v1.56.0.0 Token-reduction Phase B + AUQ paranoid safety net (#1849)

* refactor(plan-ceo-review): carve review body into on-demand section

Carve the largest skill (138,838 B) into a skeleton + one on-demand
section, the documented next Phase B target after /ship (v2_PLAN.md:216).

- sections/review-sections.md(.tmpl): the 11-section deep review, codex/
  outside-voice rules, how-to-ask, Required Outputs, registries, Completion
  Summary, Review Log, REVIEW_DASHBOARD, PLAN_FILE_REVIEW_REPORT, Next Steps,
  docs/designs promotion, Formatting Rules, and the Mode Quick Reference.
- sections/manifest.json: passive registry (CM2), one entry.
- SKILL.md.tmpl: {{SECTION_INDEX}} after the system audit, a single
  {{SECTION:review-sections}} STOP-Read after Step 0 mode selection, and a
  Section self-check. All of Step 0 (the scope/mode conversation) stays in
  the always-loaded skeleton; only EXIT_PLAN_MODE_GATE follows the section.

Measured: always-loaded skeleton 138,838 -> 80,731 B (-42%, ~14.4K tokens
off every invocation). Union (skeleton + section) 139,110 B, behavior held.

Boundary honors Codex P1: nothing review-governing (formatting rules, mode
reference, how-to-ask, required outputs) sits in the skeleton below the
STOP. Housekeeping resolvers ride in the section, matching the ship
precedent (adversarial.md carries LEARNINGS_LOG + GBRAIN_SAVE_RESULTS).

Tests (atomic with the carve — skill-docs.yml gates gen:skill-docs
freshness on every push, so source + regen + tests must land together):
- parity-harness: plan-ceo flipped to sectioned, maxSkeletonBytes 90_000
  (measured 80,731 + headroom); content/minBytes run against the union.
- skill-size-budget: plan-ceo-review added to SECTIONS_EXTRACTED.
- section-manifest-consistency: generalized to discover every carved skill,
  vars computed per-skill-case (Codex P2).
- skill-ceo-section-ordering (new, gate): per-PR static guard — STOP after
  Step 0, review body absent from skeleton, report writer in the section,
  nothing review-governing below the STOP.
- skill-e2e-plan-ceo-review-section-loading (new, periodic): refreshes the
  installed skill first (Codex P1), drives full Step 0, asserts the section
  is Read before the report.
- gen-skill-docs + skill-validation: read the skeleton+sections union for
  carved skills so relocated prose still counts.
- touchfiles: plan-ceo-section-loading registered (periodic).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* chore: bump VERSION + CHANGELOG for plan-ceo-review carve (v1.56.0.0)

MINOR: carves the largest skill into skeleton + on-demand section,
dropping plan-ceo-review's always-loaded cost 42% (138,838 -> 80,731 B,
~14.4K tokens off every invocation). User-facing release notes lead with
the measured token win.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* docs(todos): file P3 follow-up — carve the shared {{PREAMBLE}} reference blocks

Surfaced by /plan-eng-review on the plan-ceo-review carve: per-skill section
carves stay modest because the ~40-50KB shared preamble dominates the
always-loaded surface. A single preamble-reference carve would help every
tier->=2 skill at once. Records the why, the cold-vs-hot split to measure,
and the guards it needs. Not implemented this PR.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* test(auq): Layer 0 — guarantee AUQ format spec is always-loaded

Deterministic, free, per-PR keystone for the token-reduction era. For every
interactive (tier>=2) skill, asserts the full AskUserQuestion decision-brief
format (ELI10/Recommendation/Pros-cons/checks/Net/(recommended)/Stakes/
self-check) lives in the always-loaded SKILL.md skeleton, NOT only in an
on-demand section. Plus a roster guard (a carve can't silently drop the block)
and per-skill rule survival in the skeleton+sections union. 51 cases + a
negative control. Fails the instant a future carve strands AUQ-governing text
where it won't be loaded when a question fires.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* test(auq): SDK capture engine + verbose-vs-carved no-degradation A/B

Adds the reusable SDK $OUT_FILE capture engine (auq-sdk-capture.ts): drives a
skill to its AUQ and captures the verbatim text the model GENERATES, cleanly
(real-PTY mangles plan-mode AUQs via cursor escapes). Pins the skill to an
absolute path with Read/Write-only tools so the agent can't wander to the
global install. gradeAuqRecommendation normalizes a non-"because" connective
before grading so substantive reasons aren't false-flagged (without touching
the pinned shared judge).

The A/B drives the same prompt through the carved 80KB skeleton and the
pre-carve 137KB monolith and fails if carved scores worse. Result: both 7/7
format, substance 5 — proven no degradation, transcript-verified each side read
its own planted SKILL.md. Periodic tier.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* test(auq): consistency — same trigger N runs, stable format + substance

Drives the carved /plan-ceo-review AUQ N=3 times and fails if any format
element appears in one run but not another, or substance craters. Targets the
"fine one run, broken the next" failure class a single snapshot can't see.
Result: 3/3 stable, 7/7 + substance 5 every run. Periodic tier.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* test(auq): behavioral matrix across AUQ-heavy skills

Data-driven test that drives each AUQ-heavy skill (plan-eng/design/devex,
office-hours, cso, spec, design-consultation) to its first AskUserQuestion and
grades it to the plan-ceo bar: 7/7 decision-brief format + recommendation
substance >=4. One case per skill (isolated failures), env-subsettable via
AUQ_MATRIX_ONLY. Browser/design-binary skills are intentionally excluded
(comparison boards, not format-AUQs; Layer 0 covers their spec). All targeted
skills pass 7/7 with substance 4-5. Periodic tier.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* test(codex): live recommendation-substance grade for /codex

Closes the gap where /codex's synthesis recommendation was only checked
statically (template grep) and via fixtures. Drives the real /codex skill over
a flawed diff and grades the emitted "Recommendation: ... because ..." line
with judgeRecommendation (present/commits/has_because/substance>=4). The named
weak spot holds up: substance 5. Periodic tier.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* test(auq): deterministic trigger for format-compliance gate

A bare /plan-ceo-review against a repo whose work is already implemented makes
the model improvise an off-script "what should I review?" scope question that
skips the decision-brief format, which the gate test then times out waiting for.
Hand it a concrete plan to review (FORCING_FLOOR_CEO) so it reaches the real
Step 0 mode-selection AUQ that is the intended format check.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* refactor(office-hours): carve Phase 5+6 into on-demand section

Third Phase B carve (v2_PLAN.md:216, after ship and plan-ceo-review). Moves
Phase 5 (Design Doc templates) + Phase 6 (tiered relationship handoff) — the
session's output + closing tail, only reached after the conversation and
alternatives are done — into sections/design-and-handoff.md, behind a single
STOP-Read after Phase 4.5. The live conversation (Phases 1-4.5) and the
always-run Important Rules stay in the always-loaded skeleton.

Measured: always-loaded skeleton 118,280 -> 88,975 B (-24.8%). Union preserved.
The carved AUQ is identical to pre-carve (matrix: 7/7 format, substance 5),
and Layer 0 confirms the AUQ format spec stays in the skeleton — the AUQ
paranoid suite de-risked this carve end to end.

Atomic with tests + regen (skill-docs.yml gates gen:skill-docs freshness on
every push, so source + regen + tests land together; --host all regenerates
the inlined non-Claude variants):
- sections/manifest.json: passive registry, one entry.
- parity-harness: office-hours flipped to sectioned, maxSkeletonBytes 96_000
  (measured 88,975 + headroom); content/minBytes run against the union.
- skill-size-budget: office-hours added to SECTIONS_EXTRACTED.
- gen-skill-docs + skill-validation: read the skeleton+sections union for
  office-hours so relocated Phase 5/6 prose still counts.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* chore: bump VERSION + CHANGELOG for office-hours carve + AUQ suite (v1.57.0.0)

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* refactor(preamble): carve CJK-escaping manual to on-demand doc

The AskUserQuestion format block is inlined into every interactive skill (~33).
It carried the full multi-paragraph non-ASCII/CJK escaping manual inline, but
that rationale only matters when a question contains CJK text and the operative
rule already lives in the always-loaded self-check. Moved the justification to
docs/askuserquestion-cjk.md (read on demand); kept the rule + a pointer.

Corpus: Claude-host SKILL.md total 3,087,499 -> 3,057,975 B (-29,524 B, ~900 B
x ~33 skills). Layer 0 still passes — the core decision-brief format stays
always-loaded; only the rare CJK rationale moved. Atomic with the all-host
regen (skill-docs.yml freshness gate). VERSION + package.json -> 1.58.0.0.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* refactor(plan-eng-review): carve review body into on-demand section

Fourth Phase B carve (v2_PLAN.md:220). Moves the 4-section review (Architecture,
Code Quality, Tests, Performance), outside voice, required outputs, and review
report — everything after Step 0 scope — into sections/review-sections.md behind
a single STOP-Read. Step 0 (scope challenge) and EXIT_PLAN_MODE_GATE stay in the
always-loaded skeleton.

Measured: skeleton 106,984 -> 54,892 B (-48.7%). Union preserved. Atomic with
tests + all-host regen (freshness gate): parity flipped to sectioned
(maxSkeletonBytes 62K), plan-eng-review added to SECTIONS_EXTRACTED, gen-skill-docs
reads the union for relocated review/TEST_COVERAGE/dashboard prose. Layer 0 green.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* refactor(plan-design-review): carve review body into on-demand section

Fifth Phase B carve (v2_PLAN.md:220, bundled with plan-eng). Moves the 7 design
passes, required outputs, and review report — everything after Step 0 scope and
the mockup/rating phase — into sections/review-sections.md behind a STOP-Read.
Step 0, Step 0.5 mockups, the rating method, and EXIT_PLAN_MODE_GATE stay in the
always-loaded skeleton.

Measured: skeleton 112,057 -> 76,024 B (-32.2%). Union preserved. Atomic with
tests + all-host regen: parity sectioned (maxSkeletonBytes 82K), added to
SECTIONS_EXTRACTED, gen-skill-docs reads the union. Layer 0 green.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* refactor(plan-devex-review): carve review body into on-demand section

Sixth Phase B carve. Moves the 8 DX passes, required outputs, and review report
— everything after the Step 0 DX investigation — into sections/review-sections.md
behind a STOP-Read. All of Step 0 (persona, empathy, benchmark, journey trace,
roleplay) + the rating method + EXIT_PLAN_MODE_GATE stay always-loaded.

Measured: skeleton 110,621 -> 69,658 B (-37%). Union preserved. Atomic with
tests + all-host regen: added to SECTIONS_EXTRACTED, gen-skill-docs reads the
union. Layer 0 green. (No parity invariant entry for plan-devex-review.)

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* chore: bump VERSION + CHANGELOG for plan-* family carves (v1.59.0.0)

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* test: refresh ship golden baselines + gbrain-detection union after carves

Two follow-ups the carve commits should have carried (caught by the full suite,
missed by targeted subsets):
- ship golden baselines (claude/codex/factory) regenerated: the preamble CJK
  trim (v1.58) changed ship's always-loaded AskUserQuestion block.
- gbrain-detection-override probes the office-hours skeleton+section union:
  GBRAIN_SAVE_RESULTS moved into sections/design-and-handoff.md when office-hours
  was carved, so the detection assertions now check both files.

Full `bun test` green.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* test(auq): grade format-compliance gate from SDK capture, not the TUI

The real-PTY version grepped the stripAnsi'd interactive AUQ picker. Verified
directly that this cannot work: plan-mode AUQs render as a cursor picker whose
cursor-positioning escapes stripAnsi can't flatten — the picker renders fine for
a human (cursorSeen=45) but the flattened text drops ELI10:/(recommended) and
parseNumberedOptions returns 0. The test was grading a lossy projection and
failed by construction.

Rewritten to drive /plan-ceo-review via the SDK $OUT_FILE capture (the agent
writes the verbatim question it would have shown — clean text, no rendering
loss) and grade 7/7 format + kind-note + recommendation substance >=4. Same
property, reliable, environment-independent; shares the engine with the periodic
A/B and matrix evals. Result: 7/7 format, substance 5. Touchfiles key renamed
ask-user-question-format-pty -> auq-format-gate (no longer a PTY test).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* test: fix carve-broken CI evals (union reads + section fixtures)

Two CI eval jobs failed on the carved plan-* skills because they read content
that moved into sections/:

- llm-judge (skill-llm-eval): runWorkflowJudge sliced SKILL.md between markers
  like "## Review Sections" / "## CRITICAL RULE" that now live in
  sections/review-sections.md. The markers vanished from the skeleton, so the
  judge scored empty/wrong content. Fix: read the skeleton+sections union.
  Verified: plan-ceo modes / plan-eng sections / plan-design passes all PASS
  (25/25).

- e2e-plan (skill-e2e-plan): setupPlanDir copied only <skill>/SKILL.md into the
  fixture, not sections/. The carved skill's STOP pointed at a section file that
  was absent, so the model improvised a compressed report table instead of the
  canonical "| Review | Trigger | Why | Runs | Status | Findings |". Fix: copy
  sections/ alongside SKILL.md in all 6 setup sites. Verified: report test PASS,
  canonical table emitted.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* test: copy carved sections into all e2e fixtures (prevent more carve-blind CI fails)

Proactive sweep beyond the two CI logs: every e2e test that copies a carved
skill's SKILL.md into a temp fixture must also copy its sections/, or the
model hits a STOP pointing at a missing section file and improvises/degrades.

- skill-e2e.test.ts: plan-ceo/plan-eng/plan-design/office-hours copies across
  planDir/reviewDir/ohDir/benefitsDir dests now copy sections/.
- skill-e2e-plan.test.ts: the office-hours copy + the 4-skill codex-offering
  loop now copy sections/.
- skill-e2e-design.test.ts: plan-design-review copy now copies sections/.
- skill-e2e-office-hours.test.ts: both office-hours copies now copy sections/.
- skill-e2e-office-hours-brain-writeback.test.ts: GBRAIN_SAVE_RESULTS moved into
  the section, so check the regenerated skeleton+section UNION for the gbrain put
  block, ship both into the workdir, and restore both (the section regen was also
  leaking into the working tree — finally now restores it).

ship copies (single-file Step-0 slices) and review/retro (not carved) untouched.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* test: migrate section-loading E2E to lossless SDK tool-stream detection

The /ship and /plan-ceo-review section-loading tests drove a real PTY and
scraped the ANSI screen buffer for sections/<file>.md paths. That silently
saw nothing in a Conductor PTY (cursor-positioned tool renders and an
unanswered Step 0 question loop both defeat the regex), so both reported
read: [] even when the agent did the work.

They now run the skill through claude -p (the same SDK path the AUQ matrix
uses) and detect section reads from the tool-use stream — Read calls whose
file_path contains sections/<file>.md — with no rendering layer to mangle.
The run is also hermetic: the freshly-generated worktree skeleton + sections
are copied into a throwaway fixture with the absolute path pinned, so the
test validates this branch's carve without mutating the user's ~/.claude
install.

Validated EVALS_TIER=periodic: both pass (plan-ceo Reads review-sections.md;
ship Reads review-army.md + changelog.md), ~6.5 min for both vs ~23 min
combined on the old PTY path where both were failing.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* chore: consolidate branch to v1.56.0.0 (single MINOR above main)

The branch bumped VERSION several times during development (1.56 → 1.57 →
1.58 → 1.59), but none of those landed on main (main is at 1.55.1.0). Per
the "never orphan branch-internal versions" discipline, collapse all four
into a single 1.56.0.0 entry — one MINOR release covering the whole branch:
five skills carved (plan-ceo, office-hours, plan-eng, plan-design,
plan-devex), the shared AskUserQuestion preamble CJK trim, and the paranoid
AUQ no-degradation test suite + lossless section-loading tests.

VERSION and package.json set to 1.56.0.0; main's 1.55.1.0 entry preserved
below the consolidated entry. No SKILL.md drift (VERSION is not embedded in
generated bodies).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: garrytan

CHANGELOG.md

@@ -1,5 +1,65 @@
 # Changelog
 
+## [1.56.0.0] - 2026-06-03
+
+## **Five heavy skills now load their bulk on demand, the shared question preamble slimmed corpus-wide, and a paranoid test suite proves the questions never got worse.**
+
+The token-reduction program lands its biggest wave. Five of the heaviest skills — `/plan-ceo-review`, `/office-hours`, `/plan-eng-review`, `/plan-design-review`, and `/plan-devex-review` — are now a small always-loaded skeleton plus an on-demand `sections/` file the agent opens only when it reaches the work. The conversational front half (Step 0 scope, the live interview) stays always-loaded; the deep review bodies, design-doc templates, outside-voice rules, and required-output writers move behind a single STOP-Read. The shared AskUserQuestion preamble also shed its rarely-needed CJK escaping manual, so every interactive skill is a little lighter at once. And because the most user-facing surface in gstack is the question it asks you, a new paranoid test suite proves that slimming a skill never strands or degrades that question.
+
+### The numbers that matter
+
+Measured from the generated skeletons (`wc -c <skill>/SKILL.md`), regenerated for all hosts:
+
+| Skill | Before | After | Δ |
+|-------|--------|-------|---|
+| plan-ceo-review | 138,838 B | 80,731 B | -42.0% |
+| office-hours | 118,280 B | 88,975 B | -24.8% |
+| plan-eng-review | 106,984 B | 54,892 B | -48.7% |
+| plan-design-review | 112,057 B | 76,024 B | -32.2% |
+| plan-devex-review | 110,621 B | 69,658 B | -37.0% |
+
+On top of the per-skill carves, the shared AskUserQuestion preamble dropped its inline CJK manual to a one-line rule + a doc pointer, trimming ~29,524 B across the Claude-host corpus (every interactive skill, ~900 B each).
+
+The AskUserQuestion proof, measured by SDK capture (`test/skill-e2e-auq-matrix.test.ts`):
+
+| Guarantee | Result |
+|-----------|--------|
+| Carved vs verbose, same trigger | 7/7 format, substance 5 == 7/7 format, substance 5 (no degradation) |
+| Matrix across 7 AUQ-heavy skills | all 7/7 format, substance 4-5 |
+| Same trigger 3× (consistency) | stable, every format element every run |
+| AUQ format spec always-loaded | guaranteed in every skeleton (Layer 0) |
+
+Every review runs identical pass for pass; the only thing that changed is what sits in context before the work begins.
+
+### What this means for you
+
+The skills you run most start markedly lighter: a `/plan-ceo-review` opens ~42% smaller, the others 25-49%, and they pull in their review bodies only when they reach them. You will not notice any behavior change in the reviews themselves; they run section for section as before. What you get is more of the context window left for your actual work, paid back on every invocation. And every skill that asks you questions now carries a guarantee, enforced by tests: the decision brief (plain-English ELI10, an explicit recommendation with a real reason, pros and cons, the stakes) is provably in context the instant any question fires. External hosts (codex, factory, kiro, opencode) still receive the full inline skill, so nothing regresses off Claude.
+
+### Itemized changes
+
+#### Added
+- `plan-ceo-review/sections/review-sections.md` — the 11-section deep review, outside-voice rules, required-output registries, completion summary, review report writer, next-step chaining, and mode quick reference, behind a STOP-Read pointer with a passive `manifest.json`.
+- `office-hours/sections/design-and-handoff.md` — Phase 5 design-doc templates + Phase 6 tiered handoff, behind a STOP-Read pointer with a passive `manifest.json`.
+- `/plan-eng-review`, `/plan-design-review`, `/plan-devex-review` each gain a `sections/review-sections.md` carved behind a post-Step-0 STOP-Read.
+- `docs/askuserquestion-cjk.md` — full non-ASCII / CJK escaping rationale + worked example, read on demand.
+- `test/auq-format-always-loaded.test.ts` — free per-PR keystone: every interactive skill must carry the full AskUserQuestion format spec in its always-loaded skeleton, never stranded in a section. 51 cases plus a negative control.
+- `test/skill-e2e-auq-matrix.test.ts` — drives each AUQ-heavy skill to its first question and grades it (7/7 format, substance >=4).
+- `test/skill-e2e-auq-verbose-vs-carved-ab.test.ts` — proves a carved skill's question is not worse than the pre-carve monolith's, on the same trigger.
+- `test/skill-e2e-auq-consistency.test.ts` — same trigger N times, fails on any format element that flickers between runs.
+- `test/codex-e2e-recommendation-substance.test.ts` — grades `/codex`'s live recommendation substance.
+- `test/skill-ceo-section-ordering.test.ts` — gate-tier static guard: the STOP fires after Step 0, the review body is absent from the skeleton, the report writer lives in the section, and nothing review-governing sits below the STOP.
+- `test/skill-e2e-plan-ceo-review-section-loading.test.ts` — periodic backstop that asserts the carved section is Read before the report.
+
+#### Changed
+- `/plan-ceo-review`, `/office-hours`, `/plan-eng-review`, `/plan-design-review`, `/plan-devex-review` are each a skeleton + one on-demand section on Claude; the conversational front (Step 0 / Phases 1-4.5) stays always-loaded; external hosts still receive the full inline skill (no behavior change off Claude).
+- The AskUserQuestion preamble trims the inline CJK manual to the operative rule + a doc pointer; the always-loaded self-check is unchanged.
+- Parity, size-budget, section-manifest, gen-skill-docs, and skill-validation treat every carved skill consistently (content + size floors run against the skeleton + section union; skeleton-shrink assertions guard the always-loaded win).
+
+#### For contributors
+- `test/helpers/auq-sdk-capture.ts` — reusable SDK capture engine: drives a skill to its AUQ and captures the verbatim generated text cleanly (real-PTY mangles plan-mode questions), grades format + recommendation substance robust to the connective, and detects section reads losslessly from the tool-use stream.
+- `section-manifest-consistency` discovers every carved skill automatically, so the next carve is covered the moment its manifest lands.
+- The `/ship` and `/plan-ceo-review` section-loading E2E tests detect section reads from the `claude -p` tool-use stream instead of scraping the real-PTY screen buffer, so they are reliable (the PTY path silently saw nothing in some terminals) and run hermetically against the worktree carve without mutating the installed skill.
+
 ## [1.55.1.0] - 2026-06-02
 
 ## **Telemetry now tells you exactly what it records and where it stays. The project-slug helper hands the shell a safe identifier on every path.**

TODOS.md

@@ -19,6 +19,38 @@ v1.47.0.0 baselines retained in `test/fixtures/` for the v1→v2 audit trail. Th
 captured skill bytes match `origin/main` exactly (the rebasing branch left every
 SKILL.md untouched). `bun test` is green again.
 
+## Token-reduction follow-ups (Phase B, filed via /plan-eng-review on the plan-ceo-review carve)
+
+### P3: Carve the always-loaded `{{PREAMBLE}}` reference blocks into an on-demand doc
+
+**What:** The per-skill section carves (`/ship` v1.54, `/plan-ceo-review` v1.56) yield
+real but bounded wins (-42% to -59% on the carved skill) because the shared
+`{{PREAMBLE}}` (~40-50KB on every tier-3/4 skill) is the dominant always-loaded cost
+and stays inline. Move the rarely-needed preamble REFERENCE blocks (the AskUserQuestion
+split-rules and the CJK / lone-surrogate escaping reference) into an on-demand
+section-style doc the agent reads only when it hits those edge cases, leaving the hot
+path (voice, completeness principle, recommendation format) inline.
+
+**Why:** Highest-ROI remaining token target. One preamble carve helps EVERY tier-≥2
+skill at once, not one skill per PR. The eng-review on the plan-ceo carve flagged that
+per-skill carves stay modest precisely because the preamble dominates the always-loaded
+surface.
+
+**Pros:** A single change reduces always-loaded cost across the whole skill pack.
+**Cons:** The preamble is load-bearing and shared; a botched carve regresses every skill.
+Needs the same union-parity + per-push freshness guards the section carves use, applied
+corpus-wide.
+
+**Context:** Builds on the v2 section pipeline (`scripts/resolvers/sections.ts`,
+`{{SECTION:id}}` / `{{SECTION_INDEX}}`). The preamble source is
+`scripts/resolvers/preamble.ts`. Measure which sub-blocks are cold (escaping reference,
+split-rules) vs hot (voice, recommendation format) before cutting. Validate on one skill,
+then roll corpus-wide.
+
+**Effort estimate:** L (human team) → M (CC+gstack)
+**Priority:** P3
+**Depends on / blocked by:** The section pipeline (shipped v1.54). No hard blocker.
+
 ## gbrowser memory follow-ups (filed via /plan-eng-review + /codex on the v1.49 leak-fix PR)
 
 These four items came out of the memory-leak investigation that shipped

VERSION

@@ -1 +1 @@
-1.55.1.0
+1.56.0.0

autoplan/SKILL.md

@@ -371,25 +371,12 @@ so split chains are never AUTO_DECIDE-eligible — the user's option set is sacr
 **Full rule + worked examples + Hold/dependency semantics:** see
 `docs/askuserquestion-split.md` in the gstack repo. Read on demand when N>4.
 
-**Non-ASCII characters — write directly, never \u-escape.** When any
-    string field (question, option label, option description) contains
-    Chinese (繁體/簡體), Japanese, Korean, or other non-ASCII text, emit
-    the literal UTF-8 characters in the JSON string. **Never escape them
-    as `\uXXXX`.** Claude Code's tool parameter pipe is UTF-8 native
-    and passes characters through unchanged. Manually escaping requires
-    recalling each codepoint from training, which is unreliable for long
-    CJK strings — the model regularly emits the wrong codepoint (e.g.
-    writes `\u3103` thinking it is 管 U+7BA1, but `\u3103` is
-    actually ㄃, so the user sees `管理工具` rendered as `㄃3用箱`).
-    The trigger is long, multi-line questions with hundreds of CJK
-    characters: that is exactly when reflexive escaping kicks in and
-    exactly when miscoding is most damaging. Long ≠ escape. Keep
-    characters literal.
-
-    Wrong: `"question": "請選擇\uXXXX\uXXXX\uXXXX\uXXXX"`
-    Right: `"question": "請選擇管理工具"`
-
-    Only JSON-mandatory escapes remain allowed: `\n`, `\t`, `\"`, `\\`.
+**Non-ASCII characters — write directly, never \u-escape.** When any string
+field contains Chinese (繁體/簡體), Japanese, Korean, or other non-ASCII text,
+emit the literal UTF-8 characters; never escape them as `\uXXXX` (the pipe is
+UTF-8 native, and manual escaping miscodes long CJK strings). Only `\n`,
+`\t`, `\"`, `\\` remain allowed. Full rationale + worked example: see
+`docs/askuserquestion-cjk.md`. Read on demand when a question contains CJK.
 
 ### Self-check before emitting
 

canary/SKILL.md

@@ -363,25 +363,12 @@ so split chains are never AUTO_DECIDE-eligible — the user's option set is sacr
 **Full rule + worked examples + Hold/dependency semantics:** see
 `docs/askuserquestion-split.md` in the gstack repo. Read on demand when N>4.
 
-**Non-ASCII characters — write directly, never \u-escape.** When any
-    string field (question, option label, option description) contains
-    Chinese (繁體/簡體), Japanese, Korean, or other non-ASCII text, emit
-    the literal UTF-8 characters in the JSON string. **Never escape them
-    as `\uXXXX`.** Claude Code's tool parameter pipe is UTF-8 native
-    and passes characters through unchanged. Manually escaping requires
-    recalling each codepoint from training, which is unreliable for long
-    CJK strings — the model regularly emits the wrong codepoint (e.g.
-    writes `\u3103` thinking it is 管 U+7BA1, but `\u3103` is
-    actually ㄃, so the user sees `管理工具` rendered as `㄃3用箱`).
-    The trigger is long, multi-line questions with hundreds of CJK
-    characters: that is exactly when reflexive escaping kicks in and
-    exactly when miscoding is most damaging. Long ≠ escape. Keep
-    characters literal.
-
-    Wrong: `"question": "請選擇\uXXXX\uXXXX\uXXXX\uXXXX"`
-    Right: `"question": "請選擇管理工具"`
-
-    Only JSON-mandatory escapes remain allowed: `\n`, `\t`, `\"`, `\\`.
+**Non-ASCII characters — write directly, never \u-escape.** When any string
+field contains Chinese (繁體/簡體), Japanese, Korean, or other non-ASCII text,
+emit the literal UTF-8 characters; never escape them as `\uXXXX` (the pipe is
+UTF-8 native, and manual escaping miscodes long CJK strings). Only `\n`,
+`\t`, `\"`, `\\` remain allowed. Full rationale + worked example: see
+`docs/askuserquestion-cjk.md`. Read on demand when a question contains CJK.
 
 ### Self-check before emitting
 

codex/SKILL.md

@@ -366,25 +366,12 @@ so split chains are never AUTO_DECIDE-eligible — the user's option set is sacr
 **Full rule + worked examples + Hold/dependency semantics:** see
 `docs/askuserquestion-split.md` in the gstack repo. Read on demand when N>4.
 
-**Non-ASCII characters — write directly, never \u-escape.** When any
-    string field (question, option label, option description) contains
-    Chinese (繁體/簡體), Japanese, Korean, or other non-ASCII text, emit
-    the literal UTF-8 characters in the JSON string. **Never escape them
-    as `\uXXXX`.** Claude Code's tool parameter pipe is UTF-8 native
-    and passes characters through unchanged. Manually escaping requires
-    recalling each codepoint from training, which is unreliable for long
-    CJK strings — the model regularly emits the wrong codepoint (e.g.
-    writes `\u3103` thinking it is 管 U+7BA1, but `\u3103` is
-    actually ㄃, so the user sees `管理工具` rendered as `㄃3用箱`).
-    The trigger is long, multi-line questions with hundreds of CJK
-    characters: that is exactly when reflexive escaping kicks in and
-    exactly when miscoding is most damaging. Long ≠ escape. Keep
-    characters literal.
-
-    Wrong: `"question": "請選擇\uXXXX\uXXXX\uXXXX\uXXXX"`
-    Right: `"question": "請選擇管理工具"`
-
-    Only JSON-mandatory escapes remain allowed: `\n`, `\t`, `\"`, `\\`.
+**Non-ASCII characters — write directly, never \u-escape.** When any string
+field contains Chinese (繁體/簡體), Japanese, Korean, or other non-ASCII text,
+emit the literal UTF-8 characters; never escape them as `\uXXXX` (the pipe is
+UTF-8 native, and manual escaping miscodes long CJK strings). Only `\n`,
+`\t`, `\"`, `\\` remain allowed. Full rationale + worked example: see
+`docs/askuserquestion-cjk.md`. Read on demand when a question contains CJK.
 
 ### Self-check before emitting
 

context-restore/SKILL.md

@@ -367,25 +367,12 @@ so split chains are never AUTO_DECIDE-eligible — the user's option set is sacr
 **Full rule + worked examples + Hold/dependency semantics:** see
 `docs/askuserquestion-split.md` in the gstack repo. Read on demand when N>4.
 
-**Non-ASCII characters — write directly, never \u-escape.** When any
-    string field (question, option label, option description) contains
-    Chinese (繁體/簡體), Japanese, Korean, or other non-ASCII text, emit
-    the literal UTF-8 characters in the JSON string. **Never escape them
-    as `\uXXXX`.** Claude Code's tool parameter pipe is UTF-8 native
-    and passes characters through unchanged. Manually escaping requires
-    recalling each codepoint from training, which is unreliable for long
-    CJK strings — the model regularly emits the wrong codepoint (e.g.
-    writes `\u3103` thinking it is 管 U+7BA1, but `\u3103` is
-    actually ㄃, so the user sees `管理工具` rendered as `㄃3用箱`).
-    The trigger is long, multi-line questions with hundreds of CJK
-    characters: that is exactly when reflexive escaping kicks in and
-    exactly when miscoding is most damaging. Long ≠ escape. Keep
-    characters literal.
-
-    Wrong: `"question": "請選擇\uXXXX\uXXXX\uXXXX\uXXXX"`
-    Right: `"question": "請選擇管理工具"`
-
-    Only JSON-mandatory escapes remain allowed: `\n`, `\t`, `\"`, `\\`.
+**Non-ASCII characters — write directly, never \u-escape.** When any string
+field contains Chinese (繁體/簡體), Japanese, Korean, or other non-ASCII text,
+emit the literal UTF-8 characters; never escape them as `\uXXXX` (the pipe is
+UTF-8 native, and manual escaping miscodes long CJK strings). Only `\n`,
+`\t`, `\"`, `\\` remain allowed. Full rationale + worked example: see
+`docs/askuserquestion-cjk.md`. Read on demand when a question contains CJK.
 
 ### Self-check before emitting
 

context-save/SKILL.md

@@ -366,25 +366,12 @@ so split chains are never AUTO_DECIDE-eligible — the user's option set is sacr
 **Full rule + worked examples + Hold/dependency semantics:** see
 `docs/askuserquestion-split.md` in the gstack repo. Read on demand when N>4.
 
-**Non-ASCII characters — write directly, never \u-escape.** When any
-    string field (question, option label, option description) contains
-    Chinese (繁體/簡體), Japanese, Korean, or other non-ASCII text, emit
-    the literal UTF-8 characters in the JSON string. **Never escape them
-    as `\uXXXX`.** Claude Code's tool parameter pipe is UTF-8 native
-    and passes characters through unchanged. Manually escaping requires
-    recalling each codepoint from training, which is unreliable for long
-    CJK strings — the model regularly emits the wrong codepoint (e.g.
-    writes `\u3103` thinking it is 管 U+7BA1, but `\u3103` is
-    actually ㄃, so the user sees `管理工具` rendered as `㄃3用箱`).
-    The trigger is long, multi-line questions with hundreds of CJK
-    characters: that is exactly when reflexive escaping kicks in and
-    exactly when miscoding is most damaging. Long ≠ escape. Keep
-    characters literal.
-
-    Wrong: `"question": "請選擇\uXXXX\uXXXX\uXXXX\uXXXX"`
-    Right: `"question": "請選擇管理工具"`
-
-    Only JSON-mandatory escapes remain allowed: `\n`, `\t`, `\"`, `\\`.
+**Non-ASCII characters — write directly, never \u-escape.** When any string
+field contains Chinese (繁體/簡體), Japanese, Korean, or other non-ASCII text,
+emit the literal UTF-8 characters; never escape them as `\uXXXX` (the pipe is
+UTF-8 native, and manual escaping miscodes long CJK strings). Only `\n`,
+`\t`, `\"`, `\\` remain allowed. Full rationale + worked example: see
+`docs/askuserquestion-cjk.md`. Read on demand when a question contains CJK.
 
 ### Self-check before emitting
 

cso/SKILL.md

@@ -369,25 +369,12 @@ so split chains are never AUTO_DECIDE-eligible — the user's option set is sacr
 **Full rule + worked examples + Hold/dependency semantics:** see
 `docs/askuserquestion-split.md` in the gstack repo. Read on demand when N>4.
 
-**Non-ASCII characters — write directly, never \u-escape.** When any
-    string field (question, option label, option description) contains
-    Chinese (繁體/簡體), Japanese, Korean, or other non-ASCII text, emit
-    the literal UTF-8 characters in the JSON string. **Never escape them
-    as `\uXXXX`.** Claude Code's tool parameter pipe is UTF-8 native
-    and passes characters through unchanged. Manually escaping requires
-    recalling each codepoint from training, which is unreliable for long
-    CJK strings — the model regularly emits the wrong codepoint (e.g.
-    writes `\u3103` thinking it is 管 U+7BA1, but `\u3103` is
-    actually ㄃, so the user sees `管理工具` rendered as `㄃3用箱`).
-    The trigger is long, multi-line questions with hundreds of CJK
-    characters: that is exactly when reflexive escaping kicks in and
-    exactly when miscoding is most damaging. Long ≠ escape. Keep
-    characters literal.
-
-    Wrong: `"question": "請選擇\uXXXX\uXXXX\uXXXX\uXXXX"`
-    Right: `"question": "請選擇管理工具"`
-
-    Only JSON-mandatory escapes remain allowed: `\n`, `\t`, `\"`, `\\`.
+**Non-ASCII characters — write directly, never \u-escape.** When any string
+field contains Chinese (繁體/簡體), Japanese, Korean, or other non-ASCII text,
+emit the literal UTF-8 characters; never escape them as `\uXXXX` (the pipe is
+UTF-8 native, and manual escaping miscodes long CJK strings). Only `\n`,
+`\t`, `\"`, `\\` remain allowed. Full rationale + worked example: see
+`docs/askuserquestion-cjk.md`. Read on demand when a question contains CJK.
 
 ### Self-check before emitting