Token lanes carrier: fetch routing, intent, executor-backed cwd, loadable lane ids and wrapper exemption

[seathatflowsinourveins]

Summary

Closes the measured gaps in the SubagentStart token-lanes carrier, the block injected into every non-blind subagent.
Compared with main, the carrier now:

• routes page fetches through ctx_fetch_and_index and ctx_search, and does not use WebFetch for evidence;
• adds ctx_fetch_and_index to the one-call ToolSearch bootstrap, with exact loadable tool ids for each lane a task needs;
• runs output over ~5 KB through ctx_execute/ctx_batch_execute with intent, with specific ctx_search queries (limit <=3);
• states the working-directory rule the context-mode 1.0.169 executor implements: pass cwd for every language (src/executor.ts:295-306, Prepare CLOCK-R2 Windows Time changes and boot persistence checks #788);
• exempts agents told to return output unmodified from the output-routing and footer rules.
  The block grows from 3,159 bytes on main to 4,051 bytes, under a bound raised from 3,500 to 4,100 bytes. The byte itemization under Changes is relative to the intermediate build (3,490 bytes).

Changes

• adoption/hooks/claude/token-lanes-block.md: explicit cwd for every language
  with Rust's absolute-path exception; exact task-needed ToolSearch ids and
  tool-grant boundaries; specific ctx_search queries with limit <=3; full
  wrapper exemption and end-of-return accounting for other agents.
• adoption/hooks/claude/SHA256SUMS: refresh the installer digest. The carrier
  measures 4,051 UTF-8 bytes under a 4,100-byte bound, up 561 bytes:
  316 bootstrap, 122 cwd/Rust, 54 search, 69 wrapper/footer.
• docs/token-session-handbook.md: mirror every carrier line and cite its
  sources; record the dated Claude Code 2.1.283 in-process teammate
  HTTP-transport boundary without prescribing a workaround.
• docs/decisions/2026-09-27-token-lanes-subagent-start.md: date repair errata,
  correct the false schema-based correction, justify the budget and require
  per-child tool exposure for future routing counts. The original decision
  before the measured-gap addendum is unchanged.
• evidence/artifacts/token-lanes-subagent-start-20260927/measured-gaps.md:
  mark 0/15 as M4-style outside PR-H: E2E preregistration for the token-save-practice stack (frozen tasks, metrics, gate, procedure) #381's population, link the preregistration,
  correct counter provenance and separately attribute later matrix/smoke-3
  observations. The two counter .txt records added by the build, original
  counts and historical README/receipt retain their bytes.
• tests/test_token_lanes_subagent_start.py: failing-first cwd, wrapper,
  bootstrap and search controls; preserve UTF-8 budget, blind and fail-open
  coverage. Exact additionalContext delivery covers all seven shipped
  non-blind roles plus general-purpose, workflow, teammate and unknown roles.

Hook executable, installer implementation and installer tests are unchanged.
The carrier grants no tools and changes no client configuration.

Evidence

• Local integration, repair: after export TMPDIR=/var/tmp/claude-lanes,
  python3 -m unittest tests.test_token_lanes_subagent_start tests.test_install_claude_profile
  returned exit 0: Ran 64 tests in 1.023s; OK (skipped=3). All 14 carrier
  tests pass; three existing optional PyYAML checks are skipped. The sandbox
  refused creation of the requested directory as read-only; Python used
  /tmp. These covering modules passed there, not a full-suite run.
• Failing-first repair controls: cwd failed 5 assertions/subtests, wrapper
  scope/footer 3, bootstrap 1, search 1; each then passed. The old byte bound
  rejected 4,051 bytes; the UTF-8 control now rejects 4,120 against 4,100.
  The checksum control failed until its digest was refreshed. These tests
  verify delivered instructions, not model compliance.
• Historical builder integration: original web routing failed once,
  intent/cwd seven times, wrapper once; covering modules then reported 62 tests
  OK, skipped=3. The old cwd assertions encoded the wrong rule and are replaced;
  that earlier green run did not verify runtime cwd.
• Coordinator harness integration, supplied review facts: validate.sh
  FAILS=0. The contended full suite ended with 205 failures and 119 errors.
  Unit-only failing modules rerun alone with TMPDIR=/var/tmp: 1,143 tests
  OK (skipped=66) and an empty rerun failset. The base failed only
  tests.test_pre_commit_gate, absent from the unit failset. The reviewer
  attributed no regression to this unit. These are historical build results,
  not a repair validation/full-suite run.
• Local structural checks: git diff --check exit 0; all carrier lines
  match the handbook; documentation links resolve. Carrier digest:
  2a9bab79fea04ec2c626973428e7db73fad6ace1267e34c2a9553f586e1f8b76.
  Added-text scan: zero personal paths, emails, UUIDs and bearer values.
  Historical receipts, both counters and this repair's evidence manifest are
  unchanged. Read-only Python/JavaScript ctx_execute probes with explicit cwd
  saw the worktree and carrier, while script files were outside it; persistence
  is source-backed, with no write probe.
• Local measurement, coordinator-reported: smoke-2 0/15 is outside M4's
  population, with ctx_fetch_and_index exposure unknown. M3 oversized-byte
  shares remain 61.4%, 70.7%, 74.7%; ctx_execute M5-style counts 40/119, 78/214,
  16/34; wrapper copy check 2/2. Later peer smoke 3 reported ctx_search 11/20
  and ctx_execute 27/52 over 5 KB; it is not an after-measurement of this repair.
  The matrix describes exposure on 2.1.283, not general client guarantees.
• Synthetic fixtures: existing subprocess and temporary installer fixtures
  remain local integration evidence; no native transcripts were synthesized.
• Unchanged upstream tests: none run. Source/schema/release-note review is
  distinct from upstream runtime acceptance.
• Live provider execution: none by this repair builder. No after-repair
  native workflow, installation, provider usage or savings claim.

SOTA sources

• mksglu/context-mode v1.0.169 at 589d8214, executor.ts L295-312,
  Rust L290-292,
  handler L1822,
  routing.mjs L939-941:
  operative cwd (Prepare CLOCK-R2 Windows Time changes and boot persistence checks #788), correcting stale schema text. Installed source was
  checked first; release notes were read with
  gh api repos/mksglu/context-mode/releases/tags/v1.0.169.
• Same pin: fetch L3423-3478,
  intent L1733-1740,
  derived answers L1675-1690,
  threshold L1979-1980,
  ctx_search default 3, L88-94.
• Claude WebFetch,
  MCP tool search
  and SubagentStart.
  Installed CLI returned 2.1.283. Lane ids use the pinned contracts cited in
  the handbook and coordinator ToolSearch matrix observations.
• native-agent-stack 5f3a7c21, sweep.js L83-87:
  complete stdout unmodified, source of the routing/footer exemption.
• Repository evidence/artifacts/token-adoption-e2e-20260926/preregistration.json,
  thresholds.M4: stack-researcher/Codex B, routed rate >=0.9. Coordinator
  units/w3/lanes-facts.md, units/w4/e2e-facts.md, the independent review
  and retained counters provide attributed local measurements.
• Existing tools/adoption/install_claude_profile.py and tests supply the
  checksum/install path; docs/lanes.md supplies the registration-only lane
  rule. No new runtime or installation mechanism is introduced.

Review dispositions

All six findings are accepted; none is disputed.

1. Cwd blocker: executor-backed cwd, persistence and Rust guidance; refreshed
   checksum and measured budget.
2. False erratum: dated correction affirms the original bound-to-main-checkout
   warning for non-shell child calls without cwd.
3. Handoff accuracy: corrected shell-only claims, embedded red/green and
   harness results, accurate registration status. Corrected commit-build.txt
   and supplied commit-repair.txt; no Git history rewritten.
4. Wrapper scope: explicit routing/footer exemption, restored footer
   position, sweep.js citation and tests.
5. M4 scope/access: 0/15 labelled outside PR-H: E2E preregistration for the token-save-practice stack (frozen tasks, metrics, gate, procedure) #381; next remeasure records each
   child's exposure to both fetch tools. Only confirmed ctx_fetch_and_index
   exposure contributes to the routing denominator; absent/unknown is separate.
6. Counter provenance: retained-copy digest identified; tables predate the
   later ctxSrch edit; both counter copies preserved.

A1-A4 are incorporated: exact lane ids, specific queries with limit <=3, dated
teammate boundary, and unchanged non-blind context injection.

Residuals and not done

• Coordinator acceptance at the replayed head (local integration, this host): validate.sh FAILS=0 after the replay onto main. python3 -m unittest tests.test_token_lanes_subagent_start tests.test_install_claude_profile tests.test_adaptive_paper_metrics tests.test_secret_path_guard ran 126 tests OK (skipped=3). The replay resolved adoption/hooks/claude/SHA256SUMS by recomputing every entry from the files; sha256sum -c passes. The re-registration is limited to this branch's own files.

• Cross-unit residual: once the measure unit (PR-A) lands, the decision addendum should name this unit's figures after their counters (for example "fetch-tool share", "results over 5 KB") and state that PR-H: E2E preregistration for the token-save-practice stack (frozen tasks, metrics, gate, procedure) #381's overturn gate uses PR-A's published M3/M4/M5 definitions, not these scratch counts.

• The branch's first commit message still says "shell-only cwd" and "unchanged 3,500-byte budget". The squash-merge subject and body use this PR's text instead.

• Re-measure routing/containment on the next sweep and coordinator waves with
  per-child exposure and population boundaries; smoke 2 cannot retroactively
  supply an access-qualified compliance baseline.

• Private transcripts and original counter argv/timing/exit records and stdout
  were not supplied; retained methods and tables do not independently reproduce
  historical counts.

• No after-repair model run or live installation. Three PyYAML skips and the
  requested-TMPDIR sandbox restriction remain recorded.

• Coordinator owns committing, evidence registration, generated reports,
  validate.sh and full-suite acceptance. Builder left manifests/evidence.json
  unchanged and ran no prohibited Git/report-writing operations.

• The existing plugin-scoped blind-role caveat is unchanged.

Recommended lane label: lane:foundation. The coordinator's reviewed final
the branch's last commit changes manifests/evidence.json only by re-registering this branch's own files;
docs/lanes.md L147-148 keeps that work in foundation. This repair checkout starts
at 67f4a88f and leaves the next registration to the harness.

Token tools: context-mode processing and indexed source fetches/searches; RTK
for commands and rtk proxy for exact blobs. No compression or savings claim.

🤖 Generated with Claude Code