queryRewrite (default off) gates the latest-message rewrite so the
extra auxiliary LLM call is opt-in. firstTurnBaseWait and
firstTurnDialecticWait expose the turn-1 bounded waits in seconds
(0 disables). All three resolve host-block-first like every other
field. Also pins per-host timeout resolution with tests.
Move query_rewrite from the honcho plugin to plugins/memory/ and
rename the auxiliary task key honcho_query_rewrite ->
memory_query_rewrite so any memory provider can use the same
rewrite path and model/timeout config block. No behavior change.
dialecticMaxChars (default 600) is documented as the budget for the dialectic
supplement auto-injected into the system prompt every turn — a small recurring
cost that is correct to bound tightly. But dialectic_query() applied that cap
unconditionally, so explicit honcho_reasoning tool calls — where the model
deliberately spends a turn asking for a synthesized answer — were silently
truncated mid-word to 600 chars with a trailing " …", no error surfaced. The
full answer is returned by Honcho server-side; the clip happens client-side.
The auto-injection path already has its own token-based budget (contextTokens,
enforced in prefetch() via _truncate_to_budget), so the char cap's real job is a
cheap always-on guardrail for that recurring injection. Explicit tool results
are already bounded server-side by Honcho's dialectic MAX_OUTPUT_TOKENS and don't
need the injection cap — sibling tools (honcho_search, honcho_context) don't
post-clip their results either.
Add apply_injection_cap (default True, preserving current behavior) to
dialectic_query(); the honcho_reasoning tool handler passes False so it returns
Honcho's full synthesized answer. Auto-injection is unchanged. Tests cover both
the capped injection path and the uncapped tool path.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
honcho_conclude's delete action was unreachable in practice: no tool ever
surfaced a real conclusion id for the model to pass as delete_id.
honcho_search only searches the separate Message resource space, and the
SDK's ConclusionScope.list()/.query() (which do return real Conclusion.id
values) were never wired into any tool.
Adds an optional list mode to honcho_conclude (query to search, omit to
browse recent conclusions), backed by a new
HonchoSessionManager.list_conclusions(). No new tool, no changes to the
create/delete signatures or their conclusions_of() routing.
injectionFrequency='first-turn' returned empty for the entire
prefetch_context() method on turns 2+, which blocked the dialectic
supplement from being consumed and injected. The dialectic has its
own cadence (dialecticCadence) and must continue to fire and inject
independently of the base context layer.
Now first-turn mode gates only Layer 1 (base context: representation
+ card), letting Layer 2 (dialectic supplement) flow through its
normal consumption path on every turn.
Also fixes all remaining tests that passed dialecticCadence via
cfg_extra={'raw': {...}} to use the typed dialectic_cadence field.
injectionFrequency, contextCadence, and dialecticCadence were read
only via raw.get() which checks root-level keys in honcho.json.
Settings placed inside hosts.<name> (the normal per-host location)
were silently ignored, falling back to defaults.
- Add typed dataclass fields: injection_frequency, context_cadence,
dialectic_cadence with host-block-first resolution chains matching
the pattern used by all other config fields.
- Update HonchoMemoryProvider.initialize() to read from typed cfg
fields instead of cfg.raw directly.
- Fix search_context tests that mocked _honcho directly — the
.honcho property getter calls get_honcho_client() which overwrites
the backing field, so use patch.object on the property instead.
- Add injection_frequency and context_cadence config override tests.
Symptom: Honcho logs show a dialectic answer was generated, but Hermes never
injects it — intermittently.
Root cause: the dialectic supplement that queue_prefetch() fires at the end of
turn N is stored pending (fired_at=N) for consumption by turn N+1's prefetch().
But prefetch()'s trivial-prompt guard returned early BEFORE the consumption
block. So when turn N+1's prompt was trivial ('ok', 'yes', 'continue', a slash
command), the ready result was never consumed, and a few turns later the
stale-discard guard dropped it. Generated by Honcho, never seen by the model.
The dependence on 'is the consuming turn trivial?' is why the loss looked random.
Fix: trivial turns now consume and inject a ready, non-stale pending result while
still spending no new work (no base-context fetch, no new dialectic fire). A
trivial ack shouldn't generate context, but it shouldn't destroy an answer
already computed for that exact turn. Extracted the pop+stale-check into a shared
_consume_pending_dialectic() used by both the trivial path and the normal path so
they age-check identically.
Preserved (covered by new tests): genuinely stale results are still discarded on
trivial turns; trivial turns still fire no new work; a trivial turn with nothing
pending still injects nothing.
Adds regression tests for inject-on-trivial and discard-stale-on-trivial.
honcho_search routed through search_context() -> peer.context(search_query=),
which returns the peer's standing representation + card. The search_query arg
does not turn that endpoint into a search, so results were effectively
query-independent: the same representation blob regardless of the query. Factual
lookups ('what medication', 'which value did we pick') returned noise.
Rewire search_context() to call the workspace message-search endpoint
(Honcho.search) with a peer_perspective filter: RRF-ranked (hybrid semantic +
full-text) raw message excerpts spanning every session the peer was a member of,
across all authors, membership-time-scoped. This is the cross-session factual
recall primitive.
peer_perspective is chosen over the alternatives because it is the only scope
that is simultaneously (a) cross-session, (b) inclusive of assistant-authored
facts about the peer (peer-author search drops these, and they are a large part
of what you want to recall about yourself), and (c) privacy-scoped to the peer's
own sessions (plain workspace search leaks other peers' sessions).
- snippets are labeled by author so the model can tell user-stated facts from
assistant-derived ones
- max_tokens is now an enforced budget (was accepted but meaningless)
- graceful fallback to peer-authored search if peer_perspective is unsupported
- query length clamped under the embedding input cap
Replaces 3 change-detector tests that asserted the old representation-dump
behavior with 4 that assert the message-search contract + fallback path.
auto_title_session runs as a bare daemon-thread target. Any exception
escaping it hits the default threading excepthook and sprays a raw
traceback into the user's terminal mid-session. The canonical trigger
is the post-'hermes update' stale-module window: the function's lazy
imports read NEW source from disk while already-imported modules
(agent.portal_tags) are still the OLD cached version, producing an
ImportError that repeats on every auto-title attempt until the
long-running process restarts (seen live after 9ce0e67f2 added
set_conversation_context).
The public entrypoint now wraps the body in a catch-all that logs one
WARNING naming the likely cause ('restart the running Hermes process'),
routes the exception through the existing failure_callback channel
(user-visible warning in CLI, debug-suppressed in gateway per #23246),
and never re-raises. This also makes the function honor its own
docstring contract ('silently skips if title generation fails').
Add the flag-gated compute-host supervisor, delta/control protocol, PPID orphan guard, inline fail-open path, synthetic GIL-heavy turn seam, and AC-4 certify harness.
Verified on current origin/main: 343 focused tests pass; Ruff and diff checks pass; 360s AC-4 run with six heavy lanes passes at 6.11ms serving p99 with zero stalls and valid load.
Co-authored-by: Kyzcreig <9063726+Kyzcreig@users.noreply.github.com>
Bare /reasoning and /fast now render a native one-tap picker on
picker-capable platforms, with automatic fallback to the text status
card everywhere else — parity with the /model picker UX.
- gateway/slash_commands.py: generic send_choice_picker capability gate
(detected on the adapter type, like send_model_picker); selection and
typed arguments flow through one shared application path so they can
never diverge; choices built from VALID_REASONING_EFFORTS so future
levels appear automatically
- telegram: flat inline-keyboard picker (cp:<idx> callbacks), authorized
users only (same gate as approval buttons)
- discord: ChoicePickerView select menu, auth mirrors ExecApprovalView,
2-minute timeout with expiry edit
- matrix: reaction-based picker; reaction set extended to 12 slots to
fit the full effort ladder + subcommands
- locales: picker_title + choice labels in all 16 languages
- docs: ADDING_A_PLATFORM.md capability table
Closes#61110.
terminal_tool reads all settings from TERMINAL_* env vars, bridged from
config.yaml by the CLI, gateway, and TUI-PTY launchers. Processes that
skip every launcher bridge — hermes serve / the Desktop app backend's
in-process agents, the desktop cron ticker — saw an unset TERMINAL_ENV
and silently ran every command on the host even when config.yaml selects
terminal.backend: docker. A user who configured Docker isolation got
unsandboxed host execution with no warning.
Two layers:
- _ensure_terminal_env_bridged() in _get_env_config(): when TERMINAL_ENV
is unset, backfill TERMINAL_* from config.yaml via
apply_terminal_config_to_env(override=False). Explicit env always wins
(honor explicit choice; only fix the accidental fallback). One-shot,
fail-open to the historical local default.
- cmd_dashboard/serve: run the same bridge at startup so every consumer
in the backend process (in-process agents, desktop cron ticker,
tui_gateway cwd resolution) sees the bridged env directly.
Fixes#63141, #54449, #61115, #65696.
hermes update ran TWO separate pre-update backup mechanisms: the
config-gated full zip (updates.pre_update_backup, default off) and an
unconditional quick state snapshot added for #15733 that ignored the
user's setting entirely. On a large state.db (observed: 24 GB) the
'cheap' snapshot silently added ~60s to every update and ate 24 GB of
disk in state-snapshots/.
Now there is ONE mechanism, gated by updates.pre_update_backup with
three modes:
- quick (new default): state snapshot of critical small files (pairing
JSONs, cron jobs, config, auth, per-profile DBs). Files over 1 GiB
are skipped with a warning so a bloated state.db can never stall the
update again.
- full: the quick snapshot plus the HERMES_HOME zip (old 'true'
behavior; --backup forces it for one run).
- off: nothing runs — an explicit opt-out now disables the quick
snapshot too (--no-backup does the same per-run).
Legacy booleans are honored: true -> full, false -> off.
_run_pre_update_backup() now returns the quick-snapshot id so the
post-update cron-jobs restore safety net (#34600) keeps working; the
snapshot moved from the post-fetch site to the pre-mutation site,
which also covers the zip-fallback update path it previously missed.
Add current-main regression coverage for both the registered provider
profile and core GitHub Responses path while leaving live catalog loading
to the complementary catalog-resolution work in #51953.
The Copilot provider profile unconditionally mapped ``xhigh`` to ``high`` before
checking the model's catalog, so models that DO support ``xhigh`` (e.g. the
gpt-5.x family per the live /models catalog) were silently capped one level
down.
Honor the requested effort when the catalog lists it as supported, and only
downgrade when it does not, choosing the nearest weaker supported level
(xhigh->high, minimal->low, else medium, else the first supported level). This
matches the nearest-down clamp behavior used elsewhere for the ``max`` effort.
Adds tests/plugins/model_providers/test_copilot_profile.py covering forward,
downgrade, and fallback paths (catalog lookup stubbed).
The desktop self-update rebuild (`hermes desktop --build-only`, driven by
the macOS in-app updater) runs `_run_npm_install_deterministic`, which used
plain `npm ci` / `npm install`. Those honor an inherited
`NODE_ENV=production` (or npm `omit=dev`) and silently omit devDependencies.
The desktop build toolchain — tsc, vite, electron-builder — are all
devDependencies, so under `NODE_ENV=production` the install completes (exit 0)
but `tsc` is never placed, and the very next step (`tsc -b && vite build`)
dies with `tsc: command not found` (exit 127). The user sees
"UPDATE DIDN'T FINISH / Rebuilding the desktop app failed (exit 127)" even
though the git/pip update applied cleanly. NODE_ENV=production can leak in
from a shell profile, a parent process, or a packaged-app launch context.
Force `--include=dev` on both the `npm ci` and `npm install` paths. The only
callers are frontend builds (desktop / TUI / web), which always need the dev
toolchain, so this is safe across the board.
Verified empirically: under NODE_ENV=production, `npm ci` leaves tsc MISSING
while `npm ci --include=dev` installs it. Added two regression tests
(test_npm_ci_forces_include_dev, test_npm_install_fallback_forces_include_dev)
— both confirmed to FAIL when the fix is reverted.
When no agent is active, '/queue <prompt>' previously fell through
dispatch with its raw text intact instead of being treated as a
normal prompt. Rewrite event.text to the bare payload (mirroring the
/steer no-active-agent path just below) and return a usage hint when
the payload is empty.
Salvaged from PR #29290 (queue half only — the /footer mid-run
dispatch half already landed on main via #65521).
Three follow-up fixes to the salvaged reasoning_effort support, all verified
live against ollama.com /v1/chat/completions + /api/show on deepseek-v4-pro,
gemma3, and qwen3-coder:
1. Capability-gate on /api/show 'thinking'. The original ignored the
supports_reasoning flag and emitted reasoning_effort for every model. Now
gated: only models whose native /api/show capabilities list contains
'thinking' (deepseek-v4 yes; gemma3 / qwen3-coder no) get reasoning_effort.
Mirrors the LM Studio pattern — capability resolved once per (model,
base_url) in run_agent._supports_reasoning_extra_body via a cached probe
(hermes_cli.models.ollama_model_supports_thinking), threaded into the
profile hook as supports_reasoning. No live HTTP in the per-request path.
2. Disable actually disables. Ollama Cloud defaults to thinking ON and IGNORES
the extra_body.thinking:{type:disabled} shape (verified: still returned
reasoning). The only working off switch is top-level reasoning_effort:'none'.
The salvaged code returned ({}, {}) for enabled:false / effort:none, leaving
thinking ON. Now emits {'reasoning_effort': 'none'}.
3. Omit unrecognized effort. The original forwarded any unknown string verbatim
including 'minimal' (a real Hermes effort level). Ollama Cloud rejects
unrecognized values with a hard HTTP 400 (accepted set: low/medium/high/
max/none), so forwarding 'minimal' would break the request. Now omitted.
Core touches (run_agent.py, hermes_cli/models.py) add the capability probe;
the plugin profile only consumes the resolved flag. 24/24 profile tests green;
194 provider/transport tests unaffected.
LM Studio's request vocabulary tops out at "xhigh", but Hermes' generic
effort ladder has since grown two stronger levels. "max" and "ultra" miss
the _LM_VALID_EFFORTS membership test, keep the initialized "medium"
default, and are thereby conflated with unparseable input -- so asking for
more reasoning yields less than "xhigh":
high -> 'high' xhigh -> 'xhigh'
max -> 'medium' ultra -> 'medium'
This is drift, not a design choice. The valid set was an exact mirror of
VALID_REASONING_EFFORTS when the file was authored; the ladder then grew
"max" and later "ultra", and the sweep that taught every other provider
about the new levels missed this module -- it has never been touched since
it was written.
Clamp the two stronger levels onto LM Studio's declared ceiling instead,
mirroring the ceiling clamp every other provider already applies. Widening
_LM_VALID_EFFORTS would instead assert that LM Studio accepts "max" on the
wire, which is a provider-side claim this repo cannot verify; clamping
consumes only the ceiling the file already declares for itself.
The clamp is kept separate from _LM_EFFORT_ALIASES because that mapping is
also applied to the model's published allowed_options, which must not be
rewritten. A clamped value stays subject to the allowed_options check, so a
model that does not publish "xhigh" still gets the field omitted and falls
back to its own default -- exactly how a directly-requested "xhigh" behaves.
The regression test asserts monotonicity over the canonical ladder rather
than the two values alone, so the next level added upstream cannot silently
reintroduce the inversion.
Replaces the salvaged tests' dict-config path (which required a
GatewayRunner.__init__ dict-coercion hack — dropped from this salvage;
the second commit on the PR branch existed only to support it) with the
established bare-runner test pattern. Also asserts full argument
passthrough to the inner task.
When multiplex_profiles is true, background tasks spawned by /background
command failed with UnscopedSecretError because _resolve_session_agent_runtime()
was called without a profile secret scope. This fix wraps the task in
_profile_runtime_scope, mirroring the pattern used by _run_agent.
Fixes#60726
_download_url_with_cap called urlopen() after only a scheme check, so a
model-controlled URL could reach loopback services, RFC1918/CGNAT hosts,
or cloud metadata endpoints (169.254.169.254), and a public host could
302 to any of those unvalidated.
Route the fetch through the repo's canonical SSRF guard instead:
validate every hop with tools.url_safety.is_safe_url() and follow
redirects manually (httpx, follow_redirects=False, 5-hop limit) so each
Location target is re-checked before it is fetched — the same pattern
as tools/skills_hub._guarded_http_get. The streaming size cap is
unchanged. Local-fixture tests opt in via HERMES_ALLOW_PRIVATE_URLS
(the guard's documented escape hatch); new tests pin rejection of
loopback, cloud-metadata, and private-range URLs, a mocked
public→loopback redirect, and a mocked public happy path.
The salvaged attachment-toolset commit predated main centralizing the
25 MB cap as kanban_db.KANBAN_ATTACHMENT_MAX_BYTES and re-introduced a
private _MAX_ATTACHMENT_BYTES alias. Drop the duplicate: kanban_db's
store_attachment_bytes(), the dashboard upload endpoint, and the
kanban_attach_url tool all reference the one shared constant now, and
the tests monkeypatch that same name.
The kanban board has had full attachment storage and a dashboard HTTP
API (upload/list/download/delete) since #35338, but there was no agent
toolset tool and no `hermes kanban` CLI verb for attachments. Agents and
scripts that don't go through the dashboard server (or can't touch the DB
directly) had no way to create or read real attachments — only links in
comments.
Close that gap by mirroring the existing comment surface:
- `kanban_db.store_attachment_bytes()` — one shared write path (validate
name, enforce the 25 MB cap, write the blob under the per-task dir with
collision-free naming, insert the metadata row, clean up an orphan blob
if the insert fails). `_MAX_ATTACHMENT_BYTES`, `_safe_attachment_name`,
and a new `_collision_free_path` move here so the dashboard, the tool,
and the CLI all share one implementation and can't drift.
- Tools (`tools/kanban_tools.py`): `kanban_attach` (inline base64),
`kanban_attach_url` (server-side http/https fetch with the same cap),
`kanban_attachments` (list). Write tools respect worker task-ownership;
list is read-only. Registered in the `kanban` toolset.
- CLI (`hermes_cli/kanban.py`): `attach <id> <path>`, `attachments <id>`,
`attach-rm <attachment_id>`.
- Dashboard `upload_task_attachment` now imports the shared helpers and
uses `_collision_free_path` — behavior identical (still streams to disk
with the cap, still 413 on overflow).
- Docs (AGENTS.md, kanban-worker skill) and toolset membership updated.
Tests: tool round-trip + oversize + bad base64 + ownership; attach_url
against a local HTTP fixture incl. oversize-mid-stream and non-http
scheme rejection; CLI attach/attachments/attach-rm; shared-helper unit
tests; dashboard parity preserved.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
DeepSeek's own API (api.deepseek.com) reports context-cache hits as
top-level usage.prompt_cache_hit_tokens / prompt_cache_miss_tokens
(prompt_tokens = hit + miss), not the OpenAI nested
prompt_tokens_details.cached_tokens shape. Neither normalize_usage()
nor the chat_completions transport's extract_cache_stats() read those
fields, so direct DeepSeek sessions always showed 0 cache-hit tokens:
invisible in accounting, mis-billed at the full input rate, and 0%
cache display.
Both layers now fall back to prompt_cache_hit_tokens when the nested
shape is absent; the nested value wins when both are present (proxies).
Fixes#61871.
Server-side coverage for the critical-CSS shim (PR #36024 salvage):
- user theme → style block emitted with ONLY real bundle variable names
(--background-base/--midground-base from layerVars(),
--theme-font-sans/--theme-base-size from typographyVars()/index.css),
and an html,body rule expressed via those vars so runtime theme
switches never leave a stale canvas/font
- built-in / unknown / non-string active theme → no block
- malformed theme YAML and load_config() exceptions → no crash, index
still serves
- </style> breakout attempt in a theme value stays escaped
- mount_spa integration: block present in <head> for user themes,
absent for built-ins
Review follow-ups on the cherry-picked #36043 commit:
1. Guard the custom:<name> passthrough with a _get_named_custom_provider
lookup. The PR unconditionally kept the full custom:<name> string, which
broke config-less runtime custom providers (#34777 regression — entries
that exist only in the live runtime, not config.yaml): the named arm
found no entry and resolution fell through to Step 2. Now custom:<name>
only takes the named arm when a config entry actually exists; otherwise
it collapses to the anonymous-custom arm with the runtime endpoint,
preserving pre-PR behavior.
2. Drop the dead 'explicit_api_key = runtime_api_key' assignment (and its
misleading comment) in the named-entry branch. resolve_provider_client's
named-custom arm derives the key exclusively from the entry's
api_key/key_env and never reads explicit_api_key, so the assignment was
a no-op. Wiring precedence in was not justified: for a named custom
provider the runtime key IS the entry's key (set_runtime_main sources it
from the same config), so deletion is the honest option.
3. Tighten the Palantir Bearer-auth check from a loose substring match
('palantirfoundry' in normalized) to a hostname match via
base_url_host_matches(..., 'palantirfoundry.com'), so path segments or
lookalike domains containing the string no longer trigger Bearer auth.
Tests: named-custom anthropic_messages end-to-end routing (full name kept,
AnthropicAuxiliaryClient at the original /anthropic URL, no /v1 rewrite)
plus Palantir Bearer-auth positive and substring-false-positive cases.
Same bug class as the salvaged #65305/#65307: hmac.compare_digest (and
secrets.compare_digest) raise TypeError when given a str containing
non-ASCII characters, and these call sites feed it raw request input.
Compare as UTF-8 bytes everywhere:
- gateway/platforms/msgraph_webhook.py: clientState from request body
- gateway/platforms/whatsapp_cloud.py: hub.verify_token query param +
X-Hub-Signature-256 header (comment claimed 'works on str' — it
doesn't for non-ASCII)
- plugins/platforms/feishu: verification token + x-lark-signature
- plugins/platforms/raft: bridge token header
- plugins/platforms/line: X-Line-Signature
- plugins/platforms/sms: X-Twilio-Signature
- tools/code_execution_tool.py: sandbox RPC token (both loops)
Regression tests for the two gateway-core sites (msgraph, whatsapp).
The fix routes the Svix v1 comparison through _hmac_str_equal too, but the
existing non-ASCII tests only exercised the GitHub/GitLab/generic V1/V2
branches. Add a Svix case (valid svix-id + fresh svix-timestamp so it
reaches the v1,<sig> compare) with a non-ASCII signature, which raised
TypeError before the fix and now rejects cleanly.
_validate_signature backs the public webhook receiver. It compared each
attacker-supplied signature/token header (GitHub X-Hub-Signature-256,
GitLab X-Gitlab-Token, generic X-Webhook-Signature / -V2, and the Svix v1
header) against a computed hex/base64 digest with hmac.compare_digest on
two str values. compare_digest raises TypeError on a str containing
non-ASCII characters, and the header is raw client input on an
unauthenticated endpoint — so any internet client could POST a single
non-ASCII byte in the signature header and raise out of the handler,
returning a 500 instead of a clean 401. Fail-closed, but an on-demand
crash of the request path.
Route all five comparisons through a small _hmac_str_equal() helper that
encodes both sides to UTF-8 bytes before the constant-time compare
(compare_digest has no ASCII restriction on bytes). Semantics are
unchanged for valid signatures; a hostile non-ASCII header now fails
closed with a rejection instead of raising.
Adds regression tests: non-ASCII GitHub/GitLab/generic/V2 signature
headers return False (no raise), and a non-ASCII configured secret still
matches its exact token value.
Also maps drexux0@gmail.com in scripts/release.py AUTHOR_MAP.
_check_auth gates every OpenAI-compatible API server endpoint. It compared
the client's raw bearer token against the configured key with
hmac.compare_digest on two str values. compare_digest raises TypeError on
a str containing non-ASCII characters, and the token comes straight from
the Authorization header — so a request with a single non-ASCII byte in
the key (a stray unicode char, a smart quote, a pasted BOM) crashed the
handler with an unhandled TypeError. Every endpoint calls _check_auth
without a try/except, so the framework turned that into a 500 Internal
Server Error instead of the intended 401 Invalid API key.
Compare as bytes, matching web_server.py's dashboard-token check
(hmac.compare_digest(auth.encode(), expected.encode())). Encoding both
sides keeps the timing-safe comparison and its semantics identical for
valid keys while making a non-ASCII token fail closed with a clean 401.
Adds regression tests: a non-ASCII bearer token returns 401 (no raise),
and a non-ASCII configured key still authenticates against its exact
value.
Follow-up to the truncate_message split-loop floor. Two review points:
- CapabilityDescriptor.from_json trusted the wire max_message_length
verbatim, so a connector advertising 0 ('no limit') — or a buggy one
sending 0/negative — produced a descriptor whose bound flowed straight
into the adapter's MAX_MESSAGE_LENGTH and truncate_message. Normalize
it to the documented 4096 default (mirrors from_platform_entry's
'or 4096' and docs/relay-connector-contract.md), fixing the degenerate
budget at its source rather than only surviving it downstream.
- Document the truncate_message length contract for a budget too small
for one codepoint (max_length=1 with a 2-unit surrogate pair under
utf16_len): the chunk intentionally exceeds max_length by that one
indivisible codepoint, because emitting it whole preserves content
where the alternatives are data loss or an infinite loop.
Tests: from_json normalizes 0 and negative bounds to 4096 and passes a
real positive bound through unchanged; the sub-codepoint budget emits
whole codepoints with no data loss (all emojis preserved) and a chunk
that necessarily exceeds the 1-unit budget.
BasePlatformAdapter.truncate_message() splits an over-length reply into
chunks. When max_length is 0 or 1 (and the content is longer), the split
loop makes no progress and spins forever, appending empty chunks — an
unbounded hang that pins a CPU and grows the chunk list until OOM:
- headroom = max_length - INDICATOR_RESERVE - ... goes negative, and the
< 1 fallback (max_length // 2) is also 0;
- so _cp_limit is 0, the region is empty, no split point is found, and
split_at falls back to _cp_limit (0);
- chunk_body is remaining[:0] = "", remaining never shrinks, loop repeats.
The same stall is reachable under utf16_len (Telegram) whenever the next
char is a surrogate-pair emoji wider than the whole budget, so _cp_limit
maps to 0 codepoints even for max_length >= 2.
A pathological max_length is not hypothetical: the relay capability
descriptor's max_message_length is taken verbatim from the connector
(gateway/relay/descriptor.py from_json) and assigned straight to the
adapter's MAX_MESSAGE_LENGTH (gateway/relay/adapter.py), and 0 is a
documented "no limit" value there.
Guarantee forward progress: floor headroom at 1, and floor the
final split_at at max(1, _cp_limit) so at least one codepoint is always
consumed per iteration. Normal splitting is unaffected (both floors only
bite when the budget is already degenerate).
Adds regression tests that run truncate_message on a worker thread and
fail if it doesn't return: max_length 0/1/2 terminate and preserve every
character, and the utf16 emoji case terminates too.
A fallback chain entry can name its API key via key_env (or the
api_key_env alias) per the fallback-providers docs, but only the gateway
path resolved it — TUI/desktop, cron, and CLI setup fallbacks ignored it,
so a fallback provider whose key lives in a non-standard env var never
resolved on those surfaces.
Centralize the inline-api_key-then-key_env lookup in
hermes_cli/fallback_config.resolve_entry_api_key() and use it at all four
fallback resolution sites (tui_gateway, cron scheduler, gateway runner,
CLI setup mixin); the CLI mixin also gains the base_url passthrough the
other surfaces already had.
Salvaged from PR #43861 (surgical reapply — the original branch predates
the #65264 fallback restructuring).
Rebuilt from PR #61283 onto the /p/<profile>/ routing world (7aa21e336):
_profile_scope(None) now enters the DEFAULT profile's runtime scope when
multiplexing is active instead of returning nullcontext(). api_server is
a port-binding platform living on the default profile, so plain requests
(no /p/ prefix) are the primary path — with fail-closed get_secret they
crashed with UnscopedSecretError on the first credential read (#61276).
All three wrapped call sites (chat-completions executor, /v1/runs agent
construction and _run_sync) inherit the fix through the one seam.
Single-profile gateways keep the no-op. Regression tests ported from
the original PR to the _profile_scope seam.
Fixes#61276
Subset of PR #61985: _make_adapter_auth_check gains a profile_name
parameter and secondary-profile adapters (started in
_start_one_profile_adapters) bind it, so the auth callback's
SessionSource resolves the routed profile's adapter and pairing store
instead of silently falling back to the default profile. This is the
gap left open by the #65629 merge — adapter-internal auth checks (e.g.
Slack thread-context fetch) fire outside the wrapped message handler.
The PR's authz_mixin.py hunks are dropped: main's _auth_env (merged via
PR #65629) already covers the scoped allowlist reads they targeted.
