Security and Development

Dieser Inhalt ist noch nicht in deiner Sprache verfügbar.

Conformance language (MUST/SHOULD/MAY) follows BCP 14 [RFC2119]/[RFC8174] as defined in 00-overview.md. The binding source is the FlowMCP Schemas Specification v4.2.0.

Introduction

Security and development discipline form an independent grading area with high veto affinity. Three of the four Categorical-Veto triggers defined in 08-grading-model.md live in this chapter (malicious-module, api-key-domain-mismatch, illegal-content), and the fourth (ai-security-veto) is the non-deterministic counterpart that catches what the deterministic triggers miss. This chapter defines the binding rules for each.

The checks covered here feed primarily securityScore (autonomous tier), plus formattingCompliance and the outputSchemaConformance sub-dimension “pipebarkeit”. No check defined in this chapter raises the maximum attainable grade beyond B on its own. These checks contribute to the single-test and tools-aggregate-schema areas (see the 11 Areas in 08-grading-model.md).

API-Key Hygiene (Deterministic)

The following three rules are binding for every schema. A violation of rule (3) raises a Categorical Veto with triggeredBy = api-key-domain-mismatch.

API keys MUST NOT be printed via console.log (or any equivalent stdout/stderr writer). Schemas that leak credentials at runtime are scored securityScore = fail.
API keys MUST NOT appear in comments. Hard-coded examples, “TODO remove” lines, and forgotten debug values count as comments.
requiredServerParams MUST contain only key names that belong to the API’s own domain OR to the same company. A key name from an unrelated provider is a domain-mismatch.

Example. A schema for API example.xyz declares requiredServerParams: [ "FACEBOOK_API_KEY" ]. Facebook is unrelated to example.xyz; this is a domain-mismatch. The grader MUST raise categoricalVeto.triggeredBy = api-key-domain-mismatch with evidence pointing to the offending requiredServerParams entry.

A schema that uses a generic placeholder (e.g. EXAMPLE_API_KEY) for an API hosted at example.xyz is NOT a domain-mismatch — the placeholder name aligns with the domain.

External Modules and Imports

Every imported external module is part of the schema’s attack surface. The grader runs an imports scan (deterministic) and an intent judgement (non-deterministic) over each import.

Class	Determinism	Rule
Minimalist, on-purpose	deterministic + non-deterministic	Allowed. Example: `node-fetch` for an HTTP-based schema. Score-boost on `securityScore` for explicit, on-purpose imports.
Telemetry-only / tracker	deterministic (import name) + non-deterministic (purpose)	Reduces `securityScore`. NOT a Categorical Veto by itself, but a measurable penalty.
Behaviour outside the tool’s stated purpose	deterministic (import name) + non-deterministic (behaviour)	Categorical Veto `malicious-module`. Examples: trackers without user awareness, telemetry that exfiltrates user data, malware.

The imports scanner MUST list every external module with its name, version, and source (npm, file, URL). The list is part of the evidence field of the relevant grading entry.

Binding rule. A telemetry-only module that does not exfiltrate user data is NOT a veto trigger — it is a securityScore reducer. A veto requires either the deterministic match against the malicious-module list (implementation concern) OR a non-deterministic judgement under ai-security-veto (see AI Security Veto).

AI Security Veto (Non-Deterministic)

A grader (typically an LLM grader) CAN raise categoricalVeto.triggeredBy = ai-security-veto when it detects a security finding that is not on the deterministic veto list but is well-evidenced and well-reasoned.

The ai-security-veto trigger REQUIRES both:

evidence — a pointer to the concrete artefact (code snippet, output sample, third-party report) that supports the finding.
reasoning — a free-text explanation of why the artefact constitutes a security risk.

A grading entry that sets triggeredBy = ai-security-veto without populating evidence AND reasoning is INVALID (see 08-grading-model.schema.json) and is rejected at validation time with error code VET-003 (see 08-grading-model.md).

This rule deliberately keeps the AI veto open-ended in its trigger condition but closed in its evidence and reasoning obligation. The closed list of trigger names (malicious-module, api-key-domain-mismatch, illegal-content, ai-security-veto) is NOT extensible at runtime; widening the list is a gradingSystem bump.

Error Management (Deterministic)

Schemas SHOULD use the PREFIX-NUMBER error-code pattern defined by the node-error-codes skill. The pattern requires every emitted error to carry a stable prefix (e.g. RES, SKL, GRD) and a stable numeric code (e.g. RES013).

Pattern	Score impact
PREFIX-NUMBER catalogued errors with stable codes and severity classification	Score-boost on `formattingCompliance` and `securityScore`
Generic `try { ... } catch( e ) { throw e }` re-throws without code annotation	Score reduction on `formattingCompliance`
Swallowing errors without re-throw or log	Severe reduction on `securityScore` — silent failures hide security incidents

The full error-code catalogue for the grading subsystem itself (codes GRD-*, SCO-*, VET-*) is delivered in a later stage (referenced from 08-grading-model.md).

Best-Practice Reference — Preload Pattern

The Schemas-Spec v4.2.0 defines the Preload pattern in 11-preload.md. For schemas that handle large, infrequently-updated data sets, Preload SHOULD be used.

Example. A schema fetches a 2-MB reference data file that is updated daily. The schema SHOULD use Preload to cache the file once per day rather than re-fetching it on every tool call.

Behaviour	Score impact
Correct Preload use for large daily-update data	Score-boost on `securityScore` (network attack surface) and on `outputSchemaConformance`
Repeated fetch of a large file that fits the Preload pattern	Score reduction on `securityScore`
Preload on data that changes per request (anti-pattern)	Score reduction on `outputSchemaConformance`

Preload is a recommendation; the absence of Preload is NOT a Categorical Veto.

Best-Practice Reference — Shared Lists

A group’s Domain-Knowledge document (see 10-domain-knowledge.md; this is a forward reference) MAY prescribe a Shared List that all schemas in the group MUST use instead of a provider-specific convention. The canonical example is the Shared Chain Name List in the crypto domain: schemas in the crypto group MUST emit sol (Shared List) rather than solana (CoinGecko convention) when referring to the Solana chain.

A violation of a Shared-List obligation is NOT a Categorical Veto. It is a strong score penalty on domainConformance (the dimension is group-bound, see 06-determinism-and-tier.md dimension matrix). The “forbidden conventions” section of the Domain-Knowledge document is the source of truth for which provider conventions are disallowed per group.

Formatting (Deterministic)

Schema source code MUST follow the rules defined by the node-formatting skill. The three most-cited rules are:

4-space indentation. Tabs and 2-space indentation are not conformant.
No semicolons. Statement terminators are omitted; ASI (automatic semicolon insertion) is relied upon.
if() spacing. Single space between if and its condition; condition parenthesised with spaces inside: if ( condition ).

A lint script is expected as part of the grader’s setup (the spec requires its existence; the implementation is out of scope here). The lint script’s output feeds the formattingCompliance dimension.

Pipebarkeit + Output Schema (Sub-Dimension)

jq-pipe compatibility is a sub-dimension of outputSchemaConformance with low weight. A schema’s output SHOULD be navigable by canonical jq expressions (.data.results[0].balance); a schema that emits free-form prose at the top level reduces its pipebarkeit score even though no veto is raised.

This sub-dimension is deterministic: the test runs a fixed jq expression against the schema’s output and checks whether the expression yields a value. It contributes a small score adjustment to outputSchemaConformance; the weight is set by the gradingSystem version.

Veto-Trigger Overview

The following table lists all four Categorical-Veto triggers defined by gradingSpec/2.0.0. The trigger names are identical to the enum values in 08-grading-model.schema.json properties.categoricalVeto.oneOf[1].properties.triggeredBy.enum. The grader MUST NOT introduce new names at runtime.

Trigger	Class	Source
`malicious-module`	deterministic (imports scan) + non-deterministic (behaviour judgement)	Module audit (see External Modules and Imports)
`api-key-domain-mismatch`	deterministic	API-key hygiene (see API-Key Hygiene)
`illegal-content`	non-deterministic	Content review
`ai-security-veto`	non-deterministic	AI grader reasoning (see AI Security Veto)

The data model for veto entries lives in 08-grading-model.md; this chapter does NOT redefine the veto record. The four trigger names enumerated above MUST match the enum values defined there.

Env Handling Reference

Grader tools MUST NOT read .env files directly via Read/grep/Edit or any equivalent file API. The only conformant way to consume ~/.flowmcp/.env is via the helper scripts (flowmcp dev env doctor, flowmcp dev env acquire, flowmcp dev env backup, flowmcp dev env restore, flowmcp dev env diff). The binding rationale is the never-read-env-files-with-values rule, established after an incident in which several API keys were exposed by direct file operations.

Implementers of graders MUST treat any direct .env read as a securityScore failure of the grader itself (meta-rule). This rule does NOT apply to the schema under grading — it applies to the grader tooling.

Anti-Defaults

The no-silent-defaults rule is binding for this spec chapter. Every score-boost and every score-reduction described above MUST be explicit in the gradingSystem/1.0.0 implementation:

No silent fall-through to 0 for missing dimensions (n/a is the documented placeholder — see 08-grading-model.md).
No silent ||-defaulting in scoring code (the || operator MUST NOT be used to substitute a missing value; explicit conditional validation is required).
Aging defaults (14 / 30 / 180 days, the #AGING_DEFAULTS constant) MUST be exposed as named constants in the implementation, not as inline literals.

Concrete weights, thresholds, and score-boost magnitudes belong in the gradingSystem/1.0.0 implementation; this chapter does NOT enumerate them. The binding contract is the explicitness rule, not specific numerical values.

Cross-References

07-scoring-vs-grading.md — the version namespaces (scoringSystem / gradingSystem) that bind score changes from this chapter.
08-grading-model.md — the data model of the Categorical Veto entries defined here.
10-domain-knowledge.md — Shared Lists and forbidden provider conventions (forward reference).
Schemas-Spec v4.2.0 11-preload.md — the Preload pattern (external).
node-formatting skill — formatting rules.
node-error-codes skill — PREFIX-NUMBER error-code pattern.
No-silent-defaults rule — anti-defaults rule.
Never-read-env-files-with-values rule — env handling.

Depends on: 00-overview.md, 07-scoring-vs-grading.md, 08-grading-model.md
Related: 10-domain-knowledge.md, Schemas-Spec v4.2.0 11-preload.md, node-formatting skill, node-error-codes skill