Repository navigation
Releases
Releases · crunchtools/developer
v1.13.0
Added
- A josui-skills merge reaches a running seat (#42, #48).
developer-skills-syncruns at start and everyJOSUI_SKILLS_SYNC_INTERVALseconds (default 600), fast-forward only, and relinks skills without a restart. Each pass writesSKILLSSYNCdrift markers, anddeploy/holds the host push timers that feed nagios-agent'scheck_skills_sync.sh.
Changed
- claude-agent 1.2.0 (was 1.1.0): it requires tmux >= 3.2, logs keys it drops, and carries the supervisor fixes from crunchtools/claude-agent #30-#33. (#50)
Fixed
v1.12.0
Added
- Seat state for /agents (#44). A managed hook writes
~/.claude/seat-state.json: a prompt a person types marks the seat busy
and manual, with the session's first prompt as its task; a cross-session
message, which reaches the same hook, changes nothing./clearfrees the
seat and pins the new conversation's UUID, sodeveloper-worker-claude
does not resume the cleared one after a respawn. - Managed
autoCompactWindow: 500000, so seats compact well before their
conversations get huge. - An empty skills volume is cloned from
crunchtools/josui-skillsat start
(#42); an existing checkout is fast-forwarded as before.
Fixed
org.opencontainers.image.versionis this image's release, not the base
image's (it read 5.8.9 on 1.11.0). The boot check fails the build when it
andversiondisagree.
v1.11.0
Changed
- A worker seat (
JOSUI_MODE=worker) keeps one conversation, and so one claude.ai Remote Control entry, across restarts.developer-worker-claudepins a session UUID in~/.config/josui/session-id.- If that UUID's transcript holds a user turn, it runs
--resume <uuid>; otherwise it runs--session-id <uuid>. - A transcript with no conversation in it gets a new UUID rather than a respawn loop.
- If the UUID can't be saved, it warns and starts anyway.
- The tmux pane runs the wrapper. Josui Prime runs it through claude-agent's
CLAUDE_AGENT_CLAUDE. (#39)
- If that UUID's transcript holds a user turn, it runs
Fixed
- The entrypoint's usage comment names this image, not the retired
quay.io/crunchtools/josui. (#40)
v1.10.0
Changed
- claude-agent 1.1.0 (was 1.0.1), which holds Claude Code in tmux itself, on a server of its own. A texted worker (Josui Prime) no longer gets a
devsession runningclaude-agent attach --rwin a loop, which would now nest tmux. A person attaches withpodman exec -it <service> claude-agent attach --rw, and the startup line says so. Worker seats without a homeserver are unchanged.
RT #1632. The image tracks main, so :latest follows the gated :main build of this commit.
v1.9.1
Fixed
PYTEST_XDIST_AUTO_NUM_WORKERS=4in the image, so-n autocan no longer start one worker per host core and OOM the seat's Claude session. An explicit-n Nstill wins. (#36)
v1.6.0
Fixed
:latestfollows the gated:main(cascade: true). With no tag releases it
had frozen on the last image built before build-once (RT#1550).
Changed
- Builds move to constitution v1.39.1's container-release workflow: build once, gate the pushed digest (the toolchain check, run with podman against the digest), promote by digest. This image does not tag-release:
:mainfollows merges and the units pull:main. GHCR is dropped; Quay.io is the only registry. RT#1550
v1.5.1
Fixed
- The image carries a
versionlabel (1.5.0). The release sweep moves the
version in tracked files, and with none stating it outside the changelog it
refused to cut this repository's next release.
v1.5.0
Changed
- Gatehouse gets its own LLM endpoint on the gateway, not the one the L3
defense judge uses.GATEHOUSE_LLM_BASE_URLnow derives
/proxy/llm/openrouter-gatehouse/v1. The gateway serves OpenRouter twice
under two names, becausellm_keysis both the judge's credential
(resolved bydefense.provider) and what the proxy injects on
/proxy/llm/<name>— so one provider forced judging and Gatehouse onto one
key and made per-function spend unreadable. Two names, two keys, one
upstream, and all Gatehouse usage lands on the Gatehouse key wherever it
runs: seats, breetai, CI and Ashigaru. RT#1559
Fixed
- A seat runs the skills repository's default branch, whatever its volume
was seeded with. The entrypoint pulled whichever branch the checkout was
on. A volume left on a feature branch therefore pulled that branch's upstream
on every start, which never moved, and reported success: developer1 ran
skills three weeks and 91 commits behindmain, and/featurestill sent
deploys to lotor. It now refreshesorigin/HEAD, moves to that branch, and
then pulls. A move that would overwrite local changes is refused and logged,
and the checkout keeps working as it is.