portfolio-optimiser-claude/docs/extending.md
Kjell Tore Guttormsen da93a68ce7 feat(portfolio): K12 — CLI parity, doc sync, knowledge-base recipe (parity row 24) [skip-docs]
The last ungated build session: the operator now drives the whole build from the
command line, and the documents claim exactly what the code does (§1).

run.py becomes the collecting entrance. Exactly one of --bundle (one project) or
--portfolio (N projects from a schema-validated reference config, with
--verdict-dir as the portfolio-level expert inbox) is required; both and neither
are refused. --goals loads a goal contract and checks it against --ledger's
realized sum BEFORE the first model call: the §8 caps bound spend, the goal bounds
achievement, so a hard target the book already meets stops the run at exit 4
without constructing a client. A soft target reached is a flag and the run
continues; an absent ledger is an empty book, so the goal is still evaluated,
never skipped. The one declared goal also drives --value-report's goal progress —
one contract, never two figures that can disagree.

The portfolio path persists nothing (K3 returns typed results; the outbox names
pairs by run_id, which a portfolio pass has none of). Rather than accept
--out/--outbox/--run-id/--value-report/--inbox/--live-dry-run there and silently
ignore them, the entrance refuses them and says why. run_portfolio is imported
lazily — portfolio.py imports this module, so a module-level import is circular.

Three seams, each detach-proven RED:
- unwire the goal check → the run proceeds and spends → red
- unwire the portfolio branch → the configured projects never run → red
- document a flag no CLI offers → the README honesty grep goes red

That last one is the doc-sync made load-bearing: the test reads README.md,
collects every --flag it documents (excluding third-party dev-tooling lines) and
asserts each exists in the --help of a CLI the README names. The drift it exists
to close was real — README claimed 562 tests, CHANGELOG claimed 265, actual 597.

Docs synced to the code: README gains an operator-CLI section and honest goal/
portfolio descriptions, CHANGELOG is rewritten to what actually shipped, and
docs/oppskrift-kunnskapsbase.md delivers D-H point 1 — the documented team
process for building a knowledge base, with the honest 1–2 week expectation and
every factory-dependent step (verdict translation, demo path) marked NOT BUILT.

597 passed · ruff clean · mypy strict clean.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MQu2xxwedckjU56byu1aUG
2026-07-25 06:42:52 +02:00

9 KiB
Raw Blame History

Extending the ingest layer

This document explains the ingest layer as implemented in this repo (D7) and how to extend it. It is scoped to what D7 actually ships: the file (CSV) and sql (SQLite) source types. The http/MCP source type is an optional extension point that D7 does not implement — see HTTP/MCP — an extension point D7 does not build.

The layer is built from the framework-neutral ingest-spec.md alone (section references below, §n, point into it). Everything here is deterministic and runs offline: ingest makes zero model calls, touches no network, and ingest.py imports nothing from the Agent SDK — it is pure standard library.

Ingest is the automated path to a bundle: it covers the sources that are already tabular. Everything a domain expert knows that no table holds is built by hand, and that process is a separate deliverable — see the knowledge-base recipe.

Where ingest sits

manifest → connector → materialization → OKF bundle (+ index) → [existing 8-step loop, unchanged]

The method spec forbids query-time retrieval against the bundle (method-spec §3 Step 1), so data reaches the model only via OKF bundles. Ingest is the deterministic step in front of the loop that makes that true: a connector reads a real source, and the extract is materialized as an OKF bundle the loop then reads exactly as it reads a hand-curated one — by navigation, never retrieval (§1§2). This is not RAG and not a live lookup inside the loop.

The manifest — the whole coupling to a source

One JSON file couples an expert to a source (§4). It is schema-validated fail-fast before any source call (ManifestContract); a malformed manifest never starts a run. Queries are configuration, not code — declarative strings the connector interprets, never evaluated as program code.

A file manifest (from examples/ingest-golden-file/manifest.json):

{
  "manifest_version": 1,
  "source": {"type": "file", "id": "prosjekt-arkiv", "root": "fixture"},
  "bundle_summary": "Two CSV extracts from a local project archive (file source type).",
  "extractions": [
    {"id": "costs", "title": "Project costs", "query": "costs.csv",
     "okf_type": "dataset", "max_rows": 10}
  ]
}

A sql manifest differs only in source; the query becomes a single read-only SELECT (from examples/ingest-golden-sql/manifest.json):

{
  "source": {"type": "sql", "id": "portefolje-db", "connection_ref": "PORTEFOLJE_SQL_DSN"},
  "extractions": [
    {"id": "costs", "title": "Project costs",
     "query": "SELECT id, item, amount, note FROM costs ORDER BY id",
     "okf_type": "dataset", "max_rows": 10}
  ]
}

source is polymorphic on type — a Pydantic discriminated union (FileSource | SqlSource). A manifest whose source.type is anything other than file or sql (including http, or an unknown tag) fails validation fail-fast; it is never silently accepted.

The two source types D7 implements

file — a local CSV catalogue

source.root is the directory extraction paths resolve against; query is a relative path to a CSV whose first row is the header. Path resolution is boundary-checked fail-closed against root (_resolve_within) — an extraction can never read outside the catalogue.

sql — a local SQLite database

source.connection_ref is the name of an environment variable whose value is the database path — the location/credential never lives in the manifest (§4, §8), so a manifest is versionable and shareable without secrets. An unset reference fails fast. The connection is opened read-only (sqlite3 URI with ?mode=ro), and SQLite executes exactly one statement per execute, so the single-statement rule is enforced by the driver. Give the query an explicit ORDER BY: golden conformance requires a stable row order (§4).

Both connectors enforce max_rows fail-fast — an extraction over its cap is an error, never a silent truncation (§8).

Materialization — what a connector's rows become

Each extraction becomes one concept file ingest-{id}.md (§5) with:

  • a provenance frontmatter layer (§7) — type, title, source_system, source_query, ingested_at, ingest_manifest, generated: true, in exactly that order. This is a separate contract from the method spec's §9 proposal provenance — the two are never mixed. generated: true plus the manifest reference is the honesty marker: a machine-made bundle is always labelled as such (§1).
  • a markdown table body — header row = column names, data rows in source order, with §5 cell escaping. Cell typing is the connector's contract: file cells are all str; sql cells are typed (int → decimal, float → shortest round-trip form, NULL → empty string, any other type → fail, never a silent coercion). See _render_sql_cell.

ingested_at is an explicit required argument to materialize — there is no wall-clock default. That is what makes golden extractions bit-deterministic. Re-materialization removes exactly the ingest-stamped files, writes the new set, and updates index.md — curated and promoted files (and their index links) always survive (§3, §5, §6).

How to add a new source type

Adding sql on top of file touched four places; a new local source type follows the same shape (all offline, no spec change if the type already exists in §4):

  1. A source model. Add a Pydantic variant (like SqlSource) with type: Literal["yourtype"], its own fields, and the shared id grammar validator. Add it to the discriminated union in ManifestContract.source.
  2. A connector. Add a _read_yourtype(...) -> list[list[str]] returning the header row followed by data rows in source order, and enforcing max_rows fail-fast. Route to it from _read_extraction on source.type. If cells can be non-str, add a _render_*_cell that types them explicitly and fails on any unsupported type.
  3. A golden case. Add examples/ingest-golden-{type}/ with manifest.json, fixture/, ingested-at.txt, and the byte-exact expected-bundle/ (§11). This is the only fasit.
  4. Load-bearing tests. Mirror the existing seam tests so each goes red when its seam is detached (§11): golden regression, provenance stamping, navigability through the unchanged okf.py, verdict-layer reservation, and re-ingest safety. Green-but-dead is the failure mode the rule exists for.

Keep it deterministic and offline: connectors are tested only against fixtures and golden extractions, and the suite must run without credentials or network (§11).

HTTP/MCP — an extension point D7 does not build

The spec names http as an optional source type (§1, §4): a remote endpoint with a base_url and an optional named credential_ref, and "an MCP-based connector is an extension of this family." Implementing it does not require a spec change, and not implementing it does not break conformance (§1).

D7 does not implement it. There is no HTTP connector and no MCP connector in this repo. A manifest naming type: "http" is rejected fail-fast at validation (proven by test_malformed_manifest_is_rejected, which parametrizes both http and an unknown discriminator). This repo ships no network egress path and no live-source integration.

If you were to build it, three honest pointers — none of which exist in D7 today:

  • The spec's http schema (§4, §5, §8). base_url must not embed credentials; credential_ref names a runtime-resolved secret; the network opt-in flag (§8) is a run argument, never a manifest field — without it the connector MUST refuse fail-fast; the response body materializes verbatim inside a fenced code block (§5).
  • The MAF sibling's demonstration (I6). The sibling implementation (open/portfolio-optimiser) demonstrates the HTTP/MCP extension point against a local mock, behind the opt-in network flag — no live source. It is referenced here as a pointer only; D7 is built from the shared spec, not from the sibling's code.
  • The SDK vehicle for an MCP connector. An MCP-based connector would be built on the Agent SDK's create_sdk_mcp_server / @tool primitives. These are not used anywhere in this repoingest.py is pure standard library. They are named here only as the mechanism you would reach for, not as something D7 wires up.

Whatever the transport, an http/MCP connector MUST honour the same manifest, materialization, gate, and golden contracts as the local connectors (§4) — and the network gate itself is a load-bearing seam (§11): the connector must refuse fail-fast without the opt-in flag.