Repository navigation
Landing slate: 34 approved PRs across 12 plugins (security, memory-wiki, mcp-serve, rea, homeassistant, xchat, spotify, advanced-delegation, agent-importers, camofox, accent-picker, blender) - #446
Merged
Conversation
(cherry picked from commit 82a9bb0)
(cherry picked from commit bac1a5c)
…00-char limit
"Fix in chat" and "Start scan" send a session.create whose title is built from
evidence text: "Fix: <finding title> · <finding id> · <timestamp>" and
"Security scan · <repo> · <scan id>". Hermes core caps session titles at
SessionDB.MAX_TITLE_LENGTH = 100 and validate_title turns anything longer into
RuntimeStoreError('invalid_params'), so the create is refused and the finding
pane shows the bare code plus "No success was confirmed...". Any finding title
over ~66 characters could not open a fix chat at all.
chatTitle(text, suffix) clamps the descriptive part to fit the limit and keeps
the identifying suffix whole, so a chat is still named for its finding or scan
and a re-click cannot land in another one's chat. Both callers use it, and
startChat() clamps as a last resort so a future caller cannot reintroduce the
refusal.
The cut is unit-budgeted and drops a dangling half of a surrogate pair: a title
whose limit lands inside an emoji would otherwise carry a lone surrogate, which
the gateway's SQLite refuses to store, and that finding's chat would still not
open.
Titles that already fit are byte-identical to before. The e2e suite covers the
102-character title that fails today, a long repo name, the unchanged
short-title and empty-title outputs, the astral cut, and the title the "Fix in
chat" click actually puts on the wire.
Fixes #390.
(cherry picked from commit ec43acd)
(cherry picked from commit 2213d83)
…ent sessions Cron scaffolding and delegated subagents are activity records, not conversations. Their prompt text dominated the subject index and the Recent sessions list on profiles with recurring jobs. Filter non-conversational sources at the subject aggregation and overview recent-sessions call sites. The daily log keeps full coverage: it is an activity record. Fixes #379 (cherry picked from commit 22c487c)
…mit counts conversations Filtering after list_sessions_rich(limit=10) left Recent sessions empty whenever the ten newest rows were cron runs (a profile with ~15 daily jobs). Pass exclude_sources to the query instead (available since core v0.21, the plugin's requires_hermes floor). (cherry picked from commit 928323d)
(cherry picked from commit 5bb898c)
(cherry picked from commit 031d406)
… SessionDB members Core's plugin-facing hermes_state.SessionDB no longer carries set_session_hidden or db_path, so the new day-bucket tests errored on current core. (cherry picked from commit cd399ea)
…hanged sessions (#49) Every dashboard request re-read every recent session and re-ran the deterministic extractors over the full message history — tens of seconds per refresh on large profiles (#49). The per-session extraction only depends on the session row plus its messages, so memoize it as one JSON unit per session under a fingerprint of the cheap sessions-table fields the gateway already bumps on every append/rewrite/rewind (message_count, tool_call_count, transcript_revision, last_active). Warm requests re-assemble aggregates from units without reading messages; a write to one session invalidates only that session; units persist to an atomic sidecar under HERMES_HOME/cache/memory-wiki/ (profile-scoped, bounded, LRA eviction); a shared lock lets concurrent cold requests share the build. MEMORY_WIKI_CACHE=off restores the uncached path and cache=None keeps byte-identical legacy behavior. Tests: fingerprint sensitivity, no-rescan-on-warm, single-session delta, parity with the uncached path (including after restart), corrupt sidecar recovery, bound eviction, env switch. (cherry picked from commit 335fe2d)
…I_CACHE env knob, write the sidecar 0600 The slow-scan event-loop test fakes _core(), so reaching the cache module through _core()._cache_module() 500'd every route (CI red on all 3 OSes). Behaviour config does not belong in a new env var; the sidecar carries session titles/snippets, so owner-only like state.db. (cherry picked from commit c16e23d)
The landing stack makes #318's per-session units carry one bucket per message day (#262), so the cached daily log buckets exactly like the uncached scan. Tests: #250's stale-core fake accepts #318's cache kwarg, #262's single-day test reads sessions through _recent_sessions (the _load_memory_sessions helper #318 replaced), and a new multi-day cached == uncached == from-disk test pins the combination.
…\n\nmcp-serve/hermes_tools.py still advertised browser_navigate, browser_click, etc.,\nwhich were removed in hermes-agent#143146. available_tool_specs() skips unknown\nnames, so they never appeared in tools/list, but README, --help and plugin.yaml\nclaimed browser support. Drop them from EXPOSED_TOOLS and _URL_ARGS, and update\ndocs, help text and tests.\n\nCloses #304. (cherry picked from commit 01c0403)
The expectation still allowed a bare localhost:8080 image URL through vision_analyze, but main's _URL_ARGS keeps allow_bare for web_extract only. Drop that case from the passing-URLs test and pin the refusal plus the per-tool allowance explicitly. (cherry picked from commit d4ae3db)
…d names to core's registry The command description, server instructions, README intro and module docstring still promised browser tools after the browser_* removal. Note why browser_exec stays out, and add a test that every EXPOSED_TOOLS name is a registered Hermes tool. (cherry picked from commit 0af6fb3)
…essaging bridge The bridge is built unconditionally today, so a host that wants Hermes' tools but not its messaging door has no supported way to say so. `--tools-only` serves the curated six safe tools (web_search, web_extract, vision_analyze, skill_view, skills_list, text_to_speech) and skips the event bridge: no messages_send, no approval pair, no polling thread. The flag is a declaration, not a subtraction: the tempting repair (deleting the profile's args) silently mounts messages_send and the approval pair. (cherry picked from commit 4952f91)
Drop 'browser' from the --tools-only help and server instructions (core no longer ships the browser_* tools), derive the tools-only expectation from the --tools listing instead of hardcoding browser_navigate, and leave the command description line alone so this lands in either order with #310. (cherry picked from commit cd21c06)
Ghidra/JDK discovery and the Kuna, TShark, capinfos, sigrok and WABT
engine locations all pass through environment variables, but Hermes
spawns stdio MCP children with a whitelisted environment, so the
published mcp.json (forwarding only REA_PWNTOOLS_PYTHON) leaves the
default native provider unreachable: open_binary fails with
GHIDRA_INSTALL_DIR not set and a java.exe absolute-path probe error.
The documented workaround -- a same-name mcp_servers.rea override --
replaces the plugin's pinned server and silently stops update
propagation; even a disabled squatter entry prevents the plugin server
from spawning.
Forward the eight vars with the same optional-forwarding ${VAR} idiom
already used for REA_PWNTOOLS_PYTHON: unset refs render as absent, so
this is a no-op for installs that set nothing. README: set them in
~/.hermes/.env (restart Hermes so the gateway adopts them); the
override path stays documented as a per-machine pin with its
version-drift caveat.
(cherry picked from commit 4f1ef61)
…ints at .env first
The whole-value unset ${VAR} drop that keeps an unset GHIDRA_INSTALL_DIR,
JAVA_HOME or REA_ANALYSIS_PROVIDER out of REA's environment is in every stable
tree from v0.22.3 (tools/mcp_tool_config.py unset_ref); v0.22.0-v0.22.2 pass the
literal and rea-agents 6.7.0 refuses to start. The catalog entry's
requires_hermes moves to >=0.22.3 at re-pin. The recovery guide still told the
agent the config override was the only way in.
Supersedes #235 (same GHIDRA_INSTALL_DIR line). Fixes #116.
(cherry picked from commit e79e2c6)
Core reads a per-server call deadline from the package manifest (extensions.com.nousresearch.hermes.servers.<name>.timeout), not from mcp.json, so the previous placement never reached it. Declare it where core looks and say so in the README; releases without per-server deadlines ignore it and keep their own default. (cherry picked from commit eff3d12)
v0.22.3 is the first stable tag whose tree reads extensions.com.nousresearch.hermes.servers.<name>.timeout (hermes_cli/agent_plugins.py _SERVER_DECLARATION_KEYS) and carries it through the executor (registry execution_timeout); v0.22.2 was rebuilt without it and refuses the package. Checked by load_agent_plugin on the v0.22.2, v0.22.3, v0.23.2 and main trees. The skill no longer tells the agent to hand-write timeout: 1000. (cherry picked from commit f962981)
… within one second Both changes got the same message id, built from the current second, so the gateway refused the second one and the listener reconnected. The id now uses the event's time_fired, which Home Assistant sets on every event, to the microsecond. Fixes #300 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> (cherry picked from commit 1cfaa59)
…o time_fired A KeyError here would end the listener and force a reconnect; sibling tests and any non-standard event source build events without it. (cherry picked from commit fd71c56)
(cherry picked from commit 709b4d0)
…ardown Port of NousResearch/hermes-agent#80372, closed there when Home Assistant moved to this plugin. The WebSocket handshake, the explicit closes on the connect and disconnect paths and the listen-task cancel in disconnect() now have a deadline. A failed or cancelled connect runs the same bounded close on its local session, and a close that does not finish in time stays tracked instead of being awaited forever, so a hung handshake or a wedged close can no longer leave the adapter stuck (NousResearch/hermes-agent#67470). Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> (cherry picked from commit 30e8c74)
The bounded connect closes the local session when ws_connect raises, so the fake needs closed/close(). (cherry picked from commit 05019cd)
This was referenced Nov 19, 2026
Closed
Closed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Landing slate for 34 approved contributor PRs across 12 plugins. Each PR's own commits are cherry-picked in the approved order with authorship kept (no merge commits). Conflicts are resolved once, here. Each touched plugin gets one version bump.
What lands
hermes-security (0.1.4 → 0.1.5)
memory-wiki (1.1.0 → 1.1.1)
mcp-serve (1.0.0 → 1.0.1)
browser_*tools that were removed from core.--tools-onlyserves the safe tool set without the messaging bridge.rea (0.2.0 → 0.2.1, bumped by #316)
>=0.22.3, which the README states; the catalog re-pin should setrequires_hermes: ">=0.22.3".homeassistant (2.0.1 → 2.0.2)
xchat (0.1.0 → 0.1.1)
spotify (2.1.0 → 2.1.1)
spotify.default_device_name).hermes spotify loginagain now requests the new scope.advanced-delegation (1.1.0 → 1.2.0, bumped by #363)
agent-importers (1.2.1 → 1.2.2)
memory_char_limitper store, and overflow is reported.cron resumehint names the imported profile.hermes profile create. If a profile cannot be created, the import stops with exit 1.camofox (1.1.0 → 1.1.1)
ALL_PROXYtoo.accent-picker (1.0.0 → 1.1.0, bumped by #358)
ctx.storageandctx.files. The fork patch script is not shipped.blender
Conflict resolutions (written once, here)
_load_memory_sessions): a single kwargs dict carries fix(memory-wiki): keep cron and subagent runs out of subjects and recent sessions #424'sexclude_sourcesand memory-wiki: include hidden sessions; bucket day logs by message date #262'sinclude_hidden=True, plus memory-wiki: include hidden sessions; bucket day logs by message date #262's narrowTypeErrorfallback for older cores.memory_wiki_core.py, 8 hunks): perf(memory-wiki): per-session cache so refreshes skip rescanning unchanged sessions (#49) #318's unit/cache architecture is kept, and every earlier PR's behaviour is carried into it:_recent_sessions(conversations_only=True)carries fix(memory-wiki): keep cron and subagent runs out of subjects and recent sessions #424's source filter and memory-wiki: include hidden sessions; bucket day logs by message date #262's hidden-session fallback.conversations_only=False.build_memory_subjectskeeps feat(memory-wiki): add Simplified Chinese UI and Han subject search #250'sexact_slugfilter, now applied to unit candidates.days: one bucket per local message day (memory-wiki: include hidden sessions; bucket day logs by message date #262's bucketing), each with its own candidates and work items. The cached daily log therefore matches the uncached scan.cache=._recent_sessions.__init__.pydescription): keeps fix(mcp-serve): drop the ten browser_* tools removed in Hermes 0.23.0 #310's "no browser" list and mcp-serve: add --tools-only, serving the six safe tools without the messaging bridge #395's--tools-onlyclause.spotify_tools.pyimports): keeps all three PRs' imports.Verification
git rev-list --merges origin/main..HEADis empty. Every PR's head commits are present, except the two empty retrigger commits (feat(spotify): a configurable default playback device #145790040c, [accent-picker] fix: render preset swatches horizontally #52f6a0263).+line of every PR is present in the final tree, except lines that were reshaped in the memory-wiki resolutions above and fix(agent-importers): pin the cron resume hint to the imported profile #290's hint, which feat(agent-importers): give each Grok Bot its own profile #394's own later commits rewrote.hermes plugins validatepasses for every touched plugin. The install scanner givessafefor all 12 touched dirs on core main and on v0.23.2.npm run buildof the combined src reproducesdesktop/plugin.jsbyte for byte.Issues
Fixes #390
Fixes #379
Fixes #304
Fixes #116
Fixes #91
Fixes #300
Fixes #13
Fixes #19
Fixes #12
Fixes #217
Fixes #266
Fixes #265
Fixes #215
Fixes #291
Part of #228
Part of #173
Part of #264
Part of #51
Supersedes
Supersedes #376
Supersedes #388
Supersedes #413
Supersedes #424
Supersedes #250
Supersedes #262
Supersedes #318
Supersedes #310
Supersedes #395
Supersedes #222
Supersedes #316
Supersedes #301
Supersedes #431
Supersedes #432
Supersedes #323
Supersedes #278
Supersedes #437
Supersedes #191
Supersedes #142
Supersedes #145
Supersedes #144
Supersedes #363
Supersedes #434
Supersedes #276
Supersedes #272
Supersedes #271
Supersedes #290
Supersedes #394
Supersedes #255
Supersedes #325
Supersedes #292
Supersedes #52
Supersedes #358
Supersedes #435
Follow-ups after merge: re-pin each touched plugin in hermes-agent
plugin-catalog/. For rea, setrequires_hermes: ">=0.22.3". For mcp-serve, the catalog description still says "browser". hermes-agent#148472 (agent-importers memory budget, fixed by #276) is in the core repo and has to be closed there.