hermitd--v1.4.24

Added

• Chat approval on channels that relay permission prompts for spawn-session to trust a refused repository and its MCP servers, with parsed activation facts in the prompt
• Script-sent channel notices (watchdog, status reply, budget, denial and login messages) follow any configured language: the agent translates them once into .hermit/state/messages.json when the language is set and after upgrades, and every send falls back to English per message
• /hermitd:hermit-settings messages rewords any of those notices; custom wording survives upgrades and retranslation, and doctor's message-catalog check flags wording an upgrade made unusable
• hermitd agents [name] opens Claude Code's claude agents view where the agent runs, inside the container for Docker agents

Changed

• Skill reviews run after a task closes using its own open window, best effort and only where learning can write
• Reflect checks for enforcement, missed triggers and existing guidance before proposing learning changes
• Close-time learning reviews once in the main session after closure, best effort and only where learning can write, patching learned skills from reusable procedures or general primary-operator corrections
• Accepted skill improvements apply requested rewording, moves and description changes to existing text
• Recall answers distinguish stated evidence from inference and report search limits
• skills status --skill reports the latest recorded use
• Portuguese script notices come from the agent's own translation instead of shipped hand-written text, so their wording changes after the upgrade
• Channel reply rules keep each chat's content in that chat and resolve a bare go-ahead against the chat it arrived in before looking at other chats
• Tool sign-in requires an admin to finish, with a gate on account binding that raises the native permission prompt on a turn nobody typed or another prompt joined, sensitive links routed to the maintainer chat when configured, failed delivery reported, and local callback addresses redacted from the channel log
• proposal-triage no longer reads frozen session reports, and its unused prior_discussion metadata line is gone

Fixed

• Best-effort skill reviews run after closure outside the task skill too, only where learning can write; lost reviews are not resumed
• Closing one task no longer uses up the next task's best-effort review after closure; reviews run only where learning can write
• Morning briefs carry the newest record's task context instead of varying between mornings
• Scheduled briefs read the full day's records, and plain briefs read only the records they need
• Learned-skill retirement counts use recorded under worktree and plugin keys
• Proposal gate accepts SUPPRESS and DUPLICATE verdicts whose reason was dropped, so a shortened line records the verdict instead of a false gate-failed
• hermitd update and watchdog restarts no longer wait out the 15-minute quiet cap on an agent that was cleared and then left quiet
• Chat !clear and !compact report success on Claude Team and Enterprise accounts instead of "outcome unknown"
• The first artifact page republish after a restart or /clear runs in a subagent, keeping the live page's source out of the agent's context
• Auto-mode denial notices suggest telling the agent in your usual chat when the action was intended, instead of /hermit-settings or the terminal
• hermitd attach forces UTF-8 output (tmux -u), so Claude Code's bullets, arrows and status symbols render when attaching, including inside Docker; the Docker image sets LANG=C.UTF-8 after the next hermitd update
• After the agent creates or edits a project skill, it is told the change loads on its own, so it can follow the file instead of advising a restart or new session.
• Chat ! commands keep working after the agent's shell changes directory, and doctor run from a worktree or child checkout reports on the real .hermit/ instead of creating state/ in its copy
• Channel setup tests the selected channel's maintainer chat without falling back to another channel, and watchdog notices that only reached state/watchdog-events.jsonl record a failed push
• Relogin cancels and reports a sign-in link that can't reach the maintainer chat, and keeps that link out of the searchable channel log
• A boot after /clear no longer counts as a context refresh, which could skip /hermitd:watch start and the online ping

Upgrade Instructions

Run /hermitd:hermit-evolve. Its retranslation step writes .hermit/state/messages.json for an agent whose language is set and not en, so its script notices keep speaking that language; English agents get no file. Step 8 adds the messages.ts set and reset permission entries and seals Bash(.hermit/bin/hermitd-run trust-workspace check *) and Bash(bun */scripts/trust-workspace.ts check*).

Delete .hermit/state/skill-review-counter if it exists.
Use the <resolved-settings-file> Step 8 resolves from hatch_target. Run bun <plugin_root>/scripts/apply-settings.ts <resolved-settings-file> deny ask-only to merge the trust-workspace and messages.ts set/reset ask entries. Unless it reports {"status":"skipped","reason":"not_seeded"}, and only when the same file's permissions.deny already holds Edit(*.hermit/state/resident-system-prompt.md) (Hardened), run bun <plugin_root>/scripts/apply-settings.ts <resolved-settings-file> deny-add 'Bash(bun */scripts/trust-workspace.ts apply *)' 'Bash(*.hermit/bin/hermitd-run trust-workspace apply*)' 'Bash(hermitd run trust-workspace apply*)' 'Bash(*/hermitd run trust-workspace apply*)' 'Bash(bun */scripts/messages.ts set *)' 'Bash(*.hermit/bin/hermitd-run messages set*)' 'Bash(hermitd run messages set*)' 'Bash(*/hermitd run messages set*)' 'Bash(bun */scripts/messages.ts reset *)' 'Bash(*.hermit/bin/hermitd-run messages reset*)' 'Bash(hermitd run messages reset*)' 'Bash(*/hermitd run messages reset*)'. These operations are additive and idempotent; report the added entries or the unseeded skip.

Leftover check: task.ts learn verb and task.ts list --skill-review flag removed; skill_review dropped from task records, task.ts list rows and close digests.

No config.json changes required.