Token lanes: jCodeMunch route without execute, then order with written arguments - #435

Summary

The token-lanes carrier told every non-blind Claude subagent to use jcodemunch route(task, repo?, execute?). In
jcodemunch-mcp 1.108.319, route(execute=true) dispatches {"repo": repo, "query": task}: a symbol search runs on
the whole task text (counter.py L616-630). The measurements agree:

• The coordinator's native probes showed an identifier routed to a state-changing action and a sentence kept whole
  as the template query.
• Another session's Codex-lane gate (Capability gate: promptfoo codex-sdk harness for the Codex jcodemunch, ai-memory and M13 gates, reconciled with native OTel #433) found that route with execute returned neither target symbol in 6 of 6
  runs.

This PR changes both lanes' guidance to leave execute off and write the order arguments:

• The Claude carrier and its handbook mirror.
• The Codex user instructions, the managed block that both the native and the gateway lane read.
• The project Codex template's comments.

The carrier's 4,100-byte bound does not move.

Changes

• adoption/hooks/claude/token-lanes-block.md and docs/token-session-handbook.md L224:
  • Before: jcodemunch route(task, repo?, execute?), menu(query?), order(action, args) on indexed repos; (92 bytes).
  • After: jcodemunch route(task, repo: ".") (no execute) then order(action, own args) on indexed repos; (93 bytes).
  • The block is now 4,095 UTF-8 bytes (it was 4,094; the bound stays 4,100).
  • The handbook's sources line (L226) cites counter.py L584-590 and L616-630, server.py L5535-5580, and the
    probes.
• adoption/hooks/claude/SHA256SUMS: the token-lanes-block.md entry is now 5e69d1f9…, the output of
  sha256sum token-lanes-block.md in that directory. sha256sum -c SHA256SUMS passes.
• adoption/templates/codex.AGENTS.template.md: a marked section,
  <!-- native-agent-stack:jcodemunch jcodemunch-mcp 1.108.319 counter.py L616-630 -->, now closes the managed block
  after the RTK exceptions.
  • It limits itself to checkouts whose .codex/config.toml registers jcodemunch; only
    project.codex.config.template.toml does.
  • It is inert in gateway jobs, which run outside every git repository.
  • Both lanes read it: apply_codex_lane.py agents_block() for the native lane and build_args.py
codex_user_instructions() (L208-223) for the gateway lane.
• adoption/templates/project.codex.config.template.toml: two comment lines at L39-40 give the same rule and cite
  counter.py L616-630. The parsed TOML is unchanged.
• docs/decisions/2026-09-27-token-lanes-subagent-start.md: a dated addendum with these parts:
  • the need, and the Codex-lane facts above;
  • a size table and the rejected alternatives with byte counts;
  • a dated erratum and anti-pattern;
  • evidence classes, the overturn condition, and the PR-H: E2E preregistration for the token-save-practice stack (frozen tasks, metrics, gate, procedure) #381 hash note.
• evidence/artifacts/jcodemunch-route-args-20260927/: probes.json (the coordinator's probe record, byte for byte,
  sha256 b7d415ab…) and a short README.md.
• Tests:
  • tests/test_token_lanes_subagent_start.py: test_injected_jcodemunch_rule_leaves_execute_off checks the new clause
    in the injected block and the handbook mirror, and that execute? is absent. The BUDGET_BYTES comment now reads
    4,095.
  • tests/test_codex_worker_lane.py: test_jcodemunch_section_closes_the_block_after_the_rtk_exceptions checks that
    the section is in agents_block(), its marker line appears once and intact, and it sits after the exceptions and
    before the end marker. template_segments() now ends the exceptions block at the new marker.
  • tests/test_codex_agents.py: the F4 slice ends at the new marker. The three custom agents carry F4 (the RTK text
    and its exceptions) only, per the 2026-09-26 worker-lane decision.
• docs/harness-defaults.md (coordinator, 1fbe0016): an anti-pattern log row, as AGENTS.md asks for each proven
  mistake. The anti-pattern: recommending a tool's execute mode without reading how it builds the call. The rule:
  read the installed version's argument shaping and probe it natively first. The two new tests enforce it.
• docs/token-session-handbook.md "Known upstream limits behind the lanes" (coordinator, 1fbe0016): a jCodeMunch
  bullet with the same rule and the order("search_symbols", ...) form for a known name.

Evidence

All commands ran in the worktree with TMPDIR=/var/tmp/claude-jcm.

• Baseline at c8362c02. Every touched and dependent module passed.
• Failing first. Only the tests had changed:
  • python3 -m unittest tests.test_token_lanes_subagent_start exited 1: Ran 15, FAILED (failures=3). The new
    clause was missing from the injected block and from the handbook mirror, and execute? was present in the
    injected context.
  • python3 -m unittest tests.test_codex_worker_lane exited 1: Ran 59, FAILED (failures=1, skipped=8). The marker
    count was 0 != 1.
• Hash coupling. After the block edit and before the checksum update, tests.test_install_claude_profile exited
  1 (failures=2, errors=2). The installer refused, with "sha256 does not match SHA256SUMS".
• F4 coupling. After the template edit and before the slice fix, the unchanged tests.test_codex_agents exited 1
  (Ran 1, failures=3, one per custom agent).
• Acceptance, after the change. PyYAML 6.0.3 came from the local uv cache on PYTHONPATH, read-only.
  NAS_CODEX_INTEGRATION=1 ran the real codex-cli 0.157.1 against scratch HOME/CODEX_HOME, and
  CONTEXT_MODE_SECURITY_JS pointed at context-mode 1.0.169's build/security.js.

validate.py reported exactly these failures:

• SHA-256 mismatch on 9 registered files. All except SHA256SUMS, whose size is unchanged at 383 bytes, also
  report a byte count mismatch:
  • adoption/hooks/claude/SHA256SUMS
  • adoption/hooks/claude/token-lanes-block.md
  • adoption/templates/codex.AGENTS.template.md
  • adoption/templates/project.codex.config.template.toml
  • docs/decisions/2026-09-27-token-lanes-subagent-start.md
  • docs/token-session-handbook.md
  • tests/test_codex_agents.py
  • tests/test_codex_worker_lane.py
  • tests/test_token_lanes_subagent_start.py
• Not hash-listed, the two new files:
  • evidence/artifacts/jcodemunch-route-args-20260927/README.md
  • evidence/artifacts/jcodemunch-route-args-20260927/probes.json

validate.py reported no private-content findings.

A CI-equivalent full discovery, python3 -m unittest in the same environment, exited 1: Ran 6,704 tests in 836 s,
FAILED (failures=2, errors=1, skipped=680).

• test_catalog_freshness_propose test_general_publication_validator_passes_after_the_run runs validate.py and
  fails on exactly the registration set above. It passes once the hashes are registered.
• Two failures come from a host condition, not this change: this workstation has a /tmp/.git directory, which the
  repository's ancestor-.git checks treat as a repository.
  • test_order_throughput test_cli_refuses_live_base_url_from_env_file_without_network raises SafetyError
credential_file_permissions:worktree; its env file is under TemporaryDirectory(dir="/tmp").
  • test_gpt6_family_tiering_20260926 test_sigterm_records_the_running_call_as_interrupted_without_counting_it is
    refused because "the temporary directory /tmp must lie outside the state directory and every repository".
  • Both fail the same way when run alone. Their tests, blueprints/, tools/ and scripts/ are byte-identical
    between c8362c02 and this commit (git diff --quiet exits 0), and none of them reads a changed file.

Replay and registration (coordinator). replay2.sh rebased the builder's c3b97794 onto main, first e82e6be7
and then 55fc8d17 after #433 merged, as 9bcab505, with content lines identical. 1fbe0016 adds the two coordinator
lines above. 69d1f8ad updates the two dated "#433 open" passages to its merged record. The last commit, 319af22d,
registers the branch's 12 changed and added files, following the docs/lanes.md hot-file protocol. At 319af22d:

$ python3 scripts/validate.py
{"components": 69, "hashed_files": 7366, "profiles": 4, "receipts": 159, "status": "passed"}
$ TMPDIR=/var/tmp/claude-jcm NAS_CODEX_INTEGRATION=1 uv run --no-project --with pyyaml==6.0.1 python -m unittest \
    tests.test_token_lanes_subagent_start tests.test_install_claude_profile tests.test_codex_worker_lane \
    tests.test_codex_agents tests.test_render_config tests.test_adoption_docs_consistency tests.test_token_e2e_preregistration
Ran 228 tests in 72.604s
OK (skipped=1)
$ wc -c < adoption/hooks/claude/token-lanes-block.md
4095

Evidence classes:

• Upstream source read at a pinned commit, with the installed package byte-identical to it.
• The coordinator's native probes, one call each.
• Another session's workstation smoke (Capability gate: promptfoo codex-sdk harness for the Codex jcodemunch, ai-memory and M13 gates, reconciled with native OTel #433).
• Structural tests, run failing first and then passing.

This PR claims no subagent or Codex worker run with the new text and no token saving. It reinstalls nothing on the
host: install_claude_profile.py and apply_codex_lane.py were not run against the host.

SOTA sources

• jcodemunch-mcp 1.108.319, upstream commit 8f7b34abe16fb459e0bf1c04747d584216dfe32e, whose pyproject.toml reads
  version = "1.108.319". The installed counter.py (sha256 d56ba52f…) and server.py (sha256 9b6335f2…)
  are byte-identical to that commit.
  • counter.py L584-590:
    _QUERY_ARG maps search_symbols and search_text to query.
  • counter.py L616-630:
    shape_execute_args returns {"repo": repo, qarg: task}.
  • server.py L5535-5580:
    _handle_route uses that shape for args_template (L5549) and for execution (L5562). The route schema's execute
    is at L460.
  • MCP server instructions,
    server.py L340-344 and L360-368:
    "route: start here: the task in plain words, back an action to run." and "order: dispatch any action by name:
    order(action, args). Read-only by default."
• The coordinator's probes: evidence/artifacts/jcodemunch-route-args-20260927/probes.json, native jcodemunch MCP
  calls on 2026-09-27 at about 20:22Z.
• The peer gate, Capability gate: promptfoo codex-sdk harness for the Codex jcodemunch, ai-memory and M13 gates, reconciled with native OTel #433 (tools/capability-gate, merged on 2026-09-27 as 55fc8d17), is that session's measurement, with promptfoo
  0.123.1 openai:codex-sdk:
  • First smoke, 2026-09-27T20:09:50Z: route with the sentence and execute: true returned neither target symbol in
    6 of 6 runs.
  • Gate case: order search_symbols with the identifier passed 6 of 6, with exactly one completed call each
    (2026-09-27T20:16:19Z).
  • Under its repaired scoring, the gate rows passed again, 6 of 6, each with exactly one completed order search_symbols
    that returned the signature. Capability gate: promptfoo codex-sdk harness for the Codex jcodemunch, ai-memory and M13 gates, reconciled with native OTel #433 classes all these runs as workstation smoke, not receipts, and lists the pre-repair
    smokes (20:09Z-20:31Z) as superseded by the repaired-scoring rows.

Independent review

A cross-family review followed. GPT-6 Astra at max ran through the packaged landscape-sweep codex_call.sh lane,
read-only, with live web search, at 4b8c18ff. It returned agree with no findings. Its eight checks:

• the pinned upstream facts, with the installed counter.py and server.py matching the pinned blobs;
• in-memory handler probes: order keeps repo, query, kind and max_results, route without execute
  dispatches nothing, and with execute the sentence becomes the whole search query;
• the call forms and repo: "." resolution;
• v1.108.319 is still the latest release, and main at 1867c5d9 leaves the argument shaping unchanged;
• lane scope and delivery;
• in-memory literal-predicate probes of the changed tests, with the carrier at exactly 4,095 bytes;
• the registration of all 13 files.

Its residuals:

• The tests enforce wording and placement, not model compliance.
• It did not replay Capability gate: promptfoo codex-sdk harness for the Codex jcodemunch, ai-memory and M13 gates, reconciled with native OTel #433's measurements; that PR marks its early smokes as superseded.

The coordinator then checked #433's merged body and updated the two dated passages (69d1f8ad). The later commits
change only that text and the registration; the review's conclusions do not depend on them.

Residuals

• Gateway lane hash. The adoption/templates/codex.AGENTS.template.md hash changes, and the gateway lane's staged
  agents_sha256 follows it when that lane restages after the merge.
• Custom agents. The three custom agents in examples/codex-native/agents/ carry F4 only and do not get the
  jCodeMunch section. Adding it to them is a separate change. It is still unverified whether a spawned role also
  reads ~/.codex/AGENTS.md.
• Host rollout. Host installation is a separate step: install_claude_profile.py --only guard for the carrier,
  and apply_codex_lane.py for the Codex block. PR-H: E2E preregistration for the token-save-practice stack (frozen tasks, metrics, gate, procedure) #381 records the installed carrier hash at execution.
• Measurement. No child or worker run with the new wording is claimed. Re-measure with Capability gate: promptfoo codex-sdk harness for the Codex jcodemunch, ai-memory and M13 gates, reconciled with native OTel #433's gate once its
  receipts land.

Recommended lane label: lane:foundation.

🤖 Generated with Claude Code