ci/woodpecker/push/woodpecker Pipeline failed

Details

docs: add 8 architecture decision records

ADR 0001: nvim-treesitter highlights.scm as capture source
ADR 0002: language-agnostic method extraction via child_by_field_name
ADR 0003: method-level n-gram clustering before inference
ADR 0004: frequency filter with min_coverage threshold
ADR 0005: import extraction per cluster
ADR 0006: argument pattern extraction via AST node classification
ADR 0007: JSON output for LLM prompt injection
ADR 0008: BEX ensemble for grammar inference

2026-07-03 22:01:35 +02:00

2.2 KiB

Raw Blame History

7. JSON output for LLM prompt injection

Date: 2026-07-03

Status: Accepted

Context

The text table output is human-readable but not directly usable by an LLM. To use Dervish conventions in another agent or coding session, the output must be parsed, reformatted, and injected into a prompt — an extra friction step.

An LLM consuming conventions needs:

Structured data it can read directly (no parsing).
All metadata per convention (grammar, imports, args, files, packages).
Compact enough to fit in context without overflow.

Decision

Add a --json flag that outputs a structured JSON array instead of the text table.

JSON structure:

[{
  "language": ".kt",
  "conventions": [{
    "label": "every → assertEquals → verify",
    "method_count": 16,
    "algorithm": "CRX",
    "grammar": "every+.assertEquals.verify+.any?",
    "mdl_score": 8.64,
    "imports": ["import io.mockk.every", "..."],
    "packages": ["eu/corentic/springrag/agent/capability"],
    "arg_patterns": {
      "assertEquals": {
        "occurrences": 42,
        "arg_count": {"min": 2, "max": 3, "common": 2},
        "patterns": [{"count": 30, "args": 2, "types": ["lit", "var"]}]
      }
    }
  }],
  "total_methods": 665
}]

Also accepts --format json and --format text for explicit control.

Consequences

Positive:

LLM consumes the JSON directly — no parsing step needed.
All metadata in one object per convention — imports, args, files, packages all together.
--json is a single flag — the default text output remains for human review.

Negative:

JSON is more verbose than text (full import list instead of truncated preview).
No easy way to limit output size — a large codebase produces JSON that may overflow context.
- Mitigation: --include flag filters files before analysis.

Alternatives Considered

YAML output: More readable, but less universally parseable by LLMs.
CSV output: Too flat for nested data (arg_patterns, imports list).
Custom prompt template: Would need per-framework templates. JSON is framework-agnostic.
No structured output: User must pipe through jq or manual reformatting. Bad UX.

2.2 KiB Raw Blame History