Token stack inside native Codex sessions: 11/15 tools with passing checks, four real limits, and a correction to #296's QMD row

[seathatflowsinourveins]

Scope

• What this PR changes:
  • adds evidence/artifacts/token-e2e-codex-20260926/ (README and receipt.json), the Codex side of Token stack inside Ultracode subagents: 16/16 tools E2E with lifetime counters, exact comparisons and five new host receipts #296;
  • adds a dated correction to Token stack inside Ultracode subagents: 16/16 tools E2E with lifetime counters, exact comparisons and five new host receipts #296's README (its QMD row);
  • adds a Codex paragraph to docs/token-efficiency-stack.md.
• Base commit: 19092db0
• Lane: lane:foundation
• Owned paths touched: the new artifact folder, evidence/artifacts/token-e2e-ultracode-20260925/README.md, docs/token-efficiency-stack.md, and their registrations.

SOTA sources

• Codex CLI 0.155.1 (https://github.com/openai/codex): codex exec --json --output-schema -o (native help), and turn.completed.usage events for the returned usage.
• Tools, at the versions installed on the host:
  • rtk-ai/rtk 0.50.0, instruction mode through rtk init --codex;
  • mksglu/context-mode 1.0.169, the plugin;
  • chopratejas/headroom 0.37.0 with HEADROOM_OFFLINE=1 (Headroom: run the MCP server with its usage beacon off (HEADROOM_OFFLINE=1, DO_NOT_TRACK=1) #304);
  • oraios/serena at the pinned commit;
  • giancarloerra/SocratiCode 1.14.0;
  • DeusData/codebase-memory-mcp v0.11.0;
  • tobi/qmd 2.8.3;
  • ast-grep 0.45.3, MarkItDown 0.1.8, TOON 4.1.1, Repomix 1.18.1, Context Hub 0.1.4, agentsview v0.43.0;
  • akitaonrails/ai-memory 2.4.1.
• Exact counts: gpt-tokenizer 3.4.0, o200k_base, via tools/token-report/token_manifest.py compare.

Evidence-class table

Claim	Evidence class	Where
11 of 15 tools worked inside native Codex sessions, and each answer passed a deterministic check against the baseline	native_proven on this host	receipt.json tools[]
The four failures are tool or wiring limits: jCodeMunch is project-scoped (#240); QMD's index excludes adoption/; Repomix --compress dropped a declaration; Context Hub has a content gap	native_proven (retained sessions and reruns)	tools[].verdict_note
RTK +17,782 saved for this run's worktree only (41 commands)	upstream estimate, attributable	counters.rtk_gain_codex_worktree_only
9 exact comparisons, 28.2% to 99.8% removed	local_integration (exact counts; not provider savings)	tools[].exact_comparison
Usage: 7,085,015 input (6,253,312 cached), 164,906 output, 91,997 reasoning	Codex-returned	codex_usage_total
#296's QMD row retrieved the wrong document for its question	source_review	corrections_to_296

The Headroom and Context Mode (Codex) deltas are host-wide: the after snapshot came about 12 hours after the runs. They are recorded but not credited to this run.

Local commands run

$ uv run … scripts/validate.py; evidence_manifest.py --check; build_ecosystem.py --check   # passed (6584 hashed files)
$ uv run … -m unittest <27 modules that read token-efficiency-stack or evidence/artifacts>   # Ran 1589, OK (skipped=26)
$ uv run … -m unittest tests.test_landscape_sweep_harness   # Ran 71, OK alone, both on this branch and on clean main.
  # It fails only when run in one process with the other modules, which is a test-ordering interaction that also exists on main.
$ gitleaks-guarded dir evidence/artifacts/token-e2e-codex-20260926   # no leaks found

Checklist

• No workflow changes. [x] No secrets. The receipt was scanned: 0 home paths, 0 scratch paths, 0 ids. [x] No paid surface. [x] Peer files preserved.
• Independent review: a cross-family Codex review is requested next.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[seathatflowsinourveins]

Cross-family review, Codex gpt-6-astra, read-only, effort high. needs_changes → fixed. Every README and docs number matches receipt.json, the usage totals sum correctly, and the evidence classes, the four failure descriptions and the #296 QMD correction are all supported. It found three receipt defects:

• a live-clone path;
• a session-derived codebase-memory project id and the SocratiCode collection id;
• a QMD document count stated as 121 while attempt 2 read 122.

All three are fixed in the latest commit. The receipt now scans clean (0 matches for those paths and ids), and validate and the evidence manifest pass.