Review blockers (independent reviewer on #94513):
1. Reinstalling an exclude-mode catalog entry wiped the user's edited
tools.exclude, replacing it with manifest defaults. install_entry now
reads the prior exclude (like it already did for include) and re-writes
it verbatim on reinstall. Regression test added + sabotage-verified
(fails on old behavior); include-priority test added too.
2. aws-knowledge: exclude aws___retrieve_skill — vendor SKILL.md loader is
a vendor skill layer (live tools/list confirmed the tool exists).
3. betterstack: exclude list rewritten to cover the snake_case wire names
(vendor's own header examples show remove_dashboard) via globs alongside
the doc display-labels; caveat documented in the manifest — server is
OAuth-gated so pre-auth enumeration is impossible.
4. railway: exclude railway-agent (opaque server-side agent delegation,
acts outside Hermes's per-tool approval loop).
5. twelve-data: exclude oauth plumbing pseudo-tools + quota probe.
6. betterstack post_install no longer claims a fully-checked checklist —
exclude-mode bypasses the checklist; text now describes the applied
exclude list.
Live E2E: fresh temp HERMES_HOME — install applies manifest excludes,
user edit survives reinstall. 33/33 catalog tests green.
The cloudflare entry's 3,320-endpoint surface is ~43% product families a
personal/dev account never touches (Zero Trust org-fleet suite, Magic
Transit/WAN, Cloudforce One, Radar analytics, API Shield, legacy
migration surfaces). Ship a 34-pattern curated exclude list in the
manifest: 3,320 -> 1,905 tools kept, and everything Cloudflare adds
later stays enabled by default.
Mechanism, two small extensions:
- tools/mcp_tool.py: tools.include/exclude entries containing glob
metacharacters now match via fnmatch (plain names stay exact-match),
so a product family is one pattern instead of hundreds of stale
literals.
- hermes_cli/mcp_catalog.py: manifests may declare
tools.default_excluded (mutually exclusive with default_enabled);
install writes it to tools.exclude and skips the probe/checklist —
a 3,320-row curses checklist is not a UX. Prior user include
selections still win on reinstall.
Verified by replaying the real filter functions over the live-probed
3,320-tool list: 1,415 excluded, zero overmatch against a per-product
target audit; DNS/Workers/R2/D1/tunnels/Access/AI kept.
The server's default Code Mode surface (search/execute meta-tools) is
itself a tool-discovery layer; stacking it under Hermes tool_search
would mean two search hops and an opaque 2-tool surface. With
?codemode=false each of the ~3,300 API endpoints registers as its own
tool with a full JSON Schema, and Hermes's own progressive disclosure
defers and searches the complete catalog — one layer, total
information. Verified live: tools/list returns 3,320 tools, all with
input schemas. post_install documents the trade-off and how to opt
back into Code Mode.
Adds optional-mcps/cloudflare — Cloudflare's managed remote MCP server
(mcp.cloudflare.com/mcp) fronting the entire Cloudflare API (2,500+
endpoints across DNS, Workers, R2, KV, D1, Zero Trust, WAF, Pages)
through two Code Mode tools, search() and execute(), at a fixed ~1k-token
schema footprint. HTTP transport + native MCP OAuth 2.1 with DCR — no
install block, nothing to pin. post_install documents the scoped OAuth
grant, the bearer-token path for headless/CI, and Cloudflare's
product-specific servers for narrower surfaces.
Docs: mention Cloudflare in the hosted-OAuth MCP examples.
Review finding on #94633: _sanitize_label_value is lossy ('team/workspace'
and 'team_workspace' both sanitize to 'team_workspace'; >63-char keys
truncate identically), and container reuse is label-keyed — so two teams
with DIFFERENT shared keys could silently attach to one running container
(filesystem, processes, env) while their host sandboxes stayed separate.
Shared-key labels now carry a sha256 digest suffix of the raw key
(deterministic across processes; plain profile labels unchanged for
backward compat). Docs also state the first-creator-wins rule for image/
mounts on a shared container. Adds adversarial collision tests.
Follow-up on @fangliquanflq's opt-in (#84775): after the profile-scoping fix
(#94560) the container cache key is resolved in _resolve_container_task_id,
so the shared key must unify profiles there too — 'shared:<key>' for every
session of every opted-in profile AND for CLI/no-session runs. Delivery adds
the shared sandbox layout as the first translation candidate. Empty key
keeps strict per-profile isolation; SSH ignores the key entirely.
Independent review caught a compaction authority the gate missed:
post-turn micro-compaction (turn_finalizer -> _micro_compact) absorbs the
oldest exchanges into a rolling summary with no pre-compress checkpoint
hook in its path, and both compression.checkpoint_required and
compression.micro_compact could be enabled together — assistant evidence
could vanish into a summary the checkpoint filter later excludes, without
ever reaching the durable provider.
- agent_init: checkpoint_required forces micro-compaction off (warned),
mirroring the native-compaction suppression
- turn_finalizer: defense-in-depth guard at the call site (attribute is
plain mutable state a future path could flip on a live agent)
- behavioral regression test with a sabotage control (gate off proves the
harness reaches the call site; gate armed proves zero calls)
- docs + config example mention the suppression; stale v1 test header fixed
Per review: existing providers should not be retroactively re-versioned or
handed a changed payload. Version 1 is now the implicit historical
on_pre_compress() contract (best-effort, raw message list) that every
pre-existing provider is already on; the fail-closed checkpoint contract
becomes version 2. MemoryManager routes the raw transcript to v1 providers
unchanged and hands the host-normalized evidence list only to v2+ checkpoint
providers, so the plugin surface contract for shipped providers is
byte-identical with the gate off.
The fail-closed gate lived only in compress_context(), but two native
lossy owners compact without ever crossing it (review on #93996):
- codex app-server: in "native"/"off" auto-compaction mode (native is the
default) Hermes preflight is skipped and the codex agent compacts its
own thread inside run_turn() — the compress_context() rejection was
unreachable. init_agent now refuses checkpoint_required together with
api_mode=codex_app_server (BLOCKED_MISSING_PREREQUISITE, extracted as a
testable guard), and run_codex_app_server_turn() fails closed as
defense in depth before a turn can reach the codex-owned boundary.
- Responses server-side native compaction:
native_compaction_context_management() now returns None while the gate
is armed, so context_management never goes on the wire and the
checkpoint-aware Hermes compressor stays authoritative. The suppression
is logged once per process, not silently applied.
Regressions: checkpoint_required + app-server raises before run_turn()
(the session is never created); checkpoint_required keeps
context_management off the wire while the plain configuration still
produces it; the init guard refuses exactly the incompatible pair. Docs
and cli-config.yaml.example describe both bindings.
Refs #93986
Addresses the review on #93996:
- gateway: hygiene and manual /compress load the memory provider only when
compression.checkpoint_required is enabled (skip_memory=not required).
The historical fast path — no provider init, no best-effort hook — is
back for everyone who did not opt in, so default behavior is truly
unchanged.
- conversation_compression: assistant messages carrying both prose and
tool_calls keep their prose in the checkpoint evidence (the tool_calls
payload is stripped, the original message is not mutated); pure
tool-call wrappers without prose are still dropped.
- tests: legacy-database regression proving the _compressed_summary column
is added by the declarative _reconcile_columns() path on a plain reopen
(no version-gated migration needed — append_message works right after),
plus coverage for the prose-preserving filter.
- docs: providers must implement idempotent, content-keyed checkpoint
writes — a fail-closed block means the next attempt re-runs
on_pre_compress over largely the same transcript.
Refs #93986
Adds a 'Pre-Compress Checkpoints (fail-closed)' section to the memory
provider plugin guide: the versioned opt-in attribute, the operator-side
compression.checkpoint_required gate, fail-closed semantics, and the
normalized evidence contract including the persistent summary marker.
Refs #93986
Context compression is intentionally lossy. Deployments that archive
transcript evidence to an external durable store before compaction had no
way to guarantee the archive actually happened: MemoryManager.on_pre_compress
swallows provider failures by design, so a failed archive silently degraded
into data loss.
This adds an opt-in, provider-agnostic checkpoint contract:
- memory_provider: PRE_COMPRESS_CHECKPOINT_API_VERSION = 1; providers opt in
by advertising pre_compress_checkpoint_api_version. Version 0 keeps the
historical best-effort hook semantics.
- memory_manager: supports_pre_compress_checkpoint() capability probe;
on_pre_compress(require_checkpoint=True) propagates checkpoint-provider
failures and raises when no capable provider completed the checkpoint.
- conversation_compression: new compression.checkpoint_required config key
(default false, documented in cli-config.yaml.example). When enabled,
compaction fails closed with BLOCKED_MISSING_PREREQUISITE (the
uncompressed transcript is preserved) unless a checkpoint-capable provider
confirms the durable checkpoint. Providers receive normalized direct
user/assistant evidence: tool rows, system messages, tool-call wrappers,
and prior compaction summaries are filtered host-side into one stable
contract. codex_app_server compaction is rejected under the gate because
it exposes no truthful pre-compaction transcript boundary.
- hermes_state: persistent _compressed_summary column (declarative schema
migration via _reconcile_columns) so summary provenance survives process
restarts; only the resume model history carries the marker, keeping
get_messages_as_conversation on its existing contract.
- gateway: the lossy hygiene/auto-compact paths load the memory provider
(skip_memory=False) so a required checkpoint also guards those rewrites.
The gate arms only on an explicit boolean True (bare-MagicMock agents in
existing tests have truthy auto-attributes). Default behavior is unchanged:
checkpoint_required=false preserves best-effort semantics for all existing
providers. Contract tests, including a restart round-trip of the summary
marker, in tests/agent/test_pre_compress_checkpoint_contract.py.
Refs #93986
Full-screen captures carry no element tree, so the CaptureResult now has a
'note' field surfaced in the tool summary telling the model to call
capture(app='<AppName>') or capture(app='desktop') when it needs to act on
what it sees. Schema description updated to distinguish app='screen'
(composited full-screen image) from app='desktop' (shell surface with
clickable elements); docs + regression tests (14, sabotage-verified) added.
'Screenshot my screen' previously resolved the 'screen' sentinel to the OS
shell window (Progman/WorkerW) via list_windows — capturing the wallpaper +
icons layer, never the windows actually displayed. cua-driver's
get_desktop_state does a real composited full-screen grab; the
screen/fullscreen/all sentinels now route there directly, bypassing window
enumeration (which also keeps screenshots working when Windows UIA
enumeration hangs — trycua/cua#2110/#2113).
app='desktop' keeps the shell-window lane so desktop icons/taskbar stay
clickable.
Salvaged from PR #60081 by @2ndNatureAI (surgical reapply; original branch
predates the capture-routing refactor).
Commit a270c4ade's session-key fallback in _resolve_container_task_id was
added to stop cross-profile SSH environment reuse, but it wasn't backend-
gated: persistent Docker silently fragmented into one container per gateway
session, breaking the product contract (one long-lived container per profile,
shared by CLI and every session of that profile). #93950's vanishing MEDIA
attachments were downstream damage.
- persistent Docker (container_persistent: true) now keys to the profile:
literal 'default' for the default profile (same container as CLI),
'profile:<name>' for named profiles
- SSH and non-persistent Docker keep session scoping (the original leak fix
and the #82731 isolation contract are untouched)
- gateway MEDIA translation follows the profile layout and keeps the legacy
bug-window per-session sandboxes as fallback candidates, trying each until
the file resolves — old sessions self-heal, no migration
- /root/.hermes credential-surface refusal preserved across all layouts
Phase 2c on the full final diff flagged the re-export shim as
contradicting the adjacent house pattern (agent.replay_cleanup import,
which documents retiring private aliases once tests migrate). Migrate
all six tests to the canonical gateway.media_repair seam, import the
canonical name in run.py, and drop the dead 'and result' guard at the
background-task call site.
- Make repair_explicit_computer_use_media_paths fail-open internally
(cosmetic repair must never abort delivery); drop the cron-only
try/except so all three call sites are identical one-liners.
- Drop cron's redundant 'MEDIA:' pre-check (helper early-returns).
- Document the intentional lazy BasePlatformAdapter import (verified:
no cycle either way; keeps module import cheap for cron processes).
- Point the two new regression tests at the canonical
gateway.media_repair seam; pre-existing tests keep pinning the
gateway.run re-export shim.
- Docstring: matching is case-insensitive, say so.
Follow-up to the salvaged fix from PR #94439:
- Extract the repair into gateway/media_repair.py (shared module) and
re-export under the historical private name in gateway/run.py.
- Wire the repair into the two bypassed delivery surfaces: gateway
background tasks (_run_background_task_inner) and cron job delivery
(cron/scheduler.py) — both call agent.run_conversation directly and
never pass the main turn chokepoint.
- Fail closed on malformed/truncated JSON tool results: parse JSON-looking
content first instead of regex-scanning the raw string, which yielded a
doubled-backslash path artifact and rewrote the response to a path the
model never wrote.
- Deduplicate the tool_name_by_call_id builder (three verbatim copies in
gateway/run.py) into the shared module; hoist the abs-path prefix regex.
- Add regression tests: malformed-JSON fail-closed (mutation-checked) and
the compression-fallback last-user slice (incl. no-user fail-closed).
The first version of this test used inspect.getsource() on skill_manager_tool
and regex-parsed the guarded action literals. AGENTS.md bans source-text tests,
and the ban is right here: that test would pass against a guard wired to the
wrong call site and fail on a pure rename, neither of which is the thing worth
guarding.
Replaced with a behavioral assertion in the shape of the neighbouring
dry-run-banner test: stub _run_llm_review, run run_curator_review, and assert
the prompt that actually reached the model names skill_view and all four
guarded actions (edit, patch, write_file, remove_file).
It still discriminates: with the prompt block removed, action=edit and
action=remove_file no longer appear anywhere in the assembled prompt (the
toolset list only mentions patch, create, write_file and delete), so the test
fails. Runtime guard behavior stays where it belongs, in
tests/tools/test_skill_manager_tool.py.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`_background_review_read_before_write_guard` refuses a background-review
`skill_manage` write when the target file was not loaded via `skill_view` in
the same review turn (patch, edit, write_file over an existing file,
remove_file).
`CURATOR_REVIEW_PROMPT` never says so. It lists `skill_view` only under "read
the current landscape", so the reviewer goes straight to the write and every
mutation is refused. The failure is silent from the outside: the curator run
completes, writes nothing, and reads like a pass that simply found nothing to
consolidate. On our deployment that was 32 of 32 attempted writes rejected over
48h before anyone read the logs.
This adds the missing instruction to the toolset block, plus a test that fails
if a future guarded action is added to `skill_manager_tool` without being named
in the prompt — the guard and the prompt have to drift together or not at all.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The skill_manage guard (added in #55906) refuses any patch/edit of an
existing SKILL.md, or overwrite/removal of an existing support file,
unless the exact target was loaded via skill_view during the review.
Neither _SKILL_REVIEW_PROMPT nor _COMBINED_REVIEW_PROMPT ever mentioned
this, so models routinely issued the write without the pre-read, got
refused, and burned review iterations (#62397).
Both prompts now carry a Read-before-write section scoped to the
guard's actual contract: existing targets only, exact-path pre-read for
support files, transcript quotes don't count, new skills/new support
files exempt, and a bounded one-view-one-retry recovery instead of a
loop. Direction follows #60331 by @kkwills13 with the scope corrections
requested in review (existing-target-only wording, no delete claim,
bounded retry, contract tests for both prompt variants).
Fixes#62397.
- Key _user_space_cache on _conn_snapshot instead of client object identity,
so _new_client() results from the same connection share the cached user
(previously every on_memory_write triggered an uncached /api/v1/system/status
probe with a 30s default timeout)
- Thread a short timeout (0.05s) through the write-path identity probe
- Harden _tool_remember to snapshot the client before URI construction + POST,
matching the pattern already established in on_memory_write
- Remove dead instance method _user_scoped_uri (zero callers; all call sites
use the module-level function directly)
Co-authored-by: ehz0ah <haozhe4547@gmail.com>
Review follow-up to the viking://~ migration: the ~ home alias only
expands for USER/ADMIN roles. The DEFAULT dev auth mode (no
server.auth_mode, no root_api_key) resolves every request as ROOT,
which bypasses current-user expansion — the canonical parser rejects
viking://~ with 400 'Home alias URI is not canonical' (verified on a
live 0.4.16 server). A deployment upgrading to 0.4.16 with an
untouched ov.conf is in dev mode, so the ~ spelling would break
exactly the way the old uid-less one will.
Mirror the upstream first-party plugin pattern instead: resolve the
user space client-side from /api/v1/system/status (result.user,
'default' fallback) and emit explicit-uid
viking://user/<user>/memories/... URIs, which are canonical under
every auth mode (dev/ROOT, trusted/USER, api-key) and every server
version. viking://~/... input typed by the user keeps passing through
untouched. (#91995)
Upstream OpenViking removed the uid-less viking://user/<segment>
shorthand (#4196, merged 2026-08-21): reserved segments like memories
and peers no longer expand to the caller's space and the server
rejects them with HTTP 400 (NamespaceShapeError). First-party clients
were migrated to viking://~ in the same change; the Hermes plugin was
not (#91995).
Migrate every URI the plugin constructs — the profile/preferences/
entities session-start reads, the _build_memory_uri memory-mirroring
write path, and the tool-schema example — to viking://~/... README
uid-less references updated to match; canonical user-scoped forms
(viking://user/default/...) are unchanged. The ~ alias requires
OpenViking server >= 0.4.16 (#4167).
Same follow-through as the other filter-static stub updates: the three
lambdas patching filter_local_delivery_paths rejected the new keyword
and failed CI (tests/gateway/test_73771_media_resend_dedup.py).
De-silence the #93950 failure mode: when a container-absolute MEDIA path
under /workspace or /root cannot be resolved to a host sandbox file while
TERMINAL_ENV=docker, log the reason (no mounts / no prefix match / host
file missing) plus the delivering session key instead of only the generic
'Skipping unsafe MEDIA directive path' line upstream.
Persistent Docker containers bind <sandboxes>/docker/<task>/{workspace,home}
where <task> is sanitize_task_id_for_path("session:<session_key>") — but the
gateway's synthetic mounts hardcoded the literal "default" sandbox
(_default_docker_workspace_host_root / _docker_persistent_home_host_root).
For any session-scoped deployment the longest-prefix match missed, the
container path fell through to a host-filesystem resolve that could not
exist, and every MEDIA attachment was silently dropped.
The post-handler delivery pipeline also runs after
_handle_message_with_agent cleared the turn's session contextvars, so even
a correct sandbox derivation consulting ambient state would collapse onto
"default". Thread the delivering session's key explicitly through
validate_media_delivery_path -> _translate_docker_container_media_path ->
the two host-root helpers (same pattern as the TTS fix for #57049/#36685).
Default-sandbox resolution and the /root/.hermes credential exclusion are
preserved; contexts without a key keep the historical behavior.
A short Telegram drop used to fail the final reply immediately. The
answer then sat in the delivery ledger until the next gateway boot.
Wait up to 15s for the bot (or a replacement adapter) so a brief
blip delivers now, matching QQBot.
Pin the NoneType crash: TEAMS_SDK_AVAILABLE true plus a failed bind must
return False without calling App(). Also pin plugin import when the
microsoft_teams parent namespace is missing.
Teknium