The connector now delivers inbound (messages + interrupts) over the gateway's
OUTBOUND /relay WebSocket, not a signed HTTP POST to an inbound endpoint. The
gateway needs no inbound HTTP port — which is what makes hosted gateways (no
public IP) able to receive inbound at all.
- gateway/relay/adapter.py: connect() wires set_interrupt_inbound_handler(
self.on_interrupt) so connector->gateway interrupt_inbound frames bridge into
the existing per-session interrupt path (the inbound message handler was
already wired). Removed _maybe_start_inbound_receiver() + the _inbound_runner
lifecycle — there is no HTTP receiver anymore.
- gateway/relay/inbound_receiver.py: deleted (the signed-HTTP InboundDelivery
receiver).
- gateway/relay/__init__.py: removed relay_inbound_config() (dead with the
receiver gone). The delivery key is still set in-process by self-provision for
forward-compat but is no longer consumed for inbound.
- docs/relay-connector-contract.md: §3 rewritten — inbound is the WS back-channel
routed cross-instance via the connector's relay bus; §5 interrupt + §6 auth
table updated; the old signed-HTTP-POST + per-tenant-delivery-key-signing path
is documented as superseded. gatewayEndpoint noted as passthrough-plane only.
Tests: stub_connector grows set_interrupt_inbound_handler + push_interrupt;
new test_relay_interrupt case proves connect() wires BOTH inbound handlers and an
interrupt_inbound frame over the WS cancels the right session. Removed the
HTTP-receiver test; updated the crypto-shedding scan + self-provision delivery-key
assertion. 88 relay tests pass.
EXPERIMENTAL. Pairs with gateway-gateway (relay bus + WsGatewayDelivery) and the
NAS GATEWAY_RELAY_URL stamp. The cross-repo E2E (connector repo) proves the full
multi-instance path against this production adapter code.
* fix(desktop): show Hindsight memory provider
* feat(desktop): configure Hindsight memory provider
* fix(desktop): limit Hindsight modes to supported setup
* refactor(desktop): generic memory-provider config surface
Replace the bespoke Hindsight settings surface with a declarative,
schema-driven path so adding a memory provider is pure declaration —
no per-provider page, conditional, or endpoint.
- memory_providers.py: declarative registry. Each provider lists its
fields {key, label, kind, default, options, secret-vs-plain}. Hindsight's
mode is a select(cloud, local_external), so rejecting local_embedded
falls out of generic enum validation instead of a hand-written check.
- One generic endpoint pair GET/PUT /api/memory/providers/{name}/config.
GET returns declared fields + current values (secrets only as is_set,
never read back); PUT validates selects against their options, writes
plain fields to the provider config file, secrets to the env store,
and flips memory.provider.
- ProviderConfigPanel renders straight from the schema, replacing
hindsight-settings.tsx and the memory.provider === 'hindsight'
conditional in config-settings.tsx — same pattern as
toolset-config-panel.tsx off env_vars.
Scoped to memory providers; storage layout is unchanged so the runtime
Hindsight plugin reads the same config.json / HINDSIGHT_API_KEY / provider
keys as before. Tests cover the registry, endpoint behavior (defaults,
write+secret, select rejection, unknown provider, secret-never-returned),
and the generic panel.
Add platform_hints config so an admin can append to or replace Hermes'
built-in platform hint for a single messaging platform (WhatsApp, Slack,
Telegram, ...) without affecting other platforms. Enables enterprise
managed profiles to steer platform-aware skills (e.g. invoke a custom
table-formatting skill on WhatsApp where Markdown tables don't render)
while leaving Telegram/Slack/CLI behavior unchanged.
- hermes_cli/config.py: document platform_hints in DEFAULT_CONFIG
- agent/agent_init.py: load platform_hints -> agent._platform_hint_overrides
- agent/system_prompt.py: _resolve_platform_hint() applies append/replace
(replace wins; bare string = append shorthand); defensive on bad config
- tests: 16 cases covering append/replace/shorthand/isolation/malformed
Override only affects the platform-hint segment of the system prompt;
SOUL/context/memory tiers and general instructions are unchanged.
DELETE /api/sessions/{id} was the only session endpoint that didn't
resolve the id (detail, messages, rename, export all call
resolve_session_id) and 404'd when the row was already gone. The desktop
optimistically removes the sidebar row, then RESTORES it and shows the
error on any failure — so deleting a session that had just been reaped
(empty-session hygiene) or removed by a concurrent client resurrected a
ghost row and surfaced "session not found". /goal + auto-compression churn
leaves transient empty rows that race the sidebar snapshot, which is the
exact "I deleted the empty one and got 'session not found'" report.
Resolve exact ids / unique prefixes, and treat an already-absent session
as an idempotent success — DELETE's contract is "ensure it's gone". This
mirrors the bulk-delete endpoint, which already treats ghost ids as
success.
Tests: deleting an absent id is idempotent (200, not 404); delete resolves
a unique prefix; a real session still deletes.
When a worker calls kanban_create from inside a session that has a
persistent delivery channel, the originating session is now subscribed
to the new task's completion/block events automatically. The agent
that dispatched the task gets notified instead of having to poll.
- Gateway sessions (telegram/discord/slack): HERMES_SESSION_PLATFORM +
HERMES_SESSION_CHAT_ID ContextVars, set by the messaging gateway.
- TUI / desktop sessions: HERMES_SESSION_KEY in the subprocess env.
The TUI notification poller keys on platform='tui' + chat_id=<key>.
- CLI / cron / test: no persistent channel, no subscription.
Gated by kanban.auto_subscribe_on_create in config.yaml (default True).
Disable to mirror pre-feature behaviour — users who want explicit
kanban_notify-subscribe calls per task can set it to false. This
config gate addresses the design concern that got PR #19718 reverted
upstream (unconditional implicit auto-subscribe on tool-driven
kanban_create was too aggressive for orchestrator users).
HERMES_SESSION_ID is intentionally not a fallback channel — it is
set by ACP/agent subprocess telemetry for every invocation, not just
TUI, so treating it as a notification target would auto-subscribe
every CLI session and re-introduce the over-eager behaviour.
The kanban_create response now includes a 'subscribed' bool so
orchestrators can react if subscription failed (e.g. by falling
back to explicit kanban_notify-subscribe or to polling).
Includes 6 tests covering the gateway / TUI / CLI / partial-context /
gated / add_notify_sub-failure paths. All 90 tests in
test_kanban_tools.py pass; 509 broader kanban tests pass.
The new compression-tip tests poke started_at/ended_at directly via
db._conn to force deterministic lineage ordering. _conn is typed
Optional[Connection], so ty flagged .execute/.commit as unresolved on
None. Bind a local and assert it's non-None first to narrow the union.
Auto-compression ends the live session and forks a continuation child
(linked via parent_session_id). A long-lived parent keeps its own flushed
message rows, so resolve_resume_session_id()'s empty-head walk never
redirected it — resuming the parent id reloaded the pre-compression
transcript and dropped every turn generated after compression, including
the assistant's response. On the desktop this is the recurring "I sent a
message, came back, and the reply isn't there" report on large sessions:
the chat's routed id is the pre-rotation id, and both the gateway
session.resume RPC and the REST /messages read anchored on it.
Fix the resolver at the chokepoint: resolve_resume_session_id() now
follows the compression-continuation chain forward via get_compression_tip()
before its existing empty-head descendant walk. get_compression_tip() only
follows children whose parent ended with end_reason='compression' (created
after the parent was ended), so delegation/branch children never hijack a
resume. This fixes every resume caller at once (REST /messages, CLI
--resume, gateway /resume).
session.resume in tui_gateway was the one resume path that never called the
resolver — it used the raw target id directly. Route it through
resolve_resume_session_id() too (non-lazy only; lazy watch windows must
stay on their exact child branch). Resolving up front also re-anchors the
live-session fast path so a still-live rotated session is reused by its new
key instead of rebuilding a duplicate agent on the stale parent.
Tests:
- resolve_resume_session_id follows the tip even when the parent retains
messages, and is not confused by a delegation child.
- session.resume binds the agent to the continuation tip and returns the
post-compression reply.
Follow-up to the cap-removal salvage. The contributor guarded the new
unlimited default with `[:max_models] if max_models else ...`, which conflates
max_models=0 (used by slug-only callers that want an empty model list) with
None (unlimited). Tighten to `is not None` at all five slicing sites in
list_authenticated_providers / list_picker_providers, and add a regression test
asserting the three-way contract: None=full, 0=empty, N=first N.
The interactive model pickers (Desktop REST API, TUI model.options, CLI
/model) were hard-capped at max_models=50, which truncated large provider
catalogs like Kilo Gateway (336 models) to just 50 entries. This made
most models undiscoverable via the picker search box.
Changes:
- Change build_models_payload() default from max_models=50 to None (unlimited)
- Change list_authenticated_providers() default from max_models=8 to None
- Change list_picker_providers() default from max_models=8 to None
- Fix all [:max_models] slicing to handle None as 'no limit'
- Remove max_models=50 from 5 interactive picker callers:
* web_server.py: get_model_options (Desktop /api/model/options)
* web_server.py: get_recommended_default_model
* model_switch.py: prewarm_picker_cache_async
* tui_gateway/server.py: model.options JSON-RPC
* cli.py: HermesCLI model picker
- Telegram/Discord inline keyboard picker (gateway/slash_commands.py)
still passes max_models=50 explicitly — unchanged behavior.
The total_models field was already in the response payload and is now
meaningful since models.length == total_models for interactive pickers.
Fixes#48279
* feat(billing): nous_billing http client + BillingState core (phase 2b)
Phase 2b terminal-billing client foundation:
- hermes_cli/nous_billing.py: typed client for the 4 /api/billing/* endpoints
(state/charge/poll/auto-top-up). Raises typed errors (BillingScopeRequired,
BillingRateLimited, BillingAuthError) mapped from the live-verified contract;
fail-open is the caller's job. Idempotency-Key enforced client-side.
- agent/billing_view.py: surface-agnostic BillingState core + Decimal money
parsing (server emits decimal strings, not 2dp), fail-open builder,
idempotency-key gen, custom-amount validation.
- 51 unit tests (decimal parse/format, payload tiering, error->exception
matrix, fail-open, amount validation).
Plan: docs/plans/2026-06-13-001-phase-2b-terminal-billing-tui-plan.md
* feat(billing): billing:manage scope + lazy step-up re-auth (phase 2b)
- NOUS_BILLING_MANAGE_SCOPE constant.
- nous_token_has_billing_scope(): split-based scope check (no false-positive
substring match).
- step_up_nous_billing_scope(): re-runs the device flow requesting
billing:manage, reusing the held credential's portal/inference URLs + client_id
(so a preview stays a preview), persists like _login_nous but WITHOUT the model
picker. Returns True iff the minted token carries the scope (False when NAS
silently downscopes a non-admin / unticked grant).
Lazy step-up (plan D-A): normal login path unchanged; 403 insufficient_scope
from a billing call triggers this. 7 unit tests.
* feat(billing): billing JSON-RPC methods for the TUI (phase 2b)
billing.state / charge / charge_status / auto_reload / step_up in
tui_gateway/server.py. Return STRUCTURED success envelopes (result.ok +
result.error=<code>) rather than JSON-RPC-level errors, so the Ink rpc() promise
always resolves and the TUI branches on the typed billing error code
(insufficient_scope, rate_limited, no_payment_method, …) to render the right
affordance. Money serialized as decimal STRINGS + display strings. charge mints
+ echoes an idempotency_key for retry reuse. 16 unit tests.
* feat(billing): /billing CLI handler + command registry (phase 2b)
- CommandDef("billing", subcommands=buy|auto-reload|limit), added to
_SLACK_VIA_HERMES_ONLY so it routes via /hermes on Slack (keeps the 50-cap
parity test green, same as /credits).
- cli.py::_show_billing + screen helpers: all 5 screens (overview, buy→confirm→
poll, auto-reload, monthly-limit read-only). Reuses _prompt_text_input_modal /
_prompt_text_input (D-C). Non-interactive (_app is None) renders text + portal
deep-link, never prompts (R7). Decimal money end-to-end. 2s/5-min cancellable
poll loop; 429/503 = retry not failure; settled = ledger truth. Lazy step-up on
403 insufficient_scope. no_payment_method treated as mainline funnel-to-portal.
- 6 CLI tests; 156 command tests (incl. Slack/Telegram parity) green.
* feat(billing): /billing Ink TUI screens + tests (phase 2b)
- ui-tui/src/app/slash/commands/billing.ts: /billing TUI command covering all 5
screens — overview (text), buy <amt> → ConfirmReq → charge → non-blocking 2s/
5-min poll loop → settled/failed/timeout branches, auto-reload <below> <to> →
ConfirmReq → PATCH, limit (read-only). Reuses the existing ConfirmReq overlay
(D-C) — no bespoke component. Typed-error envelope branching: insufficient_scope
arms the lazy step-up confirm; no_payment_method/rate_limited/cap funnel to
portal. Client-side amount validation mirrors the server (bounds + 2dp).
- gatewayTypes.ts: Billing* response interfaces.
- registry.ts: register billingCommands.
- billingCommand.test.ts: 12 vitest cases (overview/gating/buy-confirm-poll-
settled/no_payment_method/step-up/limit/auto-reload/validation).
TUI build green; 12/12 vitest pass; slash tests pass once @hermes/ink is built.
* docs(billing): scrub private cross-repo references
NAS is a private repo — remove all references to it from the public PR:
- drop the cross-repo planning doc (planning scaffolding, not a deliverable;
the PR description documents the design)
- replace 'NAS' / 'PR #412 preview' mentions in code + test comments with
generic 'the server' / 'a preview deployment'
* docs(billing): scrub final NAS reference in step-up docstring
* docs(billing): drop dangling plan-doc refs
The phase-2b plan doc was removed in the cross-repo scrub (300afcc0b)
but two module docstrings still pointed at it. Drop the dead refs.
* feat(billing): interactive /billing overlay + step-up UX, portal-URL & token fixes
Adds the interactive /billing TUI overlay and hardens the terminal-billing
client across CLI and TUI.
- TUI: full /billing overlay state machine (overview to buy to confirm,
auto-reload, read-only monthly limit) reusing the existing confirm overlay.
- Step-up: surface the verification link in-transcript and open the browser
via the TUI's own opener (the device flow runs in the headless gateway, so a
printed URL was being dropped); run the step-up handler off the main loop and
emit the link as an out-of-band event so the gateway stays responsive.
- Step-up copy is scope-accurate ("Billing permission granted") and re-checks
/state so it never claims "enabled" when the org kill-switch is still off.
- Portal deep-links resolve to absolute URLs against the active portal base
(the server emits them relative) - fixes a bare "/billing?topup=open" link.
- Billing calls refresh an expired access token via the stored refresh token
instead of reporting a false "not logged in".
- Optimistic funnel: advise "set up a saved card on the portal" up front when
no card is on file (advisory, not a hard gate).
- Token resolution is cached briefly so the 2s charge poll loop stops
re-locking + re-reading the auth store on every tick; 401 re-resolves fresh.
- Remove the temporary demo-mode shims.
Validation: 87 Python billing tests, 88 TS tests (billing command + gateway
event handler), tsc clean, ink + ui-tui builds green.
* docs(billing): add /billing TUI screenshots for PR
* fix(cli): guard _last_invalidate on bare instances; update stale prompt-fallback test
The UI-invalidate throttle read self._last_invalidate unconditionally, which
raised AttributeError on HermesCLI instances built without __init__ (the
thread-safety test's object.__new__ shell). Guard the read with getattr.
The off-main-thread branch of _prompt_text_input was changed (#23185) to cancel
cleanly to None instead of falling back to a bare input() that would hang on the
slash-worker thread; the test still asserted the old direct-input fallback.
Update it to assert the current intended behavior: returns None, calls neither
run_in_terminal nor input(), and does not hang.
The "💾 Self-improvement review" summary (skill/memory updated) was invisible
on two surfaces:
- Desktop Electron app had no review.summary event handler — skill/memory
writes happened silently. Now appends a persistent system message to the
transcript (matching the Ink TUI's persistent-line semantics, not a
transient toast that can be missed).
- tui_gateway (backs both 'hermes --tui' and the desktop) never read
display.memory_notifications, so it always behaved as 'on' and ignored a
user who set 'off'/'verbose'. Added _load_memory_notifications() (mirrors
the messaging gateway's bool->str normalization, defaults to 'on') and
wired it to agent.memory_notifications, matching gateway/run.py and the CLI.
Delivery chain now reaches all surfaces:
background_review.py -> background_review_callback -> review.summary event ->
desktop transcript / Ink TUI line / gateway message / CLI print.
The universal PARALLEL_TOOL_CALL_GUIDANCE block already lives on main, but it
shipped with two rough edges this change cleans up:
- It duplicated the batching steer for Google models. The
GOOGLE_MODEL_OPERATIONAL_GUIDANCE block still carried its own
"Parallel tool calls" bullet, so Gemini/Gemma received the instruction
twice in one prompt. Drop the redundant bullet — the universal block is now
the single source.
- Its comment claimed "nothing in the open-source system prompt encouraged
batching," which was wrong: the steer existed for Google models only. Reword
to say the gap was that every *other* model got nothing.
- Tighten the test that asserts the steer (precedence-correct), and add an
invariant guarding against re-introducing the Google duplicate.
* Port from cline/cline#11514: encourage parallel tool calls
Add a universal system-prompt guidance block telling the model to batch
independent tool calls (reads, searches, web fetches, read-only commands)
into a single assistant turn instead of one call per turn. The runtime
already executes independent batches concurrently (read-only tools always;
non-overlapping path-scoped file ops); the open-source system prompt had
nothing steering the model to PRODUCE the batch. Fewer round-trips means
less resent context, which compounds over a long conversation.
- prompt_builder.py: new PARALLEL_TOOL_CALL_GUIDANCE block (short, static,
cache-amortised) modeled on TASK_COMPLETION_GUIDANCE.
- system_prompt.py: inject right after the task-completion block, gated by
agent.valid_tool_names + the new toggle.
- agent_init.py: read agent.parallel_tool_call_guidance (default True).
- config.py: add the default under the agent section.
- test_prompt_builder.py: behavior-contract tests (batching steer, dependent
carve-out, length bound) — invariants, not wording snapshots.
Adapted from Cline's TypeScript tool-surface guidance to hermes-agent's
Python prompt-assembly architecture and config-over-env conventions.
* fix(desktop): never persist or restore a named custom provider as bare "custom"
Custom providers vanish from the Desktop/TUI model picker with
"No LLM provider configured" — repeatedly fixed (#44062, #44109, #45578)
and repeatedly regressed (#44022, #47714) because every fix only recovered
the entry identity from a persisted base_url. When a session is
persisted/restored with the resolved provider "custom" and NO base_url, bare
"custom" leaked through verbatim; resolve_runtime_provider("custom") routes to
the OpenRouter default URL with no api_key, so the next turn/resume dies.
Bare "custom" is the resolved billing class shared by every named providers:/
custom_providers: entry — it is not a routable identity. Centralize the
"never let bare custom escape" invariant in one helper,
runtime_provider.canonical_custom_identity(), and apply it at all four leak
sites in tui_gateway/server.py:
- _ensure_session_db_row — the ORIGIN: first DB write seeds the bad row
- _runtime_model_config — live persist
- _stored_session_runtime_overrides — resume restore (heals old rows; drops
unrecoverable bare custom so resume falls back to config default)
- _make_agent — rebuild / per-turn
The helper recovers custom:<name> from the endpoint URL when present, else
from config.model.provider (the durable identity left when no base_url
survived). Regression tests in test_custom_provider_session_persistence.py
lock the no-base_url vector at every site so it cannot regress again.
The memory tool was strictly one-op-per-call. With the store running near
its char limit by design, a new add that would overflow gets rejected with
'consolidate now, then retry' -- but the model could not consolidate and add
in one call. It had to remove/replace across several turns, then retry the
add, each turn re-sending the whole conversation context. Expensive thrash.
Add an 'operations' array: a list of add/replace/remove ops applied
atomically against the FINAL char budget. The model frees space and adds new
entries in ONE call, even when an add alone would overflow. All-or-nothing:
any bad op aborts the whole batch, nothing written.
Root-cause note: the two agent-level memory interception sites
(agent_runtime_helpers.py, tool_executor.py) silently dropped any param not
in their explicit kwarg list, so 'operations' never reached the handler and
batch calls failed with 'Unknown action None'. Both now pass it through and
bridge each add/replace op to external memory providers.
Also: success response is now terminal (done=true + 'do not repeat' note,
no full-entries echo that invited re-edits); schema rewritten to lead with
the batch mechanism and an explicit one-shot stop rule (2138 -> 1476 chars).
Live-verified: near-full consolidate-and-add went 7 calls -> 1 call,
stable across 3 reps. 103 memory/approval tests + 398 background-review/
run_agent tests green; 6 new batch tests added.
PR #48372 relaxes EAP=Stop around the uv venv call so PowerShell 5.1
doesn't mistake uv's 'Using CPython ...' stderr for a terminating
NativeCommandError. But relaxing EAP also means a *genuine* uv venv
failure (exit != 0) no longer aborts on its own — Install-Venv would
continue and print 'Virtual environment ready', and in stage mode
Invoke-Stage would report ok=true, even though no venv was created.
Capture $LASTEXITCODE immediately after the relaxed call and throw on
non-zero (Pop-Location first, matching the function's other exit paths),
so the venv stage fails fast instead of falsely succeeding. This is the
explicit guard originally proposed in #48463 (devorun), composed on top
of #48372's reusable helper + regression test.
Adds a regression test asserting the uv venv exit-code capture + throw.
The dashboard MCP catalog only showed name/description/transport and a
non-clickable source. Users couldn't see what an entry connects to or runs
before installing — the exact detail the docs trust model tells them to vet.
- /api/mcp/catalog now returns transport target (url, or command+args),
auth_type, git install source/ref + bootstrap commands, default-enabled
tool hint, and post-install guidance per entry.
- McpPage renders the endpoint URL (http) or command+args (stdio), the git
install source/ref, a collapsible bootstrap-commands list, setup notes,
and the source as a clickable link when it's a URL.
- Docs: drop the 'uv pip install -e .[mcp]' quick-start step (Hermes does
not support pip installs; MCP ships with the standard install) and note
the dashboard now surfaces this detail.
- Strengthen the catalog endpoint test to assert the new inspection fields.
Epic's experimental Unreal MCP plugin embeds an MCP server inside the
Unreal Editor process, served over local HTTP (127.0.0.1:8000/mcp by
default). HTTP transport, no auth, no install block — the user enables
the plugin in-editor and Hermes connects to the URL.
Also drops test_optional_mcps_manifests_ship_in_both_wheel_and_sdist:
it asserted wheel/sdist packaging targets for pip/Homebrew/Nix installs,
which Hermes does not support — installs run from the repo checkout, where
the catalog is discovered by directory iteration with no packaging step.
* fix(tui): don't make Enter swallow trailing-space-only slash completions
Submitting a slash command in the TUI took three Enter presses: one to
complete the name (/ex → /exit), a second that only appended the trailing
space the gateway adds to keep the classic-CLI prompt_toolkit dropdown open
(/exit → "/exit "), and a third to actually submit.
The composer's submit handler accepted the highlighted completion whenever
applying it changed the input at all, so the whitespace-only delta ate an
extra keypress. Treat a completion whose only change is trailing whitespace
on an already-complete token as "already complete" and fall through to
submit. Partial-name and argument completions (a real token change) still
accept on Enter as before.
The replace/accept logic is extracted into pure helpers (applyCompletion,
completionToApplyOnSubmit) in domain/slash.ts.
* test(tui): cover Enter/completion trailing-space behavior and isolate poller queue
- completionApply.test.ts asserts completionToApplyOnSubmit accepts real
token completions (partial command name, argument) but returns null for a
trailing-space-only delta on an already-complete command, so Enter submits
instead of needing extra presses.
- test_notification_poller_delivers_completion / _skips_consumed previously
shared the process-global process_registry.completion_queue. Their events
carry no session_key, so a leaked/concurrent poller could dequeue and
dispatch them to a fixture agent without run_conversation, flaking CI
("AttributeError: '_FakeAgent' object has no attribute 'run_conversation'").
Isolate the queue per test (fresh queue.Queue via monkeypatch), matching the
sibling poller tests that already do this.
The salvaged guard allowed _rmtree_writable(SKILLS_DIR) itself. No call
site ever passes the root — every site passes a skill subdir or its .bak
sibling — so allowing the root only preserves the #48200 footgun (a dest
that collapses to the root wipes every installed skill). Require a strict
strict-child relationship and update the test that documented the
nonexistent 'full reset' capability.
Defense-in-depth fix for the silent wipe of ~/.hermes/ documented in
#48200. A `hermes update --yes` run silently destroyed a user's
.env, MEMORY.md, kanban.db, custom skills, and scripts. Two changes:
1. `_rmtree_writable` in tools/skills_sync.py now refuses to rmtree
anything outside SKILLS_DIR (the HERMES_HOME/skills/ root).
All five call sites pass paths under SKILLS_DIR, so the guard is
a no-op for current code and a loud, recoverable failure for
any future regression (bad path join, malicious bundled
manifest, stale path in scope after an exception).
2. The default `updates.pre_update_backup` flips from false to
true in hermes_cli/config.py. A few minutes of zip per update
is negligible compared to silent total data loss. Still
overridable; --no-backup still works for one-off opt-out.
Five new tests in TestRmtreeWritableScopeGuard (root path,
hermes home, sibling dir, skills root itself, subdir) plus a
flipped `test_default_enabled_creates_backup` in test_backup.py.
178/178 tests pass in the two affected files. Public method
signatures unchanged, no test-stub blast radius.
Closes#48200
Review feedback from egilewski:
1. Remove trailing whitespace from test docstring and mock patches (lines 1430, 1469, 1476, 1482)
2. Expand test coverage: also verify ANTHROPIC_API_KEY is stripped (not just OPENAI_API_KEY)
Changes:
- Remove trailing whitespace from test file
- Add ANTHROPIC_API_KEY to test environment
- Add assertion verifying ANTHROPIC_API_KEY is stripped from cua-driver subprocess env
- Syntax verified: python3 -m py_compile tests/tools/test_computer_use.py ✓
- Use _sanitize_subprocess_env() to filter Hermes-managed credentials
from the cua-driver subprocess environment (issue #37878)
- Prevents credential exfiltration to the third-party cua-driver binary
- Aligns with existing pattern used by browser-tool and other tools
- Add regression test to verify environment sanitization
The cua-driver is a lower-trust MCP subprocess per SECURITY.md §2.3.
Its inherited environment is now scrubbed by default, removing provider
API keys, gateway tokens, and platform credentials that should not leak
to third-party binaries.
Fixes#37878
The TUI banner reported fewer tools than the classic CLI for the same
config (e.g. 32 vs 38) when an MCP server connected slowly. Root cause:
the agent snapshots `agent.tools` once at build time and never re-reads
the registry. `_make_agent` briefly joins the background MCP discovery
thread (`wait_for_mcp_discovery`, ~0.75s) so fast servers land in that
snapshot, but a server slower than the bound — common for an HTTP MCP
server on first connect — lands *after* the agent is built. Its tools are
then absent from both the agent (uncallable until `/reload-mcp`) and the
banner for the whole session.
The classic CLI doesn't hit this because it re-derives
`get_tool_definitions()` at banner render time (which re-waits for
discovery), so it picks the late tools up.
Fix: after a fresh agent is built and its first `session.info` emitted,
if discovery is still in flight, schedule an off-critical-path daemon that
waits for it to finish, then rebuilds the tool snapshot and re-emits
`session.info` — the same rebuild `/reload-mcp` performs, but automatic.
Both the agent's callable tools and the banner count catch up.
Cache safety: the rebuild runs only while the session is still
pre-first-turn (`_user_turn_count`/`_api_call_count` both 0 → nothing
cached to invalidate). Once the user has sent a message we leave the
snapshot frozen rather than break the cached prompt prefix mid-conversation;
late tools then require an explicit `/reload-mcp` (user-consented), exactly
as today. No-op when discovery finished before the agent build, when the
join times out, when the registry was unchanged, or when the session was
swapped/closed while waiting.
Adds entry.mcp_discovery_in_flight() / join_mcp_discovery() accessors and
covers the matrix (added/none/post-turn/timeout/unchanged/replaced) with
unit tests.
Source-level guard (install.ps1 only runs on Windows, so there's no Linux CI
runner to execute it): the astral uv install line must be invoked via the call
operator on a resolved host variable, the bare-`powershell` literal that
produced the field-reported "The term 'powershell' is not recognized" must be
gone, and the resolver must be PATH-independent (Get-Process -Id $PID) and
pwsh-aware.
The prior assertion `all("turn1" in k or "turn2" in k for k in keys)` was
weak on two counts: it passes vacuously when keys is empty (a regression
that lost all state would slip through), and after turn 2 finalizes only
turn 1 lingers, so it only ever inspected turn 1 anyway. Replace it with an
exact check that one key survives, it is turn 1, and turn 2 never merged
into it — the real isolation invariant the test name claims.
Scoping the trace key by turn_id (the prior commit) fixed cross-turn
collisions but introduced a slow leak: _finish_trace only pops a key when a
turn ends cleanly (final response has content and no tool calls), so any
turn that is interrupted, ends on a tool call, or has empty final content
now leaves its uniquely-keyed entry in _TRACE_STATE forever. Previously the
constant per-session key was overwritten by the next turn, capping growth at
~1 entry per session.
Add an LRU cap (_MAX_TRACE_STATE) enforced by _evict_stale_locked, called
under _STATE_LOCK immediately before each insert. It evicts the
least-recently-updated entries (using the previously-dead last_updated_at
field) and ends their root span so nothing dangles. Regression test drives
50 non-finalizing turns against a cap of 8 and asserts the dict stays bounded
with the most-recent turns surviving.
The turn- and api-scoped branches each repeated the same
task/session/thread fallback ladder with only the infix differing. Extract
the shared prefix into _scope_prefix so a future scope dimension touches one
ladder instead of three. The legacy branch still returns a bare task_id (not
the task: prefix) for backward compatibility, so it stays separate.
Output key strings are unchanged; a new test pins them across every
task/session/turn/api combination since the keys are matched across hooks
and any drift would silently break trace finalization.
Cleanup pass on the salvage (behavior-preserving):
- diff_bundled_skill now uses the existing _skill_file_list() helper
instead of reimplementing the rglob/is_file/relative_to file-set
enumeration inline (twice).
- Extract _is_tracked_user_modification(origin_hash, user_hash) and use
it in BOTH the sync loop and list_user_modified_bundled_skills() so the
'kept user edit' rule can't drift between the two sites.
- _read_text_for_diff -> _read_for_diff returns (bytes, text); the binary
branch now compares the bytes it already read instead of re-reading
both files from disk.
- Drop the unused 'user_present' key from diff_bundled_skill's return
contract (no consumer or test ever read it).
- test_update_modified_notice: drop the brittle '>= 2 sites' count-floor
so consolidating the two print paths into a shared helper stays a
welcome refactor; keep the per-site 'count notice => discovery hint'
invariant (still mutation-tested).
The PR added helper-level tests for _trace_key but nothing exercised the
keys through the real hooks. This adds TestTurnTraceIsolation, which drives
on_pre_llm_request / on_post_llm_call across two turns of one gateway
session (task_id == session_id, unique turn_id, api_call_count reset per
turn) and asserts each turn opens its own root trace when the first turn
fails to finalize (tool-only final step). This test fails on the pre-fix
code (only one trace opened, turn 2 absorbed into turn 1) and passes with
the scoping fix.
Also pins the turn_id-over-api_request_id key precedence: the turn-scoped
post_llm_call carries no api_request_id, so it must still resolve to the
same key as the request-scoped hooks or finalization breaks.
Salvage follow-up to the cherry-picked feat/test commits:
- W1: the unpack/install update path in main.py printed the
'~ N user-modified (kept)' notice without the new
'hermes skills list-modified' hint that the git-pull path got.
Mirror the hint to both sites so the count is actionable
regardless of which update path runs.
- W2: 'hermes skills diff <name>' (bundled-vs-stock) now shares the
verb with the gateway write-approval 'diff <id>'. The gateway
handler's docstring + truncation message pointed users to
'/skills diff <id>' on the CLI, which now resolves a bundled skill
by that name instead. Point at the pending JSON file and note the
two diff commands are distinct.
- Add an invariant test asserting every 'user-modified (kept)' notice
in main.py carries the discovery hint (guards sibling drift).
Exercises the real sync pipeline (no mocked comparison logic): a pristine
synced skill is not flagged; an edited one is listed and diffed (modified +
added files); an unknown skill returns not-ok; and `reset --restore` clears
the modified state so revert and discovery stay consistent.
Resolves conflicts from the OpenViking churn that merged after #32445 was
opened (#48042/#47662 session-switch + write hardening, #47311/#47973):
- plugins/memory/openviking/__init__.py: keep both __init__ field groups
(the PR's _runtime_start_* alongside main's _prefetch_threads/_shutting_down).
- tests/plugins/memory/test_openviking_provider.py: keep BOTH the PR's new
setup-validation tests and main's session-switch/concurrency tests (disjoint
additions to the same region).
Two fixes layered while reconciling (contributor work otherwise preserved):
- Restore the merged tenant-header contract (#22414/#21232). The PR had changed
_VikingClient defaults to '' and made empty account/user OMIT the tenant
headers; main's contract is that empty falls back to 'default' and the
X-OpenViking-Account/User headers are ALWAYS sent (ROOT API keys need them).
Reverted the constructor to 'account or os.environ.get(..., "default")' and
updated the two PR tests that asserted the omit-when-empty behavior.
- Close a secret-file TOCTOU in the setup writers. _write_env_vars and
_write_ovcli_config wrote the api_key/root_api_key file and chmod 0600
AFTERWARD, leaving a world-readable window on newly-created files. Added
_precreate_secret_file() to create with 0600 before any secret bytes land.
* fix(dashboard): stream file uploads via multipart instead of base64 JSON
The dashboard file manager uploaded files (including backup/restore zip
archives) by reading them client-side with FileReader.readAsDataURL and
POSTing a base64 data URL inside a JSON body to /api/files/upload. For a
large backup this (a) inflates the payload ~33%, (b) buffers the whole
file plus its decoded copy in memory, and (c) reliably trips an upstream
proxy body-size/timeout limit, surfacing as a 502 with the upload
appearing to hang indefinitely (NS-501). Dashboard-only hosted users have
no shell fallback to place the archive, so backup restore was unusable.
Add a streaming multipart endpoint POST /api/files/upload-stream
(UploadFile + Form) that reads the request body in 1 MiB chunks straight
to a sibling temp file, enforces the existing 100 MB size cap as it
streams (413 on overflow, before buffering the whole file), and
atomically renames into place so a partial/aborted/over-limit upload
never clobbers an existing file. The frontend api.uploadFile now sends
multipart/form-data (raw bytes, no base64, browser-set boundary) and
FilesPage passes the File object directly; the dead readAsDataUrl helper
is removed. The legacy base64 JSON endpoint stays for backward compat.
FastAPI's UploadFile/Form require python-multipart, which is NOT pulled in
by fastapi itself, so it is added to the base deps, the [web] extra, and
the tool.dashboard lazy-install set (kept in sync).
Validated: 5 new endpoint tests (roundtrip, multi-chunk >1 MiB,
over-limit 413 without clobbering + no temp-file leak, overwrite=false
conflict, forced-root traversal containment); existing base64 tests still
pass; web typecheck + vite build clean; and a real uvicorn server E2E
(5 MB multipart upload -> HTTP 200 in 0.21s, exact byte match) plus a
30 MB TestClient roundtrip confirm constant-memory streaming end to end.
Reported via beta (NS-501).
* build(deps): regenerate uv.lock for python-multipart (NS-501)
CI ran uv lock --check / uv sync --locked which failed because the
python-multipart dependency add was not reflected in uv.lock. Regenerate
the lockfile (resolves to 0.0.20, matching the [web] extra pin) after
merging current main.
Importing a backup wrote every file from the zip over the target home
wholesale. On a hosted instance this clobbered gateway_state.json with the
source machine's last recorded run/desired state — driving the container-boot
reconciler (container_boot._read_desired_state, which only auto-starts a
gateway whose state is "running") off stale/foreign state and leaving the
gateway stuck "starting", disconnected from the Nous portal.
Add _IMPORT_SKIP_NAMES (gateway_state.json, gateway.pid, cron.pid,
gateway.lock, processes.json) and skip them by basename in run_import, so both
the root profile and named profiles preserve the target's own runtime state.
This mirrors what container_boot._STALE_RUNTIME_FILES already sweeps on every
container boot, and protects against older backups that predate the
backup-side exclusions. The import summary reports which files were preserved.
This is the second half of NS-501 (filed separately as NS-508): the upload
502 was fixed in #47663; this fixes the import-breaks-the-instance half.
The gateway half of relay Phase 3. On a MANAGED boot with relay configured and
no secret pinned, the runtime self-provisions its relay credentials IN-PROCESS:
resolve the agent's own Nous access token (resolve_nous_access_token) -> POST
the connector's /relay/provision asserting its own endpoint + route keys ->
set GATEWAY_RELAY_ID/SECRET/DELIVERY_KEY into os.environ so the immediately-
following register_relay_adapter() reads them and dials out authenticated.
No human, no enrollment token, no disk write — the creds live only in process
memory (save_env_value refuses under managed anyway, and keeping the secret off
any volume is the stronger posture). Stateless: process-env creds don't survive
a restart, so a managed container re-provisions every boot; the connector's
rotation window covers a still-connected prior instance. An explicitly-pinned
GATEWAY_RELAY_SECRET is respected (skip). Self-hosted is unchanged: humans keep
using `hermes gateway enroll`.
Endpoint provenance is gateway-asserted (GATEWAY_RELAY_ENDPOINT +
GATEWAY_RELAY_ROUTE_KEYS, env or gateway.relay_* config) — uniform code path
whether the operator sets it (self-hosted) or NAS stamps it (hosted, the only
case NAS knows the public URL). Both absent -> outbound-only provisioning
(credentials, no inbound routes). The connector scopes the asserted endpoint to
the verified tenant, so it stays within the security model.
- gateway/relay/__init__.py: relay_endpoint(), relay_route_keys(),
_provision_url(), _post_provision(), self_provision_if_managed() (never
raises — a provision failure logs and boots without relay auth).
- gateway/run.py: call self_provision_if_managed() immediately before
register_relay_adapter() in the startup path.
Tests: 12 unit (trigger logic, respect-pinned-secret, in-process env wiring,
endpoint+routes vs outbound-only, fail-soft on token/connector failure);
mutation-checked (drop is_managed guard / pinned-secret guard -> tests fail).
Cross-repo live E2E driver lands on the connector side (depends on this).
EXPERIMENTAL: relay auth scheme may change until >=2 Class-1 platforms validate.
The install method (docker/git/pip/...) describes the *running binary*, but
detect_install_method() read it from $HERMES_HOME/.install_method — a shared
DATA directory. The Docker docs deliberately bind-mount $HERMES_HOME
(~/.hermes:/opt/data) so config/sessions/memory persist and can be shared with
a host-side Desktop/CLI install.
When a containerized gateway and a host install share one $HERMES_HOME, the
home-scoped stamp is a single slot describing two installs: the published image
stamps 'docker' on every boot, the host install then reads 'docker' and the
in-app updater refuses to run 'hermes update' ("doesn't apply inside the Docker
container"). Reinstalling the Desktop app from the DMG doesn't help because the
contaminated stamp is re-read every time.
Fix (option 1 — code-scoped stamp):
- detect_install_method() reads <install tree>/.install_method first (next to
the running code, immune to the shared data dir). It falls back to the legacy
$HERMES_HOME stamp for back-compat, but IGNORES a 'docker' home stamp when
not actually containerized — so already-poisoned shared homes self-heal.
- stamp_install_method() writes the code-scoped stamp.
- install.sh stamps $INSTALL_DIR instead of $HERMES_HOME.
- Dockerfile bakes 'docker' into /opt/hermes/.install_method at build time
(inside the immutable block); stage2-hook.sh no longer writes the home stamp
and proactively removes a stale 'docker' one to heal existing shared homes.
Genuine containers still resolve to 'docker' (baked stamp, or legacy home stamp
honored when containerized). Unstamped installs in generic containers still fall
through to git/pip (preserves the #34397 fix).
* feat(relay): authenticate the connector⇄gateway WS channel
The relay gateway may be customer-managed and internet-exposed, so the
connector⇄gateway channel is itself authenticated (distinct from the
platform crypto the relay path sheds). Add gateway/relay/auth.py — a
Python port of the connector's HMAC token + delivery-signature schemes
(relayAuthToken.ts / deliverySigning.ts), verified byte-for-byte against
the connector's compiled TypeScript via cross-language test vectors.
Present an Authorization bearer on the /relay WS upgrade keyed by the
per-gateway secret (resolved from GATEWAY_RELAY_ID / GATEWAY_RELAY_SECRET
in env or config). The connector rejects an unauthenticated/invalid/
revoked upgrade with close 4401.
* feat(relay): signed-HTTP inbound delivery receiver
The connector delivers normalized inbound events to a tenant's gateway
over a signed HTTP POST, not the outbound /relay WS: the connector
instance owning a platform socket is generally not the instance a given
gateway dialed out to, so inbound targets a tenant endpoint that may
load-balance across gateway instances.
Add gateway/relay/inbound_receiver.py — verifies x-relay-signature /
x-relay-timestamp over the EXACT raw request bytes (re-serializing would
break the HMAC: JS JSON.stringify is compact, Python json.dumps spaces)
against the per-tenant delivery key verify list within a 300s replay
window, then dispatches messages to handle_message and interrupts to the
interrupt handler. Wire it into the adapter lifecycle (start in connect()
when a delivery key + bind port are configured, tear down in disconnect();
a purely-outbound dev gateway runs without it).
Refine test_relay_sheds_crypto to distinguish PLATFORM crypto (Discord
ed25519, Twilio/WeCom HMAC — still shed) from the connector⇄gateway
CHANNEL auth (intended): auth.py / inbound_receiver.py are exempt from
the platform-symbol scan but still banned from importing platform-crypto
modules, plus a positive guard that auth.py uses only stdlib hmac/hashlib.
* feat(relay): hermes gateway enroll CLI
Add the gateway half of zero-touch enrollment. `hermes gateway enroll`
resolves a fresh Nous Portal access token (the tenant-proving identity),
POSTs {enrollmentToken, gatewayId} to the connector's /relay/enroll, and
persists GATEWAY_RELAY_ID / GATEWAY_RELAY_SECRET / GATEWAY_RELAY_DELIVERY_KEY
to ~/.hermes/.env. The per-gateway secret authenticates the WS upgrade;
the per-tenant delivery key verifies signed inbound deliveries.
Refuses under is_managed() (hosted installs get the secret stamped in by
the orchestrator). Added as an 'enroll' subcommand on the existing
gateway subparser — not a new top-level command.
* docs(relay): inbound is signed HTTP, not WS; document channel auth
Fix the stale contract: §3/§5 said inbound rode the WS socket (single-
instance only, predates the multi-instance socket-ownership + channel-auth
model). Inbound + connector→gateway interrupt are signed HTTP POSTs to the
tenant endpoint. Add §6.1 documenting the two channel-auth schemes (per-
gateway WS-upgrade secret, per-tenant inbound delivery key) and how they
differ from the platform crypto the relay path sheds.
* test(relay): update build_gateway_parser callers for cmd_gateway_enroll
The enroll subcommand added cmd_gateway_enroll as a required keyword-only
arg to build_gateway_parser, but two existing parser-extraction tests still
called it with only cmd_gateway/cmd_proxy — failing CI with TypeError.
Thread the new handler through both call sites and add a test asserting
`gateway enroll` dispatches to cmd_gateway_enroll with its flags parsed.
* fix(docker): supervised gateway uses --replace to take over stale holder
Inside the s6 container image the per-profile gateway service rendered a
bare `hermes gateway run` (no --replace). When a gateway is started
OUTSIDE s6 — a stray shell `hermes gateway run`, an agent action, or the
Open WebUI helper (scripts/setup_open_webui.sh) — it grabs the
per-HERMES_HOME PID lock first. The supervised slot then execs the bare
`gateway run`, hits the "Another gateway instance is already running"
guard, exits non-zero, and s6 restarts it: a restart loop that floods the
log every ~12s and never binds. The container looks up but the gateway is
permanently down, and dashboard-only users (no shell) cannot recover.
Render the supervised run script as `gateway run --replace` so s6 is
authoritative for its slot: it reaps the stale holder via the hardened
takeover path (takeover marker + SIGTERM->SIGKILL-with-confirmation +
scoped-lock cleanup in gateway/run.py) and binds. This matches the
systemd service path, which already builds its argv with --replace
(_build_gateway_argv / 'nohup hermes gateway run --replace'), and the
intent already documented in _maybe_redirect_run_to_s6_supervision. The
existing HERMES_S6_SUPERVISED_CHILD sentinel still prevents the
run->start->run redirect recursion. Each profile is scoped to its own
HERMES_HOME and s6 guarantees one supervised instance per slot, so there
is no legitimate supervised sibling for --replace to clobber.
Reported via beta (NS-505): gateway.log showed PID 17907 'running
(manual process)' with the guard error repeating every ~12s on
v2026.6.5.
Adds a regression test asserting every gateway-run exec line in the
rendered script (default + named profile, both privilege branches)
carries --replace, and updates the existing render-script assertion.
* fix(ci): remove stray .venv symlink committed into repo
The PR's commit accidentally tracked a .venv symlink pointing at the
developer's local venv (mode 120000 -> /home/ben/nous/hermes-agent/.venv).
The CI test/e2e/build jobs run `uv venv` to create .venv and failed with
`failed to create directory .venv: File exists (os error 17)` because the
checkout already contained the symlink. All test shards aborted in <15s
during setup, before any test ran.
Untrack the symlink and add a bare `.venv` entry to .gitignore (the
existing `.venv/` rule only matches a directory, so a symlink slipped
through).
Salvage corrections on top of @XVVH's #44341:
- Make native web_search injection a 1:1 swap for an already-present client
web_search function, NOT an additive grant. The original unconditionally
appended {"type":"web_search"} on every is_xai_responses turn with any
tools, force-enabling Grok server-side search even when the user never
enabled the web toolset (bypassing Hermes web-provider config + tool-trace
plumbing). Now gated on a client web_search actually being present.
- Reconcile grok-composer context to 200000 (merged in #47908) rather than
262144; 200k is xAI's published usable context window for Composer 2.5,
262144 is the /v1/responses input+output budget.
- Update tests to match scoped behavior + add a no-web-toolset guard test.
- AUTHOR_MAP entry for #44341 salvage.
Incomplete-guard (server-side *_call items at in_progress no longer flip
has_incomplete_items) and preflight built-in-tool allowlist kept as-is.
- model_metadata: grok-composer-2.5-fast → 262144 (OAuth slug not in /v1/models)
- codex transport: inject native {"type":"web_search"} for is_xai_responses;
drop client web_search to avoid duplicate-name 400s
- codex adapter: do not treat in-progress server-side *_call items as incomplete
- tests: adapter, transport build_kwargs, model_metadata, oauth recovery
Weak open models (mimo, nemotron-class) that see tool-call XML/JSON sitting in
file contents or tool output get primed and emit their own structured tool
calls mimicking the payload — usually with an empty/whitespace name. Those
calls can't be fuzzy-repaired toward a real tool, so the dispatch loop returns
an error and the model retries. Before this fix, every empty-name error dumped
the full tool catalog back to the model, which fed the priming loop more names
to mimic and inflated context 3-4x across the retry budget.
A blank/whitespace-only tool name now gets a terse anti-priming error that
tells the model in-context tool-call syntax is DATA, with no catalog dump. A
genuinely-wrong-but-nonempty name (a real typo) still gets the full catalog so
the model can self-correct.
Not a sandbox/auth boundary issue: Hermes never parses tool-call text from
content into executable calls (structured tool_calls only; the lone text->call
parser is the Copilot ACP transport and it also rejects empty names). The
reporter's own debug dump confirms the injection never executed.
Behavior-contract test added: empty-name -> terse error, no catalog; nonempty
unknown -> catalog preserved. Exercised end-to-end via run_conversation against
an in-process mock provider.
