CF artifacts: agent-loop rehydration integration test

[ivarvong]

Summary

End-to-end integration test that proves the artifact-backed agent-loop story: provision a fresh CF artifacts repo + write token via the management API, seed initial state, boot agent 1 by cloning + mounting through VFS, modify the working tree via VFS.write_file, commit + push, drop all in-memory state, then rehydrate agent 2 from a fresh clone and verify session 1's edits persisted.

• Bumps :exgit to pull in Exgit.Workspace + its VFS.Mountable defimpl, so writes flow through the unified %VFS{} surface (no manual blob/tree/commit composition).
• Tagged :integration_network + :cloudflare; self-skips when CF_API_TOKEN / CF_ACCOUNT_ID aren't set, and deletes the provisioned repo on teardown.
• When CF_BENCH_LOG=path is set, each run appends a JSONL record (timestamp, repo name, per-op timings, total) to that path. bench/cf_aggregate.exs <jsonl> reads it and prints min/p50/mean/p90/p95/p99/max per op.

The compile-order workaround in test_helper.exs

Exgit.Workspace's VFS.Mountable defimpl is guarded by Code.ensure_loaded?(VFS.Mountable). Mix always builds deps before the host project's lib, so when :exgit compiles inside the vfs repo, our protocol module doesn't exist yet → the guard fails → no defimpl beam. Downstream consumers don't hit this — mix orders vfs before exgit via the optional edge declared in exgit/mix.exs. The breakage is specific to self-hosting the protocol's own repo. The test_helper compiles the defimpl source in-place once both modules are loaded.

Per-op timings (n=50)

Caveat: measured from my laptop over home internet, not a controlled benchmark. Numbers reflect end-user RTT to Cloudflare's edge from this host; not service SLA, not steady-state from a colocated worker. p99 values with n=50 are very noisy estimators — treat the tail as "this happened once or twice in 50," not as an SLO target.

op	min	p50	mean	p90	p95	p99	max
create_repo	925	1129	1360	2369	2730	3084	3084
mint_token	227	337	360	495	535	891	891
push_seed	690	884	1052	1211	2458	3678	3678
clone_boot	158	209	220	271	290	373	373
push_modified	103	163	163	209	216	266	266
clone_rehydrate	151	229	249	316	362	590	590
delete_repo	676	946	1074	1376	2294	3623	3623
total	3262	4156	4479	5626	7284	9142	9142

(ms)

Read of the data:

• Control plane (create_repo, mint_token, delete_repo) dominates and has heavy right-tails. p50/p99 ratios are 2.7×, 2.6×, 3.8× respectively. create_repo and delete_repo p99 ≈ 3s.
• Data plane (clone, push) is fast and tight. All four ops 100–600ms across the entire run; p50/p99 under 2.6×. Predictable enough to budget per agent turn.
• push_seed has the worst tail — p50 884ms, p99 3.7s. First push to a freshly-created repo competes with whatever post-create work the server is still doing.
• End-to-end: p50 4.2s, p95 7.3s, p99 9.1s. Rehydration is single-digit seconds.
• Steady-state turn cost (just clone_rehydrate + push_modified, no create/delete): p50 ≈ 390ms, p99 ≈ 860ms.

Test plan

• mix test --include integration_network test/integration/cloudflare_artifacts_test.exs (with CF env vars set)
• Default mix test still green (325 tests, 46 properties, 36 doctests, 0 failures)
• 50-run bench (CF_BENCH_LOG=...) + bench/cf_aggregate.exs aggregator
• CI green