docs(examples): kill -9 → resume without re-spending (the flagship)

[prashar32]

The flagship demo

The crash-resumable moat, made into a recordable, reproducible demo (examples/kill-9-resume/). A checkpointing agent runs under a governed run; the daemon is kill -9’d mid-run; on restart it reloads the run with the budget it had already spent, and re-running the agent (via resume_run, #61) picks up from its last checkpoint and finishes — without redoing the paid steps.

The proof is one number: across the crash the governor’s loop counter ends at 10, not 15.

./demo.sh — the money shot, in one command

Scripts it deterministically: start daemon → 5 of 10 steps → kill -9 the daemon → restart (resumed runs from store count=1) → resume from cursor 5 → finish → runs list shows LOOPS=10.

↻ RESUMING run 4e25ad71-…
  the governor already counts 5 spent steps — resuming at cursor 5, not redoing them.
  step  6/10 … step 10/10 done
✅ governor loop counter = 10 — exactly one per step of work … neither redone nor re-paid.

Contents

• agent.py — the resumable agent (create-or-resume via a run-id file; checkpoints each step; prints a clear “the daemon is gone — restart and re-run to resume” message when it can’t reach the daemon).
• demo.sh — the reproducible orchestrator (uses the binary so it can kill -9 the process).
• README.md — one-command run, the by-hand flow for a live recording, how it works, and the “$ not double-counting” note (Reload restores spent dollars/tokens too — shown via the proxy).
• Linked prominently from the main README (the checkpoints row + an examples pointer).

Verification

Ran demo.sh end-to-end against the built binary + the SDK: 5 steps → kill -9 → restart → resume → finish, LOOPS=10 (not 15). Also verified the daemon-crash-mid-run path prints the recovery message. py_compile + bash -n clean.

Key-free. No model call, no API key — the loop counter is the proof.