixed holder set classified to [root], taskkill /T reaped root and all
descendants. The live Desktop backend on the box still classified None
(refusal preserved). The first E2E attempt caught exactly the
trampoline/worker case the mocks missed — the live worker re-execs with
the same backend argv and a live parent — which is what part 2 fixes.
Co-Authored-By: 4adwentures <296413879+4adwentures@users.noreply.github.com>
The GUI-updater handoff race: the Desktop fires SIGTERM + app.quit() and
spawns hermes-setup, but its Python backend (`python.exe -m
hermes_cli.main serve`) can survive the teardown. The Desktop is gone --
nothing will respawn that backend -- yet the venv-holder guard refused on
it and the update dead-ended with "Hermes is still running" while the
user had zero windows open (observed twice on 2026-08-09, 01:59 and
02:17, bootstrap-installer.log).
New `_orphaned_desktop_backend_pids()` classifies remaining holders: a
serve/dashboard backend whose supervising parent is provably dead (PID
gone, or recycled -- parent created after the child) is a straggler safe
to reap. Any live-parent backend, non-backend holder, or unprovable case
keeps the refusal exactly as before. Reaping uses the new
`_stop_process_trees()` (taskkill /T /F), mirroring the Desktop's
forceKillProcessTree and install.ps1's venv sweep so the managed
.hermes-runtime interpreter child dies with its launcher (#70026).
Builds on #81327 (salvaged intact underneath): that fixed the same
parent-only-kill gap in install.ps1's venv sweep; this closes the
remaining dead-end in the `hermes update` guard itself.
E2E on a real Windows box: spawned a detached orphan with a
backend-shaped argv -> classifier returned its PID and the tree reap
killed it; a non-backend orphan and the live Desktop backend (parent
alive) both returned None (refusal preserved).
The classic CLI's light-mode detection sends an OSC 11 background-color
query and blind-waits 100ms. Terminal managers that swallow OSC 11
(herdr) made every startup pay the full 100ms for nothing, and any
in-order relay that answers slower than 100ms (SSH bridges, WSL,
loaded tmux servers) delivered the reply AFTER prompt_toolkit owned
the tty — the rgb:.../escape payload leaked into the input line as
gibberish characters.
Fix: send the OSC 11 query followed by a DA1 sentinel (ESC [ c) in one
write — the same fence pattern the Ink TUI's TerminalQuerier uses.
Terminals answer queries in order and effectively all of them answer
DA1, so the DA1 reply proves the terminal has already processed (or
ignored) our OSC 11. Fast terminals and herdr-style multiplexers now
resolve in ~1ms; slow relays get their reply consumed instead of
leaked; a hypothetical DA1-mute terminal falls back at a 1s safety
net, same clean timeout path as before.
Adds real-PTY regression tests covering the herdr-style (DA1-only),
slow-relay (+300ms reply), and fully mute emulator behaviors, each
asserting zero leftover bytes in the tty buffer. Sabotage-verified:
the slow-relay test fails against the old un-fenced code with the
exact leak payload in LEFTOVER.
On slow terminals (VPS, containers under load), the OSC 11 background
color response can arrive after TCSAFLUSH completes — leaking into
prompt_toolkit's input buffer and silently consuming the first 1–3
characters of every response.
Add a 50ms post-flush drain window that reads and discards any late
bytes via select() + os.read() before prompt_toolkit grabs the tty.
Fixes#40250
Adds the failed-result guard the salvage review called for:
_deliver_queued_first_response now takes deliver_media and the queued
follow-up call site passes deliver_media=not _delivery_result.get('failed').
A failed turn still delivers its normalized failure text (pinned by
test_run_agent_sends_normalized_failure_before_queued_followup), but its
attachments are no longer uploaded as if the turn succeeded — mirroring
the completed-turn path's 'not agent_result.get(failed)' guard.
Regression test added.
Ensure queued follow-up resends keep MEDIA-backed attachments by replaying the
first response through the gateway's text-plus-media delivery flow instead of a
plain adapter text send.
Closes the residual the contributor's own triage comment flagged: % was
excluded from the special-char class to protect the CJK LIKE fallback,
but a non-CJK query never reaches that fallback (is_cjk gates it), so
'50%' still hit MATCH raw and silently returned zero results. Strip %
whenever the sanitized query contains no CJK; the CJK path keeps its
pre-existing contract. Regression tests for both directions.
_sanitize_fts5_query's strip step only removed +{}():"^ . Every other
character FTS5's grammar rejects outside a quoted phrase reached MATCH
raw and raised, and — as the step's own comment says about the colon it
was fixed for — the execute site swallows that into zero results. Session
search silently found nothing for ordinary queries:
it's fts5: syntax error near "'"
gateway/run.py fts5: syntax error near "/"
user@host fts5: syntax error near "@"
a,b fts5: syntax error near ","
why? fts5: syntax error near "?"
e=mc2 fts5: syntax error near "="
Complete the class and assemble it with re.escape, because written as a
regex literal the backslash was eaten as an escape and never made it in
(C:\path\file still raised after the first pass).
Measured against a real FTS5 table over 651 realistic queries:
373 unparsable before, 77 after. The remainder is leading/trailing "." and
"-", which #43889 already covers.
% is deliberately left in: the CJK path falls back to a LIKE search that
needs it literal and escapes wildcards itself, so stripping it widened
those queries onto unrelated rows (test_cjk_like_escapes_wildcards).
Extends the picker fix to the read that actually uses the key: switch_model's
user-provider credential resolution (the ${VAR} api_key expansion and the
key_env fallback at the resolve-credentials step) still read os.environ raw
and passed the result to resolve_runtime_provider as explicit_api_key — so
under multiplex_profiles the actual switch, not just the picker listing,
could adopt another profile's key. Same _scoped_key_env helper, same
fail-closed semantics; identical behavior when multiplexing is off.
Adds end-to-end switch_model tests pinning that an installed scope wins over
the process environment for both read shapes.
854007d1c routed the remaining main-agent fallback key reads through
agent.secret_scope so the multiplexed gateway's per-profile scope applies.
list_authenticated_providers - which gateway/slash_commands.py calls
directly for /model - still resolved custom-endpoint and fallback-entry
credentials with raw os.environ.get(key_env), so under multiplex_profiles
one profile's picker reads whatever key the process environment happens to
hold, i.e. another profile's.
no multiplexing : profileA-key (unchanged)
scope installed : profileB-key (was profileA-key)
Route both reads through a _scoped_key_env() helper over
secret_scope.get_secret(). get_secret is identical to os.getenv when
multiplexing is off, so single-profile deployments are byte-for-byte
unchanged; a fail-closed UnscopedSecretError is treated as "no credential
visible for this profile", which is how the picker already handles a
missing key.
Scope: only the two key_env credential reads. The other environment reads
in that function are provider-presence probes (AWS creds, LM_BASE_URL),
a separate concern.
Per-turn .env adoption could rewrite agent.api_key while leaving
_credential_pool_entry_id on a previously rotated fallback. The next 429
then marked the healthy fallback exhausted via credential_id precedence
(#79156).
- Sync pool entry id after a successful env credential refresh
- First look does not stomp a pool-rotated key with the env primary
- mark_exhausted_and_rotate prefers api_key_hint when it disagrees with
credential_id
Fixes#79156
_detect_venv_python_processes() returned cmdline_raw[:120]. Gateways
autostarted via the managed-runtime interpreter carry a >120-char exe path
(.hermes-runtime\python\generation-...\cpython-3.11-...), so the truncated
cmdline ended inside the exe path, before '-m hermes_cli.main gateway run'.
The Desktop preflight's pausable-gateway exemption
(_scan_venv_blockers._is_pausable_gateway) therefore never matched, the
gateway was reported as a blocker, and every Desktop update aborted with
'Update didn't finish' even with all windows closed — the updater's own
gateway pause never got a chance to run.
Fix: return the full cmdline from the detector and truncate only at
display time (_format_venv_python_holders_message and the scan's JSON
cmdline field, after redaction).
Reproduced live on Windows 11: scan reported blocked=true for
'...cpython-3.1' (truncated); after the fix the same gateway pair scans
clear with pausable_gateways=2.
An alias that family-matches multiple catalog models (/model opus) used to
silently pick one via _model_sort_key heuristics. The heuristics have
guessed wrong repeatedly — dated snapshots like claude-opus-4-20250514
parsed as version 20,250,514 and outranked claude-opus-4-8; suffix
tiebreaks landed on the cheapest tier — and every wrong guess silently
switches the user to a model they did not ask for.
resolve_alias now raises AmbiguousAliasError whenever more than one model
matches the alias family; switch_model catches it at all three call sites
(explicit-provider path, current-provider path, authenticated-provider
fallback) and returns a failure result listing the candidates
(best-guess-first ordering, capped at 10) with instructions to pick an
exact name. A single match still resolves automatically, and DIRECT_ALIASES
exact mappings are unaffected.
The date-stamp split from #67571 is kept, demoted from selection logic to
display ordering of the candidate list.
Supersedes the auto-pick approach of #67571; credit to @Sahaun and @GottZ
for the date-stamp parser analysis that this builds on.
_model_sort_key treated YYYYMMDD snapshot stamps (e.g.
claude-opus-4-20250514) as version components, so 20250514 > 8
and resolve_alias("opus", "anthropic") returned the wrong model.
Fix: split components ≥ 19_000_101 (smallest plausible date stamp)
out of the version tuple, keeping them as a trailing tiebreaker so
bare IDs sort before their dated snapshots and newer snapshots
before older ones. Shorter numeric components (mistral-large-2411,
gpt-4-0613) keep their current behavior. No models.dev dependency
in the sort path.
_resource_attributes() in otlp_exporter.py built its own hardcoded
resource dict (service.name/instance.id/telemetry.scope only) instead
of reusing gateway_health_export.py's _runtime_resource_attributes(),
which already applies the resource_attributes allowlist from config.
Result: operator-configured attributes like deployment.environment.name
reached metrics and diagnostic logs but never spans.
Span resource building now delegates to the same
_runtime_resource_attributes() helper metrics/logs already use,
removing the duplicate implementation instead of patching it in place.
Generic thinking fields (reasoning / reasoning_content + the
reasoning_details text charge) are replayed for at most the NEWEST
assistant turn on every transport: Anthropic strips all-but-newest at
convert time, Bedrock Converse never replays thinking, and strict
chat-completions providers reject or one-space-pad the field. The tail
budget walks charged them on every message anyway, spending 19-24% of
the budget (per the issue's 1,025-message measurement) on bytes that
provably never reach the wire — so the tail cut landed early and each
compaction discarded more real transcript than configured.
_estimate_msg_budget_tokens now partitions the replay keys:
* _ALWAYS_REPLAYED_BUDGET_KEYS (codex_reasoning_items,
codex_message_items) — charged unconditionally. These ride the wire
on every retained turn (#55572), and codex_reasoning_items now also
carries native server-side compaction checkpoints (#81747).
* _NEWEST_TURN_ONLY_BUDGET_KEYS (reasoning, reasoning_content) + the
reasoning_details text charge — charged only for the newest assistant
turn via charge_stale_thinking, resolved by the three budget walks
(tail cut, raw-budget re-walk, proactive-prune boundary).
Default stays the conservative full charge for callers without
turn-position context. A partition invariant test pins that any future
_REPLAY_BUDGET_KEYS entry must be classified into exactly one class.
Direction credit: #73669 (@x7peeps) and #73730 (@webtecnica) both
attacked this; the keep_open reviews asked for provider/API-mode-aware
accounting that keeps Codex carriers charged — this implements that
shape.
get_messages() only deserializes content and tool_calls; the structured
reasoning columns (reasoning_details, codex_reasoning_items,
codex_message_items) come back as the raw TEXT they were stored as.
Feeding those rows straight back into a write, which is exactly what
the POST /api/sessions/{id}/fork handler does by piping get_messages()
into replace_messages(), hit an unguarded json.dumps() and stored the
already-serialized string encoded a second time. On replay of the fork,
json.loads() then yields the inner string instead of a list, and every
consumer's isinstance(..., list) gate silently drops it: preserved
Anthropic thinking blocks, Codex encrypted-reasoning/message-item
replay, and OpenRouter multi-turn reasoning context are all lost after
a fork, with one more encoding layer added per fork.
The /branch copy loop had the same defect from the other side: it
forwarded reasoning but none of the structured columns, and both TUI
branch writers persisted role/content alone, dropping reasoning and
reasoning_content along with them.
Route the six dumps sites in append_message and _insert_message_rows
through a shared guard that keeps already-serialized strings as-is;
structured values from the live runtime are dumped exactly as before.
Forward the reasoning fields in all three branch writers, matching the
set gateway/slash_commands.py already forwards on its own /branch path.
Adam Durham