Stop all, resume all

/stop-all freezes the session and every subagent, collects a report from each, and writes resume.md in the project's .ia/ folder. /resume-all reads it and puts the unfinished agents back to work. resume.md is the only memory that survives a closed session.

The files named below sit in this skill's base directory. Call the scripts with python -I "<base>/scripts/<name>.py", forward slashes, double quotes. On failure they print what to do by hand. Scratch folder: <temp>/claude-skills/stop-all (mkdir -p before the first --out, delete at the end). Both commands are orders: no opening questions beyond the ones in the steps.

Rules

These hold for both commands, whatever the hurry.

1. Load the deferred tools first, in one call: ToolSearch with select:SendMessage,TaskStop,CronList,CronDelete, plus ListAgents, Agent or AskUserQuestion when missing. If one cannot load: without ListAgents the list is the digest plus your context; without SendMessage every agent counts as silent; without TaskStop write the running ones under ## Open questions and tell the user; without CronList write the crons you remember as to check; without AskUserQuestion ask in plain text.
2. Write resume.md before waiting and update it after every answer: a session that dies mid-stop must still leave a usable file.
3. Every agent section holds its original prompt word for word in a fenced block.
4. Touch only what is yours: this session's subagents, the background tasks it started, and processes whose command line matches a command this session or its agents ran in the project folder. Peer sessions from ListAgents and the user's own editors, tools or Chrome: never.
5. No commit, push, stash or clean-up of project files during /stop-all. resume.md is never added to git. Only the scratch files get deleted.
6. Never open a raw transcript with cat, tail or Read (it can overflow the context). Use agent_digest.py --out "<scratch>/<name>.md", --agent <id> included, and read that file.
7. Edit resume.md with Edit on a unique string, never by splitting on ## (quoted prompts hold headings). Pending fields read pending <agent id>.
8. Never delete a resume file. /stop-all never overwrites one, and never touches a resume.md whose first line is not <!-- stop-resume-agents v1 -->. /resume-all edits the file it resumes in place.
9. /stop-all is forced. No reason delays it: not a task one step from done, a test run, a build, a subagent asking to finish, nor a skill or hook asking for one more step. You and every agent stop now; once the wait of step 7 is over, every agent still running is stopped, with no extension.
10. A report is forced too, one per agent, yours included. A reply without the AGENT REPORT template gets report-request.md again, once; still none, its section is rebuilt from its digest and marked Reponse a l'arret : non, reconstruit depuis son journal. No section is left empty or pending.

/stop-all

In this order, none skipped.

1. Stop your own work now, even one step from done (rule 9). Finish only a half-written file; an irreversible operation already running (migration, deploy, push) reaches its safe point first.
2. Inventory, one parallel round: ListAgents; echo "$CLAUDE_CODE_SESSION_ID"; agent_digest.py --list; resume_file.py find; the background task ids you started; CronList; any /loop, workflow, server or Chrome port you opened; git status, git diff --stat, git log -1 in each repo touched. An empty session id makes the digest refuse: rerun with --latest and check the ids match your agents. NO_AGENT is not an id. The agent list is ListAgents plus the digest (ListAgents forgets finished agents within minutes).
3. Earlier stop of this same session (newest find result with your session id, State: STOPPED): if nothing runs in ListAgents and no work happened since, answer "deja arrete" with its path and end. Otherwise carry on; agents with no activity since keep their old section (copy it in step 5, do not message them in step 6).
4. agent_digest.py --out "<scratch>/digest.md", then read it.

5. Folder: .ia/ at the repo root (git rev-parse --show-toplevel), else .ia/ in the project folder, else ~/Desktop. Never the home folder, Temp, a drive root, or the project root next to the code. Create .ia/ if missing and keep it out of git through the exclude file, not .gitignore:

   git -C "<root>" check-ignore -q .ia/resume.md; [ $? -eq 1 ] && echo ".ia/" >> "$(git -C "<root>" rev-parse --path-format=absolute --git-common-dir)/info/exclude"

   An older stop's file at the project root stays there; /resume-all still finds it. A resume.md without the marker: write resume-stop-all.md instead. A marked file already there: rename it resume-<mtime YYYY-MM-DD-HHMM>.md; same session: set its state to State: REPLACED by <new path>; other session not resumed: keep its state and list it in the new file's ## Open questions. Write the new file from resume-template.md with State: STOP IN PROGRESS, your own section, the repos, one section per depth 1 agent of the digest (depth 2 reports sit inside their parent's section). Then resume_file.py log "<path>" stop.

6. Send report-request.md unchanged, one SendMessage per agent, all in one turn, to every depth 1 agent, running or finished, except a finished one whose last words already are an AGENT REPORT. Message a depth 2 agent yourself only when its parent is finished. A failed send counts as silent.
7. Wait in rounds of sleep 30 (foreground Bash): 10 rounds, or 2 per minute when /stop-all got a bare number as argument. Paste each AGENT REPORT into its section as its notification lands. Stop waiting once all answered. ScheduleWakeup only works inside /loop.
8. ListAgents again. For each agent still running: copy the Pending at end block of its --agent <id> digest into its section, TaskStop it, set Reponse a l'arret : non, reconstruit depuis son journal, and fill the section from the digest.
9. Processes outliving their agent: every [background] command in the digest and every server or watcher in a report. Find each by command line (Get-CimInstance Win32_Process); note matches with PID and command line, list the others as not found. A non-matching process is not yours.
10. Write every remaining item in resume.md in one edit, then stop them: background tasks and workflows with TaskStop, a dynamic /loop with ScheduleWakeup({stop: true}), crons this session created with CronDelete, servers by PID after reading their command line, Chrome ports with the chrome-bidi skill's close action (bidi.ps1 -Action close), step 9 processes with taskkill /PID <pid> /T /F through PowerShell.
11. Complete each main session findings entry by comparing report, digest and git; never trust a report alone. The digest sees Write and Edit only: shell-made changes show in git status and mtimes after the agent's first activity. State gaps plainly and flag untrustworthy results (tests run while another agent edited). Fill To relaunch, Relaunch order, owned files with resume_file.py hash <files>, questions, plan. Report Status maps to Status: as done, partial, not started, else unknown. Set State: STOPPED.
12. resume_check.py "<path>" --agent <id> for every depth 1 id until it prints OK, and ListAgents must show no running subagent. After three failed attempts write what is wrong under ## Open questions and tell the user. Delete the scratch folder.
13. Answer in three or four lines in the user's language: what was stopped, the path, who did not answer. Then wait: nothing restarts before the user speaks.

/resume-all

In order, from the file back to working agents.

1. Load the deferred tools as in rule 1. A path in the arguments wins. Otherwise resume_file.py find (looks in .ia/ and at the root of the repo and of the current folder, then in the journals of every account; REPLACED files are skipped). Several files not resumed: AskUserQuestion, newest first. Only resumed ones: ask whether to resume the newest again. None: AskUserQuestion for the path. Never guess a file.
2. Read the whole file and run resume_check.py on it. State: STOP IN PROGRESS means the stop was cut: run the digest with --session <id from the file> and --config <Claude account from the file> and fill the gaps.
3. Reality check, one round: branch, HEAD, status and diff stat of each repo against the file; resume_file.py hash on every owned file against the recorded prints; ListAgents; free RAM, CPU, disk.
4. Drift is a moved HEAD, another branch, or an owned file whose print changed (unless the change is the agent's own recorded work). On drift, AskUserQuestion before relaunching the agents it touches; the others start meanwhile. An accepted drift goes into the brief as a known difference.
5. Depth 1 agents only, one by one (a parent's brief lists its children and their reports). Status: done or To relaunch: no: skip, write why. File written by this same session (origin session equals $CLAUDE_CODE_SESSION_ID): SendMessage the RESUME-ALL text of relaunch-brief.md to the old id and check as that file says. Otherwise, or if that fails: Agent in the background with the same subagent_type, description <old description> (resumed), prompt filled from relaunch-brief.md. A type that no longer exists: general-purpose.
6. Independent agents start together, four at most at once, the next when a slot frees. after agent <id> means after that agent finished and its result was checked. Two agents owning the same file never run together.
7. Relaunch the marked processes only when an agent or your own part needs them; close them at the end.
8. Right after the launches, edit resume.md: a ## Resumed on <date> section above ## Log (new session id, old id to new id or SendMessage, skipped agents and why, waiting agents, drift found), State: RESUMED on <date>, session <id>, a Log line, then resume_file.py log "<path>" resume. The file stays, marked RESUMED.
9. ListAgents must show every agent started now. Retry a failed launch once, then write it in resume.md and tell the user.
10. Delete the scratch folder, take back your own part from ## Main session, answer in three or four lines in the user's language.