@onenomad/przm-cortex-pipeline-code-dossier

Code dossier pipeline. Runs a 3-pass LLM-driven architectural extraction over a checked-out repo and emits a small number of high-signal memories — a brief-typed architectural dossier plus one decision per ADR and one reference per significant entry point.

This is a deliberate alternative to @onenomad/przm-cortex-pipeline-code, which walks every file and emits hundreds of low-signal chunks suitable for code search. The two pipelines coexist; callers (MCP tools, adapters) choose between them via a mode parameter.

Why it exists

A team using Cortex to "remember a repo" doesn't usually want code search — they want to know what the repo IS. The dossier pipeline captures architecture, key modules, design decisions, and tech stack at the level a new teammate would summarize them on a wiki page. The output reads like a hand-written architectural overview because that's what the synthesis prompt asks for.

Output shape

Each run() invocation produces a PipelineMemory[]:

| Type | Count | Content | | ----------- | ----------------------- | ------------------------------------------------- | | brief | 1 | The synthesized architectural dossier (markdown). | | decision | one per docs/ADR-*.md | The ADR file verbatim. | | reference | one per entry point | The entry point file verbatim. |

When ctx.llm is not available, the pipeline gracefully degrades to a single brief memory containing the raw structural payload (no synthesis). The caller can still re-run later under a configured LLM to upgrade quality.

Inputs

import type { CodeDossierInput } from "@onenomad/przm-cortex-pipeline-code-dossier";

const input: CodeDossierInput = {
  repoPath: "/abs/path/to/checkout",
  sourceIdPrefix: "github:OneNomad-LLC/cortex",
  project: "cortex",            // optional
  tags: ["dossier"],             // optional
  sourceUrl: "https://github.com/OneNomad-LLC/cortex", // optional
};

Re-derivation gating

computeInputsSha(input) returns a stable SHA-256 over the structural payload (sorted file paths + content hashes). Callers use it to decide whether to re-run synthesis; the pipeline itself always runs when invoked.

Prompts

Prompts live as markdown files in src/prompts/. Per ADR-007 they are never inlined in code — review and tuning happen in the .md files alongside the implementation.