polyglot-recipes

name	polyglot-recipes
description	> Use when this capability is needed.

polyglot-recipes — cross-language seams as ripvec sees them

Be brief. Cite docs/AGENTIC_PATTERNS_4_0.md Part IX (NC1c, NC8, M5), Part X (NC16, M22), Part XI §XI.4 (M6 trichotomy) and BUG_DATABASE.md §§4-5 (I#54a/b polyglot edge bugs). aurora is the M14 polyglot canonical (HCL+Python+SQL, validated through 3 cycles); the constitutional polyglot sentinel.

§0 Graph position

Specializes the five hub orientations for cross-language seams. This skill fires when no single per-language skill (c-recipes, python-recipes, javascript-recipes, jvm-recipes, go-recipes, rust-recipes) captures the task because the interesting edges cross language boundaries. Most often escalates to ripvec:drift-auditor because polyglot dark-matter manifests as systemic dead-zone drift — ripvec can't see the seam, so the seam silently rots.

§1 Polyglot character

What ripvec sees on a polyglot corpus:

• Cross-language string-as-API is everywhere. HCL outputs become Python config values become SQL @{variable_name} interpolations become Jinja {{var}} becomes the URL the JS fetch() hits. None of these are typed in ripvec's eyes — every hop is a string. M5 anchor.
• M5 Dark-Matter Coefficient (NC16): rg-files(name) / index-files(name). On aurora, bucket_name: 11 files contain the string, 1 is in the symbol index. Ratio 11:1 = ripvec is blind to 10 of the 11 files for this name. Quantifiable polyglot blindness.
• M6 Oracles-Are-Typed (Codd/Pearl/Hickey trichotomy): every cross-language API point fits one of three shapes:
  • Codd-form: relational projection (HCL terraform_remote_state.X.outputs.Y)
  • Pearl-form: intervention/do-operator (Python sets SQL @{var})
  • Hickey-form: value/identity (tfvars strings; protobuf bytes)
  • Lensing decision: each oracle form needs a different recipe.
• NC8 Runtime-Variable-Injection Tracing: a 4-hop chain (env → tfvar → Python config dataclass → SQL template var) reproducible by grep-composition even when ripvec can't trace it symbolically. Anchor recipe for polyglot detective work.
• NC1c "not in index" IS the M5 diagnostic. When find_similar(symbol_name=X) returns "not in index" but rg finds X across the corpus, that's a polyglot dark-matter signal.
• I#54a/b polyglot edge bugs are open. HCL→HCL via terraform_remote_state and SQL FROM/JOIN edges are partial- resolved via semantic search; full LSP edge resolution is a 4.2.0+ target. M22 Polyglot Incompleteness Cost says: when bug deferral lets the dark-matter coefficient grow, productivity decays at compound rates.

§2 Working recipes (polyglot earned its keep on these)

Recipe	Trigger	Tool sequence	Caveat	Cite
NC16 Polyglot Dark-Matter Probe	"Quantify how blind ripvec is on this name"	mcp__ripvec__find_similar(symbol_name=X) → if "not in index" → cross-language mcp__ripvec__search(corpus="all", query=X) → compare hit-counts; ratio = dark-matter coefficient	High ratio (≥5) = serious polyglot blindness; pair with workaround	Part X §X.4 NC16
NC8 Runtime-Variable-Injection Trace	"How does this Python config value reach the SQL query?"	Pick variable name (e.g., athena_silver_schema); mcp__ripvec__search(corpus="all", query="athena_silver_schema") → enumerate hits across HCL/Python/SQL files → narrate the 4-hop chain by file order	Reproducible by name-grep even where ripvec can't trace symbolically; M5 use-case canonical	Part IX §IX.4 NC8
M6 Codd-form Oracle Probe (HCL remote_state)	"Where does Terraform module A read from module B's outputs?"	mcp__ripvec__search(query="terraform_remote_state", root=infrastructure/) → for each hit, the .outputs.X path IS the Codd projection	I#54a partial: search works, lsp_references returns silent empty on HCL refs (deferred 4.2.0)	M6, BUG_DATABASE §4 I#54a
M6 Pearl-form Oracle Probe (Python→SQL @{var})	"Where does this template var get bound?"	Pick @{var_name} from SQL → mcp__ripvec__search(corpus="all", query="var_name") → look for variables={var_name: ...} Python assignment	This IS Pearl's do(var_name=value) intervention on the SQL DAG	M6, Part XI §XI.4
M6 Hickey-form Oracle Probe (typed value/identity)	"Decomplect this tfvars string from its consumer types"	mcp__ripvec__search(query="bucket_name") then mcp__ripvec__lsp_references on each site to classify consumer-type (ARN vs S3-URI vs path)	M6 Hickey-form: the value travels typed-free; consumer-types diverge over time = M22 cost manifesting	M6, M22
NC7 Spec-as-Implementation-Oracle	"Does the spec doc match the impl?"

§3 Known engine gaps for this language family

Per docs/BUG_DATABASE.md (verified Cycle 11 W3 / 4.1.9).

Bug	Status	Symptom	Workaround	Cite
B-0017 / I#54a HCL terraform_remote_state edges	Open (P2)	lsp_references on HCL .outputs.X returns silent empty; semantic search works	Use mcp__ripvec__search instead of lsp_references for HCL edges	BUG_DATABASE §4
B-0018 / I#54b SQL FROM/JOIN edges	Open (P2)	SQL table references not resolved to schema definitions	Cross-language string search; the M5 dark-matter pattern	BUG_DATABASE §4
B-0019 HCL module "X" { source = "../X" } blocks	Open (P3, subsumed by I#54a)	Module composition edges missing	Workaround via search; treat as I#54a	BUG_DATABASE §4
B-0020 SQL chunker zero-symbols (per-CREATE granularity)	Partial fix	CREATE TABLE → struct now indexed per-file; per-CREATE granularity still missing	One file = one schema oracle; live with file-level granularity	BUG_DATABASE §4
B-0021 HCL locals { } blocks chunked as single	Open (P3)	Individual local-vars not addressable	Use mcp__ripvec__search(query="local.X")	BUG_DATABASE §4
B-0023 HCL locals { ... } collapsed to single chunk	Open (P3)	Same family as B-0021	Same workaround	BUG_DATABASE §4
N3 anti-pattern: string-API across languages	Open architectural	Treating untyped string passing as if typed	Apply M5 + M6 trichotomy; document each oracle's form explicitly	M5, M22

§4 Polyglot-specific BPMN — the M6 Codd-form HCL→Python→SQL trace flow

flowchart TD
  U[User: "How does the silver schema name<br/>get from Terraform to the SQL query?"] --> PV[Pick a variable name<br/>e.g. athena_silver_schema]
  PV --> SA[mcp__ripvec__search<br/>corpus=all, query=athena_silver_schema]
  SA --> CLA{Hits include .tf .py .sql .j2 ?}
  CLA -->|Yes — polyglot trace| EH[Enumerate per file kind]
  CLA -->|HCL only| HCL_only[Single-language;<br/>route to c-recipes / language-specific]
  EH --> HF[For each HCL hit:<br/>read locals{} or output{} or var{} surrounding]
  HF --> PF[For each Python hit:<br/>read read_tfvar / config dataclass / @click.option]
  PF --> SF[For each SQL hit:<br/>read CREATE TABLE / @{var} interpolation / FROM]
  SF --> JF[For each Jinja/.j2 hit:<br/>read {{var}} expansion]
  JF --> CH[Chain narration: HCL output → Python read → SQL @{var} → Jinja {{var}}]
  CH --> DM[Compute M5 dark-matter coefficient<br/>rg-files / index-files ratio]
  DM --> DMR{Ratio ≥ 5?}
  DMR -->|Yes| ALERT[**Polyglot dark matter detected**<br/>route to ripvec:drift-auditor<br/>cite M22 polyglot incompleteness cost]
  DMR -->|No| CLEAN[Trace is clean;<br/>document M6 oracle form per hop]
  HCL_only --> H6[H6 Fill-Graph Dual<br/>combine find_similar+lsp_references<br/>per c-recipes / jvm-recipes appropriate]

§5 Cross-corpus calibration

Polyglot is exercised primarily by aurora (canonical IaC polyglot, per Part X §X.4):

• aurora (~few hundred HCL+Python+SQL files) — M14 polyglot canonical. M5 dark-matter coefficient measured (bucket_name: 11:1). NC8 4-hop chain reproducible. NC7 spec-vs-impl drift quantifiable.
• Cycle 6 Wave 6 / opus/aurora finding: M22 Polyglot Incompleteness Cost — bug deferral on I#54a/b lets the dark-matter coefficient grow; aurora's bucket_name was 11:1 in Wave 5 → 11:0 in Wave 6 (the darkness deepened, not improved).
• (Cycle 12+ candidate) add a second polyglot dipole partner: candidate is a Django+PostgreSQL+templates app, or a NestJS app with protobuf/graphql.
• Diagnostic rule: if a polyglot finding reproduces on aurora AND a second polyglot corpus, it's M5/M6 architectural; if on aurora only, suspect IaC-specific (HCL chunker, SQL grammar).

§6 Heritage citations

Polyglot's earned heritage:

• Codd, E.F. A Relational Model of Data (1970) — HCL terraform_remote_state.X.outputs.Y is literally a relational projection (SELECT outputs.Y FROM remote_state WHERE name=X). M6 Codd-form anchor.
• Pearl, J. Causality (2009) — Python setting SQL @{var} is Pearl's do(var=value) operator on the SQL DAG: an intervention not an observation. M6 Pearl-form anchor.
• Hickey, R. Simple Made Easy (2011); The Value of Values (2012) — typed-free string-as-value flowing across language boundaries is complecting value with identity. M6 Hickey-form anchor; M22 incompleteness-cost roots here.
• Wittgenstein, L. Philosophical Investigations §43 (1953) — "the meaning of a word is its use in the language" — polyglot strings have meanings determined by their consumer types, not their producer. M5 dark-matter is literally Wittgensteinian.
• Brooks, F. No Silver Bullet (1986) — polyglot architecture is essential complexity (the cost of right-tool-for-the-job) AND accidental (the cost of cross-language semantic erosion); M22 quantifies the accidental cost.
• Box, G.E.P. All Models Are Wrong (1976) — the polyglot oracle trichotomy (Codd/Pearl/Hickey) is a model not a theorem; use it where it earns its keep.
• Naur, P. Programming as Theory Building (1985) — polyglot codebases require cross-language theory; the M22 cost is the theory-rot tax when ripvec can't see one of the languages involved.

---

Source: fnordpig/my-claude-plugins — distributed by TomeVault.