language-discipline — internal English, human-facing AIF_HOOK_LANG-gated
Class A rule with a 3-category model: internal machinery is English-only (CI principle test 22), operator-facing output follows AIF_HOOK_LANG, and match-data stays bilingual — translating it away silently breaks Russian-input recognition.
Census id: F18 · Family: F — rules corpus + generated-rule tooling + shipped audit scripts · Provenance: framework pin
b069c59328aa8e08671d3a19fd25efa618dfc1c5(stagingFETCH_HEAD, read 2026-09-11; re-anchored from census pina1337cb301). Everyfile:linebelow is at that pin. Raw reference drafted by stage E4 — factual, unpolished by design.
| Field | Value |
|---|---|
| Status | raw reference (census row F18: MISSING → drafted E4; census satellite D3) |
| Ships-to | clone only (setup.d/20-agents.sh:44 «(rules/ is not shipped)») |
| Fires-at | «writing any internal machinery or human-facing output.» (.claude/rules/language-discipline.md:14) — paths:(3) + edit-time inject |
What it bans
Mixed-language internal machinery, and match-data translation. .claude/rules/language-discipline.md:8 «# Language discipline — internal English, human-facing AIF_HOOK_LANG-gated»; the 3-category model (:22-26): category 1 «Internal machinery — comments, logic, AI-facing SKILL instructions, tool args, code, repo artifacts» → «English, always»; category 2 «Human-facing output» → «AIF_HOOK_LANG-gated: ru → Russian, else English»; category 3 «Activation / match metadata» → «bilingual / language-specific kept». The category-3 trap: :28 «Category 3 is the trap: removing those Russian tokens silently breaks recognition of Russian operator input/output, with no upside.» Class header: :13 «> Class: A — companion principle test shipped at [packages/core/principles/22-internal-english.test.ts] (2026-06-16).»
Never (fires)
The anti-patterns: .claude/rules/language-discipline.md:51 «- #russian-in-machinery — Cyrillic comment/prose in a hook/script/skill body. Counter: translate to English; principle 22 Surface 1/2 catches it.»; :52 «- #headers-localized-prose-not — localizing shell tokens but writing prose in the wrong language (the 2026-06-16 pipeline bug).» (the origin incident: :18 «/pipeline localized table headers but wrote English prose even under AIF_HOOK_LANG=ru»); :53 «- #match-data-translated-away — "making it English" by deleting category-3 Russian tokens, breaking Russian-input recognition.»
Always (clean)
The two human-facing sub-channels: :34 «- 2a shell-emitted (hooks/helpers echo, no LLM in the loop): via lang/{en,ru}.sh packs, selected by ${AIF_HOOK_LANG:-en} with a hard EN fallback.» and :35 «- 2b LLM-authored (the model writes prose): the model learns the active language from (i) the always-on line injected by [inject-session-bootstrap.sh] … and (ii) the pipeline AIF_OUTPUT_LANG render-time signal, and MUST write ALL operator-facing prose in it». Precedence: :37 «AIF_HOOK_LANG overrides the prompt-language default when set; unset → English.»
Enforcement channels
- Rule-index row:
.claude/rules/00-rule-index.md:26«|language-discipline.md| A | writing any internal machinery or human-facing output. | paths:(3), edit-time inject |» - Edit-time inject, 3 globs:
.claude/rules/language-discipline.md:10«», inject text at:11. - CI principle test:
packages/core/principles/22-internal-english.test.ts— «Surface 1 (machinery shell, zero-tolerance) + Surface 2 (SKILL.md bodies after frontmatter, allowlisted)» (:45); a deterministic Cyrillic grep, zero API calls (:62). - Always-on injection (not a gate): the B1 output-language line in
inject-session-bootstrap.sh— «judgment-shaped prose-following, so injection, not a gate» (:46). - Parity checks:
.claude/hooks/lang/check-parity.sh+.claude/skills/pipeline/lang/check-parity.sh(en/ru key parity,:47). - Not backend-rendered — no FF diagnostic applies (FF7001/FF7002,
packages/core/diagnostics/registry.ts:321/:329); honest status: CI grep + injection + parity scripts.
kickoff-staging-placement — dispatch-input placement
Class A rule: dispatch kickoffs are tracked files read from the staging branch — merge the kickoff to staging BEFORE dispatching; plus the two render-as-success traps (CANON-symlink swallow, kickoff-name near-miss) with their shipped gates.
memory-codification — codify durable conventions
Class B rule: a durable behavioural convention written to agent memory is codified into the repo in the same step and the memory entry reduced to a pointer — CI is structurally unreachable, so the ceiling is B by construction.