feat(codex): surface cached/reasoning token breakdown in Challenge and Consult modes

[nova28]

Why (in your own words)

codex exec --json's turn.completed event already reports a five-field usage
breakdown — input_tokens, cached_input_tokens, cache_write_input_tokens,
output_tokens, reasoning_output_tokens — but the Challenge and Consult mode
parsers in /codex only read input_tokens + output_tokens into a single
tokens used: N line and silently drop the other three. Anything downstream that
wants to actually price a codex call (cache hit rate, reasoning spend vs response
spend) has nothing to read. This adds a second, machine-parseable line
(CODEX_USAGE: input=... cached_input=... cache_write=... output=... reasoning_output=...)
printed right alongside the existing one, so the human-readable output is
byte-for-byte unchanged and any downstream tooling can key off the new line.

Live evidence

$ codex exec "say hi in one word" -s read-only --skip-git-repo-check --json < /dev/null
{"type":"thread.started","thread_id":"01a0249b-250c-7e13-a32d-82b9746dd022"}
{"type":"turn.started"}
{"type":"item.completed","item":{"id":"item_0","type":"error","message":"Skill descriptions were shortened..."}}
{"type":"item.completed","item":{"id":"item_1","type":"agent_message","text":"Hi"}}
{"type":"turn.completed","usage":{"input_tokens":21294,"cached_input_tokens":6912,"cache_write_input_tokens":0,"output_tokens":5,"reasoning_output_tokens":0}}

That raw capture piped through the exact parser code now in codex/SKILL.md.tmpl
(Challenge-mode block):

$ cat codex-verify-out.jsonl | python3 -u -c "<the new parser code>"
Hi

tokens used: 21299
CODEX_USAGE: input=21294 cached_input=6912 cache_write=0 output=5 reasoning_output=0

tokens used: 21299 is unchanged from before this PR (21294 + 5). CODEX_USAGE is
the new line, sourced from fields that were already in the JSON object and simply
weren't being read.

Note on obj.model: while investigating this I checked test/helpers/providers/gpt.ts,
which has a defensive if (obj.model) modelUsed = obj.model; on the same
turn.completed event. I could not find a model field anywhere in a real capture
(checked every event type: thread.started, turn.started, item.completed,
turn.completed) on codex-cli 0.146.0. Left that alone — not touching an existing
assumption I can't independently confirm, and out of scope for this PR either way.

Scope

• Changed: codex/SKILL.md.tmpl — Challenge mode and Consult mode (new-session)
  JSONL parsers both print a CODEX_USAGE: line on turn.completed, in addition to
  the existing tokens used: N line. Consult mode's resumed-session path shares the
  same parser by reference ("same python streaming parser as above" in the template)
  so it picks this up without a separate edit. Documented in the ## Cost Estimation
  section. Regenerated codex/SKILL.md via bun run gen:skill-docs (verified with
  --dry-run afterward — all files report FRESH).
• Verified live by: a real codex exec --json call (codex-cli 0.146.0, shown
  above), the extracted parser code run standalone against that capture (shown
  above), bun test test/skill-validation.test.ts (331 pass / 0 fail — includes the
  existing expect(content).toContain('tokens used') assertion, unaffected since that
  string is untouched), and the full codex test surface: codex-e2e*.test.ts,
  codex-hardening.test.ts, codex-model-probe.test.ts,
  codex-resume-flag-semantics.test.ts, codex-under-codex-detection.test.ts,
  codex-web-search-flag.test.ts, codex-generation-model.test.ts (76 pass / 20
  skip / 0 fail, 96 tests across 10 files).
• Did NOT test: Review mode is unaffected by design (native codex review never
  uses --json, so it has no turn.completed event to read this from at all — noted
  explicitly in the doc addition). Did not add a new automated test asserting the
  CODEX_USAGE: line's exact format, since the existing test suite validates rendered
  skill content/structure rather than executing the embedded Python against live
  JSONL — happy to add one if a maintainer wants a specific shape/location for that.

Liveness proof (required)

Checklist

• Liveness screenshot attached: GSTACK PR typed live into a real surface (not edited onto the image) — pending, see note above

- [x] This is not a generated-file-only diff (I edited the source/template and regenerated) - [x] No ETHOS.md edits, and no changes to voice / founder perspective / YC references - [x] New public command / external service / host adapter has an accepted issue linked (or N/A) — N/A, no new command/service/adapter - [ ] Linked issue or reproduction: #

[trunk-io]

Merging to main in this repository is managed by Trunk.

• To merge this pull request, check the box to the left or comment /trunk merge below.

After your PR is submitted to the merge queue, this comment will be automatically updated with its status. If the PR fails, failure details will also be posted here