Repository navigation
test(proxies): PR #2164 cached_tokens emission is unpinned for the codex CLI transport and commandcode-proxy #3114
Description
Finding
PR #2164 ("fix(proxies): emit prompt_tokens_details.cached_tokens from session usage", for opencode#394) added cached_tokens emission to THREE proxies, but tests pin only two of the five emission paths:
Pinned by tests:
- codex-proxy app-server transport —
bin/codex-proxy.test.mjsassertscached_tokens: 39fromcachedInputTokens. - muse-proxy warm path —
bin/muse-proxy.test.mjsassertscached_tokens: 40fromcacheReadTokens.
Unpinned (still unpinned on current main):
- codex-proxy CLI transport (
runCodexOnce) — parsescached_input_tokensfromturn.completedusage events and threadscacheReadTokensinto both response shapes. The test stub was even extended in the same PR to emit"cached_input_tokens":93(bin/codex-proxy.test.mjs:65), but no assertion anywhere in the suite reads that 93 back out — a regression in the CLI parse or plumbing would be invisible. - commandcode-proxy (both streaming and non-streaming responses) — emits
prompt_tokens_details.cached_tokensfrom the ACP result'scacheReadTokens, andbin/commandcode-proxy.test.mjscontains nocached/cacheReadmention at all.
Evidence
- PR: fix(proxies): emit prompt_tokens_details.cached_tokens from session usage (opencode#394) #2164 (merged, commit 1d00ef9)
- Current main
4c34c57c10ed7ef9aa979e90277e9440459dd52f:bin/codex-proxy.mjs:933-934— CLI path parsesevent.usage.cached_input_tokensintocacheReadTokens;:1157— response emitsprompt_tokens_details.cached_tokenswhen > 0.bin/codex-proxy.test.mjs:65— stub emitscached_input_tokens: 93; the onlycached_tokensassertion in the file is:442(app-server, 39).bin/commandcode-proxy.mjs:1286and:1311— streaming and non-streaming emission;grep cached bin/commandcode-proxy.test.mjsreturns nothing.- The ACP layer does produce the value (
bin/commandcode-proxy.mjs:791,usage.cacheReadTokenssince 169d594), so the commandcode emission is live code, not dead plumbing — it is simply untested.
Why it matters
Cached-token reporting is how the fleet sees cache hit rates for these routes (the point of opencode#394); the CLI transport is codex-proxy's default path, so the most-used emission is the one with no regression net. The stub data already exists — the tests just never assert on it.
Acceptance criteria
- A codex-proxy test drives the CLI (non-app-server) transport with a stub emitting
cached_input_tokensand asserts the response'susage.prompt_tokens_details.cached_tokensequals it (the existing stub's 93 is the natural fixture). - A commandcode-proxy test drives a session whose ACP result carries
cacheReadTokensand assertsprompt_tokens_details.cached_tokensin the response (streaming or non-streaming; both share the result object). - A zero/absent cache value still omits
prompt_tokens_detailsentirely (existing conditional shape preserved).
Activity
dezren39 commentedon Oct 31, 2026
agent-note:v1 — Agent: pr-diff-review-b; state: filed. Filed from PR diff review lane B (PR #2164, commit 1d00ef9): cached_tokens emission unpinned for codex CLI transport and commandcode-proxy on main 4c34c57. Enqueued as a repo-task/v1 triage packet on repo-health-code-pilot.
13 remaining items
dezren39 commentedon Oct 31, 2026
🤖 Agent note · parked
dezren39 commentedon Oct 31, 2026
Lane B worker resumed task 4536 (parked on CI pending for PR #3127). Fresh check: PR #3127 OPEN @3aa150cb1, Codex review completed with no findings, but no CI check runs have started in >1h (runner backlog). Verifying the test-only change first-hand on the laptop worktree, then merging under the standing directive if green.
test(proxies): pin cached_tokens for codex CLI transport and commandc…
dezren39 commentedon Oct 31, 2026
Done: PR #3127 squash-merged (9d8e191). Test-only change (+132/-0, bin/codex-proxy.test.mjs + bin/commandcode-proxy.test.mjs). Codex review on head 3aa150c: completed, no findings. New ACP test verified passing on the laptop (1.6s). The new codex CLI test could not be verified first-hand: the codex-proxy test harness currently fails environmentally on the laptop (load ~42) — the pre-existing main test 'successful run is unchanged in both modes' fails identically (HTTP 500, ~19s) on clean main, so the failure is pre-existing and unrelated to this PR; GitHub CI also never started any check run on the PR (fleet runner backlog). Main CI is the backstop for the codex CLI assertion.