Leaf finding: lobotomized overrides break CC 2.1.168 — orphan vars + boot crash (Adoption record evidence)

[dividedby]

What

The first real run of the Integration gate (RealAdoptionEnvironment, #22) against a live Claude Code 2.1.168 install produced a FAIL verdict: the lobotomized overrides, applied to 2.1.168, leave unbacked Orphan variables and the patched binary crashes on boot. This is a real adoption breakage in the leaves, surfaced exactly as the gate is designed to.

Per the contributor-cockpit model (CONTEXT.md → Control plane), the control plane prepares a verified finding — it does not direct-push to the leaves (skrabe). This issue is that prepared finding; the maintainer decides how to surface it to the leaf owner.

Evidence — Adoption record (HITL run, 2026-06-08)

• Boot crash (runtime, claude -p): ReferenceError: IS_TRUTHY_FN is not defined — the patched binary fails to start.
• Static orphan validator (vs. bundled prompts-2.1.168.json identifierMap): 7 orphans —
  GLOB_TOOL_NAME, GREP_TOOL_NAME, READ_TOOL_NAME, SHELL_TOOL_NAME, IS_BASH_ENV_FN, USE_EMBEDDED_TOOLS_FN, EXIT_PLAN_MODE_TOOL_NAME.
• Gate verdict: pass: false, bootVerifyPassed: false, non-zero exit.

{
  "pass": false,
  "versions": [{
    "ccVersion": "2.1.168",
    "fourZeros": {
      "pass": false,
      "failedPatches": [],
      "missingSystemPrompts": [],
      "orphanVariables": ["GLOB_TOOL_NAME","GREP_TOOL_NAME","READ_TOOL_NAME","SHELL_TOOL_NAME","IS_BASH_ENV_FN","USE_EMBEDDED_TOOLS_FN","EXIT_PLAN_MODE_TOOL_NAME"],
      "bootVerifyPassed": false
    }
  }],
  "date": "2026-06-08T02:04:13.648Z"
}

(The restoreDrill block in the raw record is stubbed/placeholder — #23 — and not evidence.)

Note

The runtime crash (IS_TRUTHY_FN) and the static validator's 7 names are different sets — both correctly conclude "broken," but the divergence is itself a validator-source-alignment issue tracked separately. Either way, the adoption is genuinely broken on 2.1.168.

Suggested action (human decides)

• Surface to skrabe as a prepared finding (orphan vars + boot crash on 2.1.168), with this record as PR evidence per the gate's purpose (PRD PRD: release-adoption control plane (integration gate + release-detector substrate) #2, user stories 9–10).
• Likely a leaf fix: realign the affected overrides' placeholders against 2.1.168's identifierMap (renamed/inlined upstream).

Related

• Surfaced by RealAdoptionEnvironment adopt path — real --apply / claude -p boot-verify / orphan validator (HITL) #22 (RealAdoptionEnvironment adopt path) / PRD PRD: RealAdoptionEnvironment — real shell-out adapter behind the gate seam (HITL) #20.
• Validator-source-alignment follow-up: see companion issue.
• Leaf test broadening: Roadmap: leaf test broadening (tweakcc-fixed + lobotomized) #8.

🤖 Generated with Claude Code

[dividedby]

End-to-end gate verification (real CC 2.1.168) — confirms broken, characterizes the cause

Ran the real Integration gate against an installed CC 2.1.168 with both prepared leaf PRs checked out and live — skrabe/tweakcc-fixed#4 (rebuilt dist) + skrabe/lobotomized-claude-code#4 (realigned overrides on disk). Restore drill passed (clean rollback; install untouched).

Verdict: pass: false · bootVerifyPassed: false. The two prepared PRs are a correct but partial fix — 2.1.168 still does not boot. Root cause is now precisely characterized:

1. The boot crash is the _FN()/_OBJECT.member expression class

The orphans that actually crash are expression-form inside template-literal interpolation slots, not bare placeholders:

${!IS_TRUTHY_FN(PROCESS_OBJECT.env.CLAUDE_CODE_DISABLE_BACKGROUND_TASKS)&&...?`…`}
${ADDITIONAL_DREAM_GUIDANCE_FN()}   ${ATTACHMENT_OBJECT.blockingError.command}

tweakcc-fixed#4's skip-guard regex (?<!\\)\$\{(NAME)\} matches only a bare ${NAME}, so it cannot skip these — the crashing prompt is spliced in and ReferenceError: IS_TRUTHY_FN is not defined fires. This matches ADR 0005's original observation (Boot-verify caught IS_TRUTHY_FN; the static check did not).

2. The override realign is far from complete

lobotomized#4 realigns 5 files. Genuine unfixed orphans remain in untouched overrides:

• AGENT_TOOL_NAME — 14 files
• IS_TRUTHY_FN — tool-description-agent-usage-notes.md, agent-prompt-schedule-slash-command.md
• PROCESS_OBJECT, TODO_TOOL_OBJECT, PR_PREAMBLE_STUB, CONDITIONAL_WHAT_HAPPENS_NOTE

Implication

The two PRs should not be presented to skrabe as "fixes 2.1.168." They are a partial realign plus a guard that structurally can't catch the crash class. A green 2.1.168 adoption needs (a) the patcher guard extended to expression-form slots and (b) a full override realign covering the inventory above. See #31 for the detection-scope follow-up.

[dividedby]

Confirmed pre-existing & version-independent — surfaced by the 2.1.169 gate boot-verify

While running the 2.1.169 Integration gate (HITL, #58), the patched binary crashed at boot with ReferenceErrors from lobotomized overrides whose ${VAR} references resolve in no recent CC version. This is the breakage #26 describes — and it is pre-existing, not a 2.1.169 regression:

Override (tracked, opus-4-8)	pinned ccVersion	references	.168 map	.169 map
tool-description-askuserquestion.md	2.1.156	${PROMPT_VAR_0}, ${PROMPT_VAR_1}	ENTER/EXIT_PLAN_MODE_TOOL_NAME	ENTER/EXIT_PLAN_MODE_TOOL_NAME
tool-description-croncreate.md	2.1.144	${CRON_DURABLE_FLAG}	CRON_DURABILITY_SECTION (renamed)	CRON_DURABILITY_SECTION

Both slot names match neither .168 nor .169 — the prompts were restructured/renamed after the override's pinned version and the overrides were never realigned. So applying these overrides breaks 2.1.168 too (root-cause / re-baseline this issue asked for). Crash form: ${PROMPT_VAR_0} leaks literally into an evaluated bundle template literal → ReferenceError: PROMPT_VAR_0 is not defined at startup.

Two sub-findings:

1. Static orphan-validator under-reports. PROMPT_VAR_0 boot-crashed but was not in the gate's advisoryOrphans list (which showed IS_TRUTHY_FN, PROCESS_OBJECT, CRON_DURABLE_FLAG, CONDITIONAL_WHAT_HAPPENS_NOTE). The validator misses some runtime-fatal orphans — worth a coverage issue against the orphan-validator.
2. Runtime vs tracked. A third crash (EMPTY_STRING, memory-synthesis) was runtime-clone pollution only — the tracked override correctly uses ${OPTIONAL_TAIL_NOTE}. (See runtime-vs-tracked-override-dirs.)

Recommend un-parking. Fix = realign the conflict-flagged overrides (workflow / croncreate / askuserquestion / explore / memory-synthesis) to current slot names; mechanical for pure renames, content reconciliation for restructured prompts (curation). The skip-unresolved-placeholder detector (#45 / skrabe/tweakcc-fixed#4) is the runtime guard that would degrade these gracefully instead of crashing.

[dividedby]

Mechanical subset shipped → skrabe/lobotomized-claude-code#5

Re-baselined the orphan inventory against the current lobotomized origin/main (post skrabe convergence: 767011b upstream identifier vocab, 54e0d34 croncreate 7-slot) and the authoritative 2.1.169 identifierMap (tweakcc-fixed adopt/cc-2.1.169:data/prompts/prompts-2.1.169.json). Most of the original 2.1.168 inventory (AGENT_TOOL_NAME×14, IS_TRUTHY_FN, PROCESS_OBJECT…) is already resolved. The stale PR #4 (fix/realign-explore-askuserquestion-2.1.168, pre-convergence base) is superseded.

Done — leaf PR #5 (safe + mechanical)

Realign orphaned named-prompt slots (note: the patcher's bc60baa skip-guard means these no longer boot-crash — bare placeholders in the identifierMap union are skipped, so the override silently stopped applying; realigning re-activates it):

• agent-prompt-memory-synthesis (4-8): ${EMPTY_STRING} → ${OPTIONAL_TAIL_NOTE}
• tool-description-enterplanmode (4-7): ${CONDITIONAL_WHAT_HAPPENS_NOTE} → ${CONDITIONAL_WHAT_HAPPENS_NOTE_FN()} (mirrors correct 4-8)

Escape latent boot crashes (not in identifierMap → guard can't skip → ReferenceError when spliced):

• data-anthropic-cli (4-8): \${VERSION} (4-7 sibling already escaped)
• skill-cowork-plugin-authoring-mcp-discovery (4-8): \${GITHUB_TOKEN}, \${DATADOG_API_KEY}, \${DATADOG_APP_KEY}

Still open — needs curation / patcher altitude (ready-for-human)

• Restructured conditionals (content reconciliation, not rename):
  • tool-description-croncreate (4-7): the ${CRON_DURABLE_FLAG?A:B} conditional became a single bundle-owned ${CRON_DURABILITY_SECTION}; the 4-7 file carries its own custom durable/session prose that can't be injected the same way. (4-8 was already realigned by 54e0d34.)
  • tool-description-workflow (4-8): 5 slots renamed+restructured (WORKFLOW_TOOL_NAME/… → WORKFLOW_INVOCATION_QUALIFIER/RESEND_NOTE/ISOLATION_TYPE/WORKTREE_NOTE/GROUP_GLYPH).
• Inline-blob minified identifiers — the inline-* overrides (both 4-7 and 4-8) reference version-specific minified names (${ZD}/${dY}, ${bY8}/${bV8}, ${QW}/${cY}, …) pinned to their extraction's bundle. Realigning means re-extracting current minified names from the target bundle (extractor work), not a hand rename — and it's a recurring fragility worth a design note, not a one-off fix.

Static authoring-drift pre-check only (all 5 targeted orphans clear, no new orphans). Runtime boot-verify (HITL Integration gate) remains the authority and hasn't been run on PR #5. Leaving this issue open; #58's lobotomized half stays blocked until the curation remainder + a green full-override gate.