Repository navigation
Releases
v0.2.3
Fixed
- You are told when long commands can no longer move to the background. Moving a long command to the background uses two Hermes functions (
tools.interrupt.request_yieldandconsume_yield) that are not on Hermes' published plugin API. If a Hermes release moves them, the plugin used to turn the feature off silently, and the main agent still believed long commands would move to the background. Now the plugin logs a warning, its prompt stops promising it, and the startup check tells the main agent once. Asked upstream for a public way to do this: NousResearch/hermes-agent#156468. - Tests ready for the next Hermes release. Hermes is removing
tools.interrupt.is_thread_yield_requested(NousResearch/hermes-agent#156130). The end-to-end test no longer uses it. The plugin itself never did.
Upgrading
hermes plugins update super-agentThen restart the gateway. No config changes are needed.
v0.2.2
New
- Running subagents hear your corrections. When you send a message while subagents of the chat are running, the main agent gets one short line naming them (goal and id) and is told to pass a change to one of their tasks on with
delegate_task(action='steer')instead of starting a second subagent. Hermes does not remind it which subagents are running, so after a long chat it could lose track. Async results and the plugin's own wake turns do not get the line.
Fixed
- Restarts are recognised even when the PID repeats. A restarted container often gives the gateway the same PID, and after a reboot an old PID may belong to another process. Subagents that were running then were never marked lost, so you were not told their results were gone, and they kept showing as running for hours. The plugin now also checks the process start time and boot id (on Linux). Other platforms keep the PID check.
Upgrading
hermes plugins update super-agentThen restart the gateway. No config changes are needed.
v0.2.1
New
- The main agent asks you after repeated failures. Once 2 subagents in a chat have ended without finishing (
failed,blocked,partial,interrupted) since your last message, the nextdelegate_taskis returned once with what they ran into. The main agent is told to tell you what was tried, what stopped it and what it proposes, and to let you decide. Your next message resets the count. Questions (needs_decision), tasks the main agent stopped itself and tasks lost to a restart are not counted. New settingfailed_task_limit(default2,0turns it off). - A subagent that ends without its report is not taken as done. Hermes calls such a run
completed. The plugin now records it asno_report, quotes its last words and the files it changed, and tells the main agent to check the real state instead of passing on "done" or sending the same plan again. hermes super-agent statusshows the settings in effect, the startup host check and the latest subagent tasks from the ledger: running ones first, then whether you have heard each result and whether a question is still open.--jsonprints JSON and--limit Nshows more tasks. It only reads. It is not available underplugins.isolation: host, where Hermes does not register plugin CLI commands.- Startup host check. On the first turn after the gateway starts, the main agent is told once about settings that weaken the plugin: a governed platform where
delegate_taskruns inline and blocks the chat, an unset or highdelegation.max_iterations, or a task ledger that could not be written to disk.
Docs
- New README banner, an image for each promise, a demo GIF, and a short README for the plugin's page in the Hermes catalog.
Upgrading
hermes plugins update super-agentThen restart the gateway. No config changes are needed.
v0.2.0 — Hermes Super Agent
hermes-proactive-agent is now Hermes Super Agent (hermes-super-agent, plugin id super-agent). The behaviour is the same; the name changed.
Upgrading from proactive-agent 0.1.x
- Install and enable the new plugin, then disable the old one:
hermes plugins install JerryLiu369/hermes-super-agent/super-agent hermes plugins enable super-agent hermes plugins disable proactive-agent - In
~/.hermes/config.yaml, renameplugins.entries.proactive-agenttoplugins.entries.super-agent. - To keep your task history, copy
<hermes home>/plugin-data/proactive-agent/ledger.dbto<hermes home>/plugin-data/super-agent/ledger.dbbefore the first start. - Environment overrides are now
SUPER_AGENT_<KEY>. The oldPROACTIVE_<KEY>names still work.
Chats that were running before the upgrade keep working: the plugin still recognises its old [proactive-agent] notes and markers.
v0.1.17: recommended subagent step budget raised to 90
proactive-agent 0.1.17
Recommended subagent step ceiling raised to 90, terminal background detach transitioned cleanly to Hermes' tool_execution middleware, and subagent / cron routing decoupled from private internal modules.
- Recommended subagent iteration ceiling raised to 90. The recommended
delegation.max_iterationshost configuration is now 90 (up from 60), accommodating complex multi-step worker tasks without prematurely cutting off iterative execution. - Background command detach through
tool_executionmiddleware. Long-running foreground terminal commands (>30 s) are now yielded to the background via Hermes' standard middleware mechanism rather than internal monkeypatching, cleanly preserving command output and notification handlers. - Decoupled execution context inspection. Caller identification for subagents, cron jobs, and governed sessions now relies strictly on public identifiers (
task_idprefix, session platform) without importing unexported Hermes modules. - Turn budget awareness. Context compression events during turns no longer count against the main agent's per-turn time budget, and
vision_analyzeremains available in the 120 s pause window. - Reporting format without schema friction. Worker reports cleanly request structured JSON reporting without rigid output schema rejection overhead.
Upgrade:
hermes plugins update proactive-agentRecommended host settings in ~/.hermes/config.yaml:
delegation:
max_iterations: 90
plugins:
entries:
proactive-agent:
allow_gateway_injection: trueRequires Hermes Agent 0.22 or later (tested on v0.22.7+) and Python 3.10+. All 97 tests pass on a real Hermes install, including full end-to-end plugin loader suites.
v0.1.16: results reach you, decisions come back
proactive-agent 0.1.16
Two gaps closed: a subagent result no longer gets lost when you write just as it arrives, and a subagent that needs a decision now brings the question back instead of guessing.
- A result reaches you even if you interrupt it. When a background subagent finishes and you send a message while the main agent is reporting, the main agent used to answer only your message. Now the result stays marked unreported until a reply you did not cut into. The next tool result reminds the main agent to answer you and report the result in the same reply. If it replied without a tool, the plugin wakes it once more for a short follow-up, or silence if it already covered it. Otherwise your next message carries the reminder. A result gets at most 2 wakes and is not chased after a day. The state lives in the ledger, so a gateway restart keeps it.
- Subagents bring decisions back instead of guessing (#7). A subagent cannot ask you anything, so it used to guess or give up. Now, when it reaches a choice that is not its to make (two workable approaches, a production change, spending, deleting data), it ends with status
needs_decision, a question and options. The defaultoutput_schemacarries these fields, and Hermes asks for a correction if the question is missing. The main agent answers the question itself when the chat, memory or facts it can check settle it, and asks you in one short message when it is your call. Then it sends a new subagent with the answer and the earlier task's record. If it has not followed up by your next message after you heard the question, it is reminded once. - Refused steps are not worked around. When Hermes' approval or one of your hooks refuses a subagent's step, or the step waits for an approval nobody can give during the run, the subagent is told right away not to retry it or find another way to the same result, and to bring the question back.
- Mid-turn messages are recognised in every form Hermes uses. A message that redirects the running turn now restarts the turn clock and joins your words in the next briefing, just like a steered one. Hermes' async-delegation rows and the plugin's own wake no longer count as your words, and Hermes' background memory review no longer resets the live turn.
- Docs rewritten. README and README_zh now read as plain explanation: what it is, install, why, what it does, principles and settings. The Chinese README calls the main agent "助手".
llms.txtandllms-full.txtremain the machine-readable reference.
Upgrade:
hermes plugins update proactive-agentTo let the plugin wake the main agent for an interrupted report, also set this (it is off by default in Hermes):
plugins:
entries:
proactive-agent:
allow_gateway_injection: trueRequires Hermes Agent 0.22 or later (tested on v0.22.7) and Python 3.10+. All 91 tests pass on a real Hermes v0.22.7 install, including the end-to-end tests through Hermes' plugin loader. These cover a needs_decision report validated by Hermes' own output_schema checker and Hermes' own approval-refusal results.
v0.1.15: talks only to the main agent
proactive-agent 0.1.15
The plugin now talks only to the main agent, the turn budget uses Hermes' real tool names, and the code and docs are reorganised around the four things the plugin promises.
- Nothing goes to your chat directly. A worker stopped for repeating itself is now reported to the main agent only. Its next turn opens with the note and it tells you in one line what got stuck and what it will do instead. The plugin itself no longer posts any message.
- Turn budget uses Hermes' registered tool names. After 120 s the tools that stay available are checked against what Hermes actually registers:
todo_list(nottodo), andsend_messageis dropped because Hermes does not register it for the main agent.skill_manageis available again, so the main agent can still save what it learned while wrapping up. An end-to-end test now fails if any of these names drifts from Hermes' tool registry. - Cron runs and subagents are recognised only by what Hermes uses. These are
cron:<job>:<run>task ids,cron_<job>_<time>sessions,sa-<n>-<hex>subagents, Hermes' cron contextvars, and a truthyHERMES_CRON_SESSION(truenow counts, not only1). Guessed markers were removed. - One rule for "the file was changed". A write counts unless Hermes reports it as
error, both for live tool calls and for the subagent's final tool history. - One module per promise.
secretary.pykeeps the chat responsive,handoff.pybriefs subagents,workers.pystops ones that spin, and the newreports.pywithledger.pychecks and records results (including restart recovery). - Docs rewritten (README, README_zh, llms.txt) around the five problems, the four promises and the design rules. They now explain that repeats are counted over the whole run, and that Hermes' own loop guard only warns subagents, which is the gap the repeat stop fills.
Upgrade:
hermes plugins update proactive-agentRequires Hermes Agent 0.22 or later (tested on v0.22.7) and Python 3.10+. All 63 tests pass on a real Hermes v0.22.7 install, including the end-to-end tests through Hermes' plugin loader.
v0.1.14: task ledger
proactive-agent 0.1.14
Task ledger release. Workers now leave a record that the main agent and later workers can build on, and the plugin no longer posts on a timer.
- Task ledger. One record per delegated task: goal, status, files it changed, corrections sent with
action='steer', its live transcript path, and how it ended (summary, evidence, blockers). Written by the plugin's hooks only, with no model calls, in<hermes home>/plugin-data/proactive-agent/ledger.db. Each chat keeps its last 20 finished tasks for up to 7 days. - Follow-up workers start from what the last one found. Hermes cannot resume a finished subagent, so the next worker's briefing lists the subagents running now (and the files they change) and this chat's recent unfinished tasks.
- Restarts are covered. Tasks that were running when Hermes stopped are marked lost, and the main agent is told their side effects are unknown and which files they had changed. Result and restart notes wait in the ledger until the chat's next turn, so a restart, or a CLI process loading the plugin first, does not lose them.
action='stop'is recorded at once. A stopped task leaves the running list right away, its partial work still reaches the next worker, and its result note says it was stopped on purpose.- File clash check. A
delegate_taskthat names a file another running worker is changing is returned once per turn. - No more timed progress lines. The 10-minute progress messages and
progress_interval_secondsare gone (a leftover key inconfig.yamlis ignored). The only message the plugin posts itself is the alert when a worker is stopped for repeating itself. - Turn budget: after 120 s,
skill_manageis paused too, andprocess_manageis limited to list, poll, log and kill. - Fixes:
/newnow cleans the session that ended (old_session_id) instead of the new one, and an empty session id no longer clears every chat. Worker reports are read only from a real JSON report, so quoted JSON in plain text cannot change a task's status.
Upgrade:
hermes plugins update proactive-agentRequires Hermes Agent 0.22 or later (tested on v0.22.7) and Python 3.10+. All 61 tests pass on a real Hermes v0.22.7 install, including the end-to-end tests through Hermes' plugin loader.
v0.1.13: plugin in its own directory
proactive-agent 0.1.13
Layout release: no behaviour changes.
- The plugin has its own directory.
plugin.yamland all modules live inproactive-agent/. Hermes installs only that directory, so tests and docs no longer land in~/.hermes/plugins/. - New install command. The old README command with a
github:prefix never worked (Hermes does not accept that prefix). Use:
hermes plugins install JerryLiu369/hermes-proactive-agent/proactive-agent
hermes plugins enable proactive-agent- Tooling: ruff, configured in
pyproject.toml, replaces flake8.
Requires Hermes Agent 0.22 or later (tested on v0.22.7) and Python 3.10+. Verified by installing from GitHub into a fresh HERMES_HOME on Hermes v0.22.7.
v0.1.12: tidy layout, tooling and docs
proactive-agent 0.1.12
Housekeeping release: no behaviour changes.
- Manifest:
plugin.yamldeclaresrequires_hermes: ">=0.22",license,homepageandtags, and lists hooks only under the canonicalprovides_hooks. - Tooling:
pyproject.tomlcarries real metadata and adevdependency group (uv run --group dev pytestorpip install --group dev); one.flake8config for local runs and CI. - CI: Python 3.10 to 3.14.
- Tests: one test file per module, with shared helpers.
- Docs: README and README_zh reordered (install first, contents, how it works, features, configuration, development with a project layout), with a live release badge.
Install
hermes plugins install github:JerryLiu369/hermes-proactive-agent
hermes plugins enable proactive-agentRequires Hermes Agent 0.22 or later (tested on v0.22.7) and Python 3.10+.