/hermit-evolve
Evolves hermit configuration and templates after a plugin update. Detects version gaps and runs the upgrade (migrations, templates, new settings) in an isolated subagent. Run after updating the plugin.
Claude Code
$ npx -y skills add gtapps/hermitd --skill hermit-evolve --agent claude-code
Codex
$ npx skills add gtapps/hermitd --skill hermit-evolve -a codex -y
Cursor
$ npx skills add gtapps/hermitd --skill hermit-evolve -a cursor -y
Gemini CLI
$ npx skills add gtapps/hermitd --skill hermit-evolve -a gemini-cli -y
GitHub Copilot
$ npx skills add gtapps/hermitd --skill hermit-evolve -a github-copilot -y
OpenCode
$ npx skills add gtapps/hermitd --skill hermit-evolve -a opencode -y
Windsurf
$ npx skills add gtapps/hermitd --skill hermit-evolve -a windsurf -y
SKILL.md
hermit-evolve.SKILL.mdname: hermit-evolve
description: Evolves hermit configuration and templates after a plugin update. Detects version gaps and runs the upgrade (migrations, templates, new settings) in an isolated subagent. Run after updating the plugin.
Evolve Hermit
Upgrade the project's hermit configuration after a plugin update.
Commands
- `evolve-adopt`: `bun ${CLAUDE_PLUGIN_ROOT}/scripts/evolve-plan.ts .hermit adopt`
- `evolve-apply`: `bun ${CLAUDE_PLUGIN_ROOT}/scripts/evolve-plan.ts .hermit apply`
- `leftover-check`: `bun ${CLAUDE_PLUGIN_ROOT}/scripts/evolve-plan.ts .hermit leftover-check`
Execution routing
Every run of this skill (interactive or unattended) delegates steps 0–9 to the `hermitd:evolve-runner` subagent, so the upgrade's transient churn (changelog slice, migration execution, file diffs) never lands in the calling session. The main loop keeps only step 10 (summary + operator notification).
> **Tool note:** `hermitd:evolve-runner` is a **subagent** — invoke it via the Agent tool, never the Skill tool. The `plugin:name` form it shares with skills does not imply the Skill tool.
- **If you are running AS the `evolve-runner` subagent**, skip this section and execute steps 0–9 directly (you are the delegate — do not re-dispatch).
- **Otherwise (main loop, any mode):**
1. **Determine execution and delivery separately** and remember both for step 10:
- Positional argument `unattended` ⇒ execution *unattended*, delivery *automated-maintainer*.
- No `unattended` argument + an inbound `<channel source="...">` tag ⇒ execution *unattended*
(never block on `AskUserQuestion`), delivery *direct-channel-reply*.
- Otherwise ⇒ execution *interactive*, delivery *inline*.
Only the explicit `unattended` argument authorizes a proactive maintainer notification. A direct channel request must answer the channel that invoked it. 2. **Check the runner reference:** run `test -f "${CLAUDE_PLUGIN_ROOT}/skills/hermit-evolve/reference.md"`. If it fails, **abort** — log `"hermit-evolve aborted: plugin root unresolved; cannot dispatch evolve-runner."` and stop. Do not dispatch with a broken path. 3. **Record `<now>`** for step 10 as the current UTC time (`date -u +%FT%TZ`), then **dispatch** the `hermitd:evolve-runner` subagent via the Agent tool. Pass the report contract (below); the runner body supplies its own plugin root and Commands entries. Do **not** execute steps 0–9 yourself. 4. **Go to step 10** with the subagent's returned report.
Delegated mode
Steps 0–9 (in `reference.md`, read only by the `evolve-runner` subagent) are executed with no `AskUserQuestion` — the subagent cannot pause to ask. Each step's **"Delegated mode:"** note states the non-interactive behavior. The rule in every case: never guess on a destructive choice, never block.
- For any interactive choice with a safe non-destructive default (new settings, file deletions, `## Plan` strip, template conflicts, an `### Upgrade Instructions` step that names a single non-destructive command such as `deny ask-only`), take the default silently and report the outcome.
- For a genuine either/or with **no safe default** (an `### Upgrade Instructions` migration step in 2b/7 that poses a choice and names no default), **defer**: skip that step and record a verbatim deferred-migration block in the report (see the report contract in step 10). Never guess. Step 10 resolves it — interactive asks the operator; direct-channel and automated-maintainer execution relay it through their respective delivery routes. Never apply a deferred instruction that would write `.claude/settings*.json` from a channel session.
10. Report
After a successful upgrade, arm `/hermitd:later add 1d "doctor has no new failures after upgrading to <version>" | .hermit/bin/hermitd-run doctor-check .hermit --verdict --since <now>` (hermit origin, `--timeout-s 120`: the doctor run probes credentials and docker and gets the same budget as its routine). `<now>` is the time recorded before dispatch, so a failure that already existed before the upgrade does not count against it while one first seen during the runner's steps still does. The resolver form survives the next plugin update; a baked plugin path points at a cache directory that may be gone by the time the claim is checked.
**Step 10 runs in the main loop** (not the subagent), consuming the `evolve-runner`'s returned report. The subagent's report is the single source for what follows.
Build the Templates, Bin wrappers, Docker entrypoint and Docker templates lines from the runner's `evolve-apply` JSON receipt plus its completed `remaining` resolutions. A parked file is not a merged file.
**Report contract** — the subagent's final message is exactly this; non-deferred runs carry no deferred block, so the common payload is tiny:
Upgrade: vOLD -> vNEW | core current vNEW | blocked: <reason>
Settings added: <keys | none>
Templates: <refreshed/restored/kept-N/conflicts-parked-N | none>
Bin wrappers: <restored/replaced(.bak) | none>
Docker entrypoint: <refreshed | conflict-replaced(<backup path>) | migrated(<N> moved, <M> in <patch path>) | n/a>
Docker templates: <name merged(3-way[; reviewed[, n lines fixed] | ; n conflicts resolved]) | kept(bootstrap, upstream not merged: <path>) | conflict(<n | review | reason>): upstream copy at <path>; ... | report-only(<names>) | none>
Docker rebuild: <needed + order | base-patched | no>
Sibling hermits: <one or more of the following per sibling, space-separated, or "none">
<name vOLD->vNEW> (confirmed by finalizer — only from siblings_confirmed)
<name current> (no version gap)
<name path-unresolved> (in _hermit_versions but no project-effective plugin-list match)
<name SKIPPED-by-finalizer> (finalizer's siblings_skipped — never report as upgraded)
Siblings detected but not activated: <name ... | none>
Siblings warnings: <one line per siblings_warnings entry | none>
Permissions added: <entries |
Read more
name: hermit-evolve description: Evolves hermit configuration and templates after a plugin update. Detects version gaps and runs the upgrade (migrations, templates, new settings) in an isolated subagent. Run after updating the plugin.
Evolve Hermit
Upgrade the project's hermit configuration after a plugin update.
Commands
- `evolve-adopt`: `bun ${CLAUDE_PLUGIN_ROOT}/scripts/evolve-plan.ts .hermit adopt`
- `evolve-apply`: `bun ${CLAUDE_PLUGIN_ROOT}/scripts/evolve-plan.ts .hermit apply`
- `leftover-check`: `bun ${CLAUDE_PLUGIN_ROOT}/scripts/evolve-plan.ts .hermit leftover-check`
Execution routing
Every run of this skill (interactive or unattended) delegates steps 0–9 to the `hermitd:evolve-runner` subagent, so the upgrade's transient churn (changelog slice, migration execution, file diffs) never lands in the calling session. The main loop keeps only step 10 (summary + operator notification).
> **Tool note:** `hermitd:evolve-runner` is a **subagent** — invoke it via the Agent tool, never the Skill tool. The `plugin:name` form it shares with skills does not imply the Skill tool.
- **If you are running AS the `evolve-runner` subagent**, skip this section and execute steps 0–9 directly (you are the delegate — do not re-dispatch).
- **Otherwise (main loop, any mode):**
1. **Determine execution and delivery separately** and remember both for step 10:
- Positional argument `unattended` ⇒ execution *unattended*, delivery *automated-maintainer*.
- No `unattended` argument + an inbound `<channel source="...">` tag ⇒ execution *unattended*
(never block on `AskUserQuestion`), delivery *direct-channel-reply*.
- Otherwise ⇒ execution *interactive*, delivery *inline*.
Only the explicit `unattended` argument authorizes a proactive maintainer notification. A direct channel request must answer the channel that invoked it. 2. **Check the runner reference:** run `test -f "${CLAUDE_PLUGIN_ROOT}/skills/hermit-evolve/reference.md"`. If it fails, **abort** — log `"hermit-evolve aborted: plugin root unresolved; cannot dispatch evolve-runner."` and stop. Do not dispatch with a broken path. 3. **Record `<now>`** for step 10 as the current UTC time (`date -u +%FT%TZ`), then **dispatch** the `hermitd:evolve-runner` subagent via the Agent tool. Pass the report contract (below); the runner body supplies its own plugin root and Commands entries. Do **not** execute steps 0–9 yourself. 4. **Go to step 10** with the subagent's returned report.
Delegated mode
Steps 0–9 (in `reference.md`, read only by the `evolve-runner` subagent) are executed with no `AskUserQuestion` — the subagent cannot pause to ask. Each step's **"Delegated mode:"** note states the non-interactive behavior. The rule in every case: never guess on a destructive choice, never block.
- For any interactive choice with a safe non-destructive default (new settings, file deletions, `## Plan` strip, template conflicts, an `### Upgrade Instructions` step that names a single non-destructive command such as `deny ask-only`), take the default silently and report the outcome.
- For a genuine either/or with **no safe default** (an `### Upgrade Instructions` migration step in 2b/7 that poses a choice and names no default), **defer**: skip that step and record a verbatim deferred-migration block in the report (see the report contract in step 10). Never guess. Step 10 resolves it — interactive asks the operator; direct-channel and automated-maintainer execution relay it through their respective delivery routes. Never apply a deferred instruction that would write `.claude/settings*.json` from a channel session.
10. Report
After a successful upgrade, arm `/hermitd:later add 1d "doctor has no new failures after upgrading to <version>" | .hermit/bin/hermitd-run doctor-check .hermit --verdict --since <now>` (hermit origin, `--timeout-s 120`: the doctor run probes credentials and docker and gets the same budget as its routine). `<now>` is the time recorded before dispatch, so a failure that already existed before the upgrade does not count against it while one first seen during the runner's steps still does. The resolver form survives the next plugin update; a baked plugin path points at a cache directory that may be gone by the time the claim is checked.
**Step 10 runs in the main loop** (not the subagent), consuming the `evolve-runner`'s returned report. The subagent's report is the single source for what follows.
Build the Templates, Bin wrappers, Docker entrypoint and Docker templates lines from the runner's `evolve-apply` JSON receipt plus its completed `remaining` resolutions. A parked file is not a merged file.
**Report contract** — the subagent's final message is exactly this; non-deferred runs carry no deferred block, so the common payload is tiny:
Upgrade: vOLD -> vNEW | core current vNEW | blocked: <reason> Settings added: <keys | none> Templates: <refreshed/restored/kept-N/conflicts-parked-N | none> Bin wrappers: <restored/replaced(.bak) | none> Docker entrypoint: <refreshed | conflict-replaced(<backup path>) | migrated(<N> moved, <M> in <patch path>) | n/a> Docker templates: <name merged(3-way[; reviewed[, n lines fixed] | ; n conflicts resolved]) | kept(bootstrap, upstream not merged: <path>) | conflict(<n | review | reason>): upstream copy at <path>; ... | report-only(<names>) | none> Docker rebuild: <needed + order | base-patched | no> Sibling hermits: <one or more of the following per sibling, space-separated, or "none"> <name vOLD->vNEW> (confirmed by finalizer — only from siblings_confirmed) <name current> (no version gap) <name path-unresolved> (in _hermit_versions but no project-effective plugin-list match) <name SKIPPED-by-finalizer> (finalizer's siblings_skipped — never report as upgraded) Siblings detected but not activated: <name ... | none> Siblings warnings: <one line per siblings_warnings entry | none> Permissions added: <entries |
Run an always-on Claude Code agent on your own machine, for you or your team. Start and orchestrate sessions from chat in any project it can reach. Take over the same session in the Claude app or the CLI.
Other skills on hermitd.
More skills from other plugins.
browser-use
Direct browser control via CDP for web interaction: automation, scraping, testing, screenshots, and site/app work. Also covers…
agent-browser
Browser automation CLI for AI agents. Use when the user needs to interact with websites, including navigating pages, filling…
Questions about hermit-evolve.
- Which coding agents can use hermit-evolve?
- hermit-evolve installs in Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, OpenCode and Windsurf with the skills CLI, using each agent's flag shown on this page. Flowy's auto-invocation, which fires a skill when your prompt matches, runs in Claude Code only.

