Repository navigation
Plan mode, subagent reporting, generate_image, mid-turn compaction, Windows fixes - #10
Merged
Merged
Conversation
The agent can now draw or edit images through the provider's
OpenAI-compatible images endpoints: POST {base_url}/images/generations
(JSON) for a prompt, POST {base_url}/images/edits (multipart, image[])
when workspace images are given. It uses the chat requests' bearer key
and per-model routing headers (X-JuCode-Group), a
read timeout of at least 180 s, and accepts b64_json or url results.
gpt-image models are not sent response_format; others ask for b64_json.
Images are saved in the workspace under write's rules (workspace and
sandbox confinement, subagent write isolation, approval in manual mode)
at the requested path or images/<timestamp>-<slug>.<ext>; an existing
file is never overwritten, a numeric suffix is added instead.
The model is the call's `model`, else the new `image_model` config key,
else the first configured model whose name contains "image". The tool is
offered only when such a model resolves on a responses or chat endpoint
(not anthropic, codex or azure) with an API key.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A new approval mode, plan (like Codex's and Claude Code's): only read-only tools run (reads, search, web, provably read-only shell commands, MCP tools marked readOnlyHint, subagents in plan mode too); everything else is refused with a note to plan instead. The model delivers its plan with propose_plan (goal, a File | Change | Why table, steps, risks, verification), which emits proposed_plan and ends the turn. approve_plan approves it and runs it in the chosen mode, or revises it with feedback. Plans are saved in the session and replayed with their latest status. The TUI cycles through plan with Shift+Tab. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A native subagent ran on its own thread and only told front-ends that it started and finished, so the desktop showed nothing of it. Each subagent now keeps a bounded trace (its task, messages, reasoning, tool calls with input and output, its latest action as one line, usage; at most 300 items and 256 KB). The engine answers the agent-trace ops the desktop already uses for Claude Code and Codex: agent_runs (also pushed while agents work, twice a second at most) and subagent_transcript. subagent_lifecycle carries the agent's label, model and the spawn_agent call it came from. Agents of earlier turns stay listed. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The root binary's headless stats matched every event and broke the build. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A rename over config.json fails on Windows while the desktop or another engine has it open (PermissionDenied); a model pick then did not save. Retry briefly before giving up. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A message sent while the agent works used to stop the turn (killing running tools and subagents) and start again. It now goes into the running turn: the model reads it before its next request, after the current tool calls finish, and it is saved and shown as the user's message. If the turn ends before the model reads it, it runs next. A message with images still restarts as before. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The ripgrep tool failed on every machine without rg (a stock Windows has none). It now searches in-process when rg cannot start, printing what rg prints for the options the tool takes. A title model that refuses its lightest effort (gpt-6-astra takes no none) left every conversation untitled; the title request now retries once with the next effort up. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A long turn of tool calls only checked the context window when it started, so it ran up to the limit without compacting. The turn now stops between requests once the last one passed the compaction threshold, and the core compacts and continues it (the overflow path, with its own notice). Without subagent_models, spawn_agent offers every chat model of the provider (image models left out); the agent picks per task and takes the one the user names. write_stdin to a session that is not running says that a finished bash command already returned its output. A timed-out command in the sandbox suggests escalate for programs that write outside it. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The catalog lists the JuCode marketplace, Anthropic's skills, and now
Ikaleio, Superpowers (obra) and Composio, read from each repository's
SKILL.md files wherever they sit (skills/ or the top level), an hour at a
time. A source that cannot be read is a warning. JUCODE_SKILL_SOURCES
("Name=https://github.com/o/r;…", or empty for none) replaces the
community list. Frontmatter reads folded and literal block values.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
ToolArgumentsDelta lets the plan stream while the model writes it; response.reasoning_text.delta is read as reasoning. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The parsers emit tool-call argument fragments; the core reads the title and plan out of propose_plan's half-written JSON and sends plan_draft events with the text added since the last one. The finished plan follows as proposed_plan with the same id, so a client grows one plan in place. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The main turn stopped for compaction whenever the last request's input passed the threshold, even when nothing before the current user message could be folded; the retry then had no plan and the turn failed. It also compared API usage (system prompt and tools included) with a tokenizer count, so it could fire on every request, and closed running subagents. Now the core arms it only for a turn that starts uncompacted with earlier history to fold, and measures how much the input grew since the turn's first request against the headroom left under the threshold. The continuation compacts at its start and is not armed again, keeps the turn's subagents and unread steers, and a mid-turn stop always continues the turn (compacting if it can) instead of failing it.
git grep -O/--open-files-in-pager runs a program; git reflog expire and delete rewrite history (only reflog and reflog show run); --textconv and cat-file --filters run filters. sort's -o inside a short-option cluster (-uo) and --compress-program, file -C/--compile, tree -o in a cluster and -R (it writes 00Tree.html files), and rg --hostname-bin are refused too.
A message steered into the running turn reached the model without the user_prompt_submit hook that a new turn runs; it now runs first, and a block drops the message with the same error. A turn interrupted or failed before reading a steered message lost it from view, and the next turn requeued it behind the message that started that turn. The core now requeues what the session has not recorded (including a message the worker read whose event was dropped) as soon as the turn stops, ahead of later messages, and shows it as pending.
An empty subagent_models list offered every chat model of the provider; it is back to meaning subagents run on the main agent's own model, as config documents. With a list, the guidance still tells the agent to use the model the user names, and is left out when there is nothing to pick.
The desktop catalog fetched three community repositories by default. They are now listed only when JUCODE_SKILL_SOURCES names them, so a default install shows the JuCode marketplace and Anthropic's index alone; the docs describe the variable and the catalog's sourceName field.
The TUI shows a proposed plan but cannot approve it, so a session cycled into plan mode ended every turn on a plan that stayed pending.
The retry covers Windows refusing to replace a file another process has open; on Unix a permission error is permanent and now fails at once instead of after five seconds of retries.
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
2 participants
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.
The CLI side of the last two days' work, ported onto JuCode's
mainwith JuCode's names, hosts and skill marketplace kept. One commit per change, each with its own description.Depends on JuCode-Team/llm-provider-kit#1: the submodule points at that PR's head (
51fd509). Merge the kit PR first.What's in it
propose_plan→proposed_planevent). The plan runs once the user approves it. While the model writes the plan, it streams to clients asplan_draftevents.agent_runsandsubagent_transcriptops/events; headless output counts them.generate_imagetool. Draws or edits images through the provider's OpenAI-compatible/images/*endpoints.model, elseimage_modelin config.json, else the first configured model named like an image model.write.rg(it searches in-process when rg can't start).JUCODE_SKILL_SOURCESreplaces the community list; set it empty for none.Checks
cargo fmt --all --checkcargo clippy --workspace --all-targets -D warningscargo test --workspace: all passing.🤖 Generated with Claude Code