PostToolUse(Task) hook — catches a session/usage-limit signal in a subagent result and injects a checkpoint directive so work-in-progress is saved before more quota is spent.
When a subagent (Task tool) completes, the hook scans its result for a session/usage-limit signal (hit your session limit, hit your usage limit, or session limit + resets). On a match it first auto-snapshots uncommitted work (0.1.4), then emits a non-blocking additionalContext instructing the agent to:
- Auto-snapshot (done by the hook) —
git diff HEADis saved to.claude/limit-snapshots/<ts>.patchplus a.status.txtsidecar (branch · HEAD ·git status --porcelain· untracked list). Non-destructive: no commit, no push, no stash, no index mutation — safe in a shared/concurrent working tree (a shared.git/indexmust never be auto-mutated), nogit add -A, no branch switch. Recover a lost tracked change withgit apply <patch>. Untracked files are listed (not captured) in the sidecar. - Report how far it got — per work item, what is committed (with SHAs) vs uncommitted, and where work stopped.
- Commit + push uncommitted work — a limit hit mid-edit loses uncommitted work.
- Write
.claude/RESUME.md— a resume manifest: done items, remaining items, the limit reset time, the next concrete step. - Stop parallel fan-out — switch to sequential foreground work, one committed increment at a time.
No match → silent pass-through.
Parallel subagent fan-outs burn quota fast. When the limit hits, N agents fail at once and uncommitted work is lost. Sequential work with a commit per unit caps the loss at a single unit, and a resume manifest lets a post-reset session pick up cleanly.
bin/_limit_guard.hexa, invoked via hexa run from hooks/hooks.json. No Python, no shell shim.
There is none — no env var, no config file, no exception list. A bypass switch would just be auto-flipped to silence the guard, defeating its purpose. The injected directive is non-blocking guidance; if limit-guard is wrong for your workflow, uninstall the plugin.