Gate A U2 and U3: Claude and Codex kernel measures for the #381 token E2E (call states, M15, ledgers, D7, joins)

[seathatflowsinourveins]

Scope

• What this PR changes, in one or two sentences: adds the Claude-side kernel measures of the PR-H: E2E preregistration for the token-save-practice stack (frozen tasks, metrics, gate, procedure) #381 token E2E (U2: injected context from every hook event, loaded-not-called per server, context bytes with JSON and tabular detection, incomplete returns, M14 call states with a private per-child call ledger behind --call-ledger, M15 per-server infrastructure-error rates, heredoc-aware shell parts (B8), per-actor usage with advisor iterations (B9), and the rtk-claude part classification) and the Codex-side measures (U3: injected versus inherited marker, grouping by custom-agent role, spawn fork_turns / role / route joins, nested code-mode attribution, the rtkAgent replay with D7, and the Codex half of the call ledger). The two units rewrote rtkParts in the same file, so the merge is resolved once here (one rtkParts with both units' behavior) and verified as one tree; the Codex ledger tests, which need U2's callLedger, now run instead of skipping.
• Base commit: f77612b6 (main after bc's PyJWT relock; the branch's last merge commit brings it in)
• Lane: lane:foundation (Gate A units U2 and U3; owner native-agent-stack-2d)
• Owned paths touched: examples/claude-native/workflows/{child-usage.mjs,test-child-usage.mjs,README.md,SHA256SUMS}, tools/skill-usage/{skill_usage.py,README.md}, tests/{test_token_measurement.py,test_skill_usage.py} (and additive lines in tests/test_child_usage_suite.py where a unit needed them), evidence/artifacts/pra-u2-differential-20260929/ and evidence/artifacts/pra-u3-differential-20260929/ (new), manifests/evidence.json (hot file, last commit)

SOTA sources

• openai/codex rust-v0.157.1 (36650394): codex-rs/protocol/src/{items.rs:372-381 (WebSearchItem has no status), models.rs:1182-1203, 1060-1165, 1938-1949, 2173-2182, protocol.rs:772-779, 1403-1415, 1570, 4393-4402}, codex-rs/rollout/src/policy.rs:56, 94-145 (what a rollout persists in legacy and paginated history), codex-rs/core/src/tools/handlers/multi_agents_v2/spawn.rs:233, 253-263, codex-rs/core/tests/suite/code_mode.rs:721-760, 3436-3752 (nested returns go to the code-mode JavaScript), codex-rs/core/src/tools/context.rs.
• rtk-ai/rtk v0.50.0: src/hooks/decision.rs:61-97, 196-204 and src/hooks/permissions.rs:56-67, 141-175 (rtk hook check --agent claude|codex: the claude checker merges project and home Bash rules, the codex one has no RTK-side rules; the rewrite decision is the same).
• mksglu/context-mode v1.0.169 src/server.ts:825-836, 891-897 (trackResponse prepends the outdated-version notice to every response, successes included) and its tool error shapes.
• Claude Code 2.1.284 and 2.1.285 (hook attachment types read from the binary, docs for hooks and usage monitoring), the Anthropic Messages API usage and advisor iterations documents (retained copies in the U2 evidence sources.json), POSIX.1-2024 XCU 2.7.4 (here-documents) and 2.9.x, bash(1) 5.2 and dash behavior observed on this host.
• Every retained copy and line anchor with its retrieval date: evidence/artifacts/pra-u2-differential-20260929/sources.json and evidence/artifacts/pra-u3-differential-20260929/README.md.

Evidence-class table

Claim	Evidence class	Command / receipt
The merged tree passes both units' suites: the kernel, child-usage, skill-usage, command-position oracle, preregistration, freeze-snapshot, landscape-sweep and saturation-ledger modules	local_integration	python3 -B -m unittest tests.test_token_measurement tests.test_child_usage_suite tests.test_skill_usage tests.test_command_position_oracle tests.test_token_e2e_preregistration tests.test_freeze_snapshot tests.test_landscape_sweep_harness tests.test_saturation_ledger → Ran 705 tests in 423 s, OK (skipped=4): the four skips are environment ones (CONTEXT_MODE_SECURITY_JS unset, no bash 3.2 binary, shellcheck not installed, jsonschema not installed); the rtk-gated and parser-gated tests ran
The CodexCallLedger tests, seven of them against the real callLedger (five were skipped in U3 alone), run and pass	local_integration	python3 -B -m unittest tests.test_skill_usage.CodexCallLedger → Ran 16 tests, OK (0 skipped)
The node suites pass	local_integration	node test-child-usage.mjs 216/216, test-contract-mutations.mjs 74/74, test-envelope.mjs 254/254, test-usage-receipts.mjs 18/18, test-codex-envelope.mjs 0 FAIL; node check-syntax.mjs … and node --check child-usage.mjs exit 0
The merged kernel gives the U2-only kernel's Claude output on real data: no path differs, none is added or removed	local_integration (private host transcripts reduced to counts)	lanes sweep of the frozen window 2026-09-26T00:00Z to 18:37Z: 635,560 paths, 635,560 equal, 0 changed, 0 added, 0 removed, 0 unexpected; the same with --rtk-check over 2026-09-26T00:00Z to 05:00Z: 182,796 paths, all equal; run mode over 20 real Workflow runs: 226,429 paths, all equal, exit codes equal (measured again on the repaired head 7f9263e1) (evidence/artifacts/pra-u2-differential-20260929/differential.mjs with an empty allow-list)
U2's own differentials at its final kernel (frozen W_AA copy, 232 workflow runs, 12 --rtk-check chunks): 0 unexpected paths, invariants hold, the call ledger is byte-identical to a run without it	local_integration	U2 evidence README and evidence/artifacts/pra-u2-differential-20260929/ (recorded before this merge, at kernel c1d9ad16)
U3's host differential: every difference between its base and head reports fits the design; the Claude output is byte-identical at the default agent	local_integration	evidence/artifacts/pra-u3-differential-20260929/README.md, claude-byte-identity.py
The merged tool's Codex lane report over the host store (window 2026-01-01 to 2026-09-29, --rtk-check) against U3's alone: 0 leaves removed, 122,680 added (U2's call_states, m15, eligible_call_states and carrier fields on Codex actors), 441 changed values, all under rtk_parts (status, unknown_calls, ineligible_parts, proxy_parts, d7), which is B8's call-level unknown rule (U2) reaching the Codex replay	local_integration (private host rollouts reduced to leaf-path counts)	skill_usage.py --lanes --codex-root <store> --since 2026-01-01T00:00:00Z --until 2026-09-29T00:00:00Z --now 2026-09-29T00:00:00Z --json --rtk-check at both trees, compared by leaf path
Independent reviews: U2 one high (M15 scored calls that did not run as succeeded) and five low, the high fixed and the independent re-verify passed 14 of 14; U3 one medium (the Codex ledger was undisclosed as missing) and five low, the ledger delivered and the re-verify passed 15 with 2 unverified (the ledger fixtures, which this merge now runs, and the outer-JS rule, accepted below)	source_review	the Gate A workflow records of both units
Independent review of the merge itself (Opus evidence-reviewer, read-only, integration only)	source_review	3 medium and 2 low findings, all repaired in this PR (below)
GPT-6 cross-family review	NOT run (native and pooled GPT-6 capacity exhausted); recorded as a missing round	-

Local commands run

$ python3 -B -m unittest <the eight modules above>
Ran 705 tests in 423.125s

OK (skipped=4)
$ python3 -B -m unittest tests.test_skill_usage.CodexCallLedger
Ran 15 tests / OK
$ python3 scripts/validate.py
{"status": "passed", ...}
$ git diff --check e45328d3..HEAD
(no output, exit 0)

Review and process record: the pre-repair independent command re-run agent of both units finished without calling StructuredOutput and counted as failed; each unit's repair stage re-ran its acceptance list and an independent re-verify agent followed. Two U2 commits (1a189bdb, 01af2f62 before rebasing) hold a kernel that does not parse, fixed by the next commits; the PR is squash-merged. Unit commits carry an Opus 5.5 trailer because a builder of that model wrote them.

Integration decisions made here (each with a test that failed first or a differential): (1) one rtkParts with U2's structure and U3's agent parameter and D7 counters, D7's per-call counters held until the call is known readable so an unknown call adds nothing to any part counter, as B8 requires for the other counters; D7's status follows the merged rtk_parts.status; (2) a completed web.search item now carries native_status: completed (U3 open question 1, option (a): its item_completed event is its completion; all 3,650 such rows, 7.4 percent of the host ledger, read as cancelled_or_unfinished before), with a real-kernel test that failed first; (3) the outer-JS rule of tools/skill-usage/README.md ("The outer exec code") is accepted by the PR-A owner, which releases U3's commits 7 and 8: a count of outer HTTP mentions, no reading of the isolate where the client's isolation is known.

Merge review and repair (one review, one repair round; failing-first tests are the commit 05100fe2, the repair 7f9263e1, six mutants of the new guards all caught): (1) the Codex bridge had added unresolved commands to unknown_calls after the kernel finished, forced status incomplete and left unknown_call_share stale, so one unknown call read differently depending on which layer counted it; the kernel now takes measureTranscript option unresolvedBash (non-negative integer, never beyond the Bash calls not already unknown) and counts inside rtkParts, so B8's 5 percent rule, the share and the D7 status come from one place; (2) D7's status had been silently relaxed to B8: decided to keep the harder reading (any unknown call or unreviewed log or find part leaves d7 incomplete, rtk_parts.status keeps B8's tolerance; both READMEs say so and that an amendment may relax it); (3) the aggregate d7 status is the strict vote, independent of how calls are split across actors (a test splits 110 calls over one and two rollouts); (4) a failed hosted web_search_call no longer reads completed; (5) stale README statements on the ledger and native_status. Residuals (bounded review loop: one review, one repair round): (1) the shell splitter still reads $[1<<2] as a here-document operator (shellParts('echo $[1<<2]\nls -la') returns one part; 0 of 37,951 Bash calls in the frozen W_AA files hold the form; fixing it changes the rtk-check differential); (2) six M15 error classes stay unassigned until Amendment 4 (grading reads rate_upper_bound and over_threshold meanwhile, which fixes the high finding), and every_error_classified counting of invalid and indexed exits is an Amendment 4 decision; (3) the plain ( ) subshell here-document case, the 5m/1h split edge for a message straddling since, and stale item-6 docs in scripts/saturation_ledger.py and tools/sota-convergence/landscape-sweep/usage_record.py (outside these paths); (4) the V1 SubAgentActivity shape is pinned by a fixture, unverified against a real V1 rollout, and a legacy-history rollout's McpToolCallEnd, WebSearchEnd and non-completed SubAgentActivity events are not read (this host holds no legacy rollout); (5) fork_turns is Option<String> with deny_unknown_fields upstream, so a literal integer fails the spawn (note for the #381 protocol owner; the join reads it as last-N only); (6) two store lines did not parse in U2's host runs (2 of 5,283 files, cause not examined); (7) CI skips the rtk and parser replay tests (they ran here with rtk 0.50.0 and the tree-sitter install).

Decision record

No new decision record: the choices are recorded in the workflows README ("U2 measurement fields"), tools/skill-usage/README.md and the two evidence directories' READMEs and sources.json.

Host evidence

Not applicable (no files under evidence/hosts/).

Checklist

• New/changed GitHub Actions are pinned to a full commit SHA with a version comment (no workflow changes in this PR).
• New/changed workflows declare top-level permissions: contents: read (no workflow changes).
• No secrets are printed, logged or committed; the reports and the ledger hold no ids or paths (the ledger is a private create-only file outside every work tree, and a test refuses a path inside the checkout).
• No new paid hosting, subscription or billing surface was introduced.
• Peer-owned untracked files and worktrees were preserved.

🤖 Generated with Claude Code

[seathatflowsinourveins]

Codex report on the repaired head 7f9263e1, merged tool against U3 alone (skill_usage.py --lanes --codex-root <store> --since 2026-01-01T00:00:00Z --until 2026-09-29T00:00:00Z --now 2026-09-29T00:00:00Z --json --rtk-check at both trees; 1,436 actors paired by position, same count and order; leaf-path counts only, no ids or text):

• Actors whose rtk_parts fields differ: status 102, unknown_calls 102, ineligible_parts 90, proxy_parts 14, explicit_rtk_on_excluded_or_sensitive 1, d7.wrapped_exceptions 1. No other rtk_parts field differs for any actor.
• unknown_calls moved down for all 102 of them (by 1 to 32 calls): B8's call-level, heredoc-aware splitting (U2) reads calls that U3 alone counted as unknown. None moved up.
• rtk_parts.status flipped incomplete to measured for all 102, and every actor's merged status equals B8 applied to its own counts (0 inconsistent); unknown_call_share is fresh for every actor (0 stale).
• d7.status flipped incomplete to measured for 99 actors, and every merged d7.status equals the strict rule (measured only with 0 unknown calls and 0 unreviewed log or find parts): 0 inconsistent.
• Groups after the merge: workers 3,782 Bash calls, 0 unknown, status measured, d7 incomplete (unreviewed log or find parts); negative_controls 3,150 calls, 0 unknown, measured, d7 incomplete.
• 0 leaves removed against U3 alone; the leaves added are U2's fields on Codex actors (call_states, m15, eligible_call_states and carrier fields).
• Repaired-tree checks: 705 tests OK (4 environment skips), node suites 216/74/254/18 and 0 FAIL, Claude sweep 635,560 of 635,560 paths equal, run mode 226,429 of 226,429 paths equal over 20 real Workflow runs.

[seathatflowsinourveins]

Foundation defaults handoff for native-agent-stack-2d / PR #528, following docs/lanes.md and the 2026-09-29 spend-attribution decision.

The user approved routine Codex gpt-6.1-sol / ultra coordination and gpt-6.1-sol / max primary workers, with Astra/max for consequential architecture, conflicting primary evidence, complex cross-system changes, or failure after one bounded Sol repair. Explicit model choices and the selected Astra researcher/verifier role definitions remain authoritative. Claude remains Opus judgment / native max session effort / Ultracode, saved xhigh fallback, no global effort override, and Fable advisor by default with native consent.

Sources: installed Codex 0.159.2; tagged child defaults and role precedence; Ultra versus Max native dispatch; Claude 2.1.285 and its 2.1.284 Ultracode/effort separation. Native Opus 5.5/max analytical peer review returned conditional approval with observable escalation and bounded decision records; this does not qualify tool integration or establish model superiority.

Please retain ownership of Gate A role caps, selective omitClaudeMd and carrier narrowing. In particular, assess the unconditional Serena carrier loads and the six uncapped roles using the frozen experiment protocol; no new limits were inferred or written by this defaults change. Generic routing changes are new inputs for future trials, not a retroactive modification of dated Astra role qualification. PR #381 is closed preregistration, so this packet is attached to the active owner work here.

PR #528 already contains separate executor, advisor and combined accounting with incomplete-iteration checks. Please carry the accepted implementation into canonical agent-lab first, run its supported .claude/workflows/test-child-usage.mjs, and vendor byte-identically with refreshed hashes. Current upstream lead: agent-lab eced71c572671fc3e381381c068d05b2a7a5bb22; its advisor handling is not equivalent to the completeness checks proposed here. I will consume the owner-managed result rather than create competing edits to these files. No request to expand this foundation work into trading or optional services.