---
id: ATLAS-WP-0003
type: workplan
title: "Discovery connectors"
domain: infotech
repo: config-atlas
status: finished
owner: codex
topic_slug: custodian
created: "2026-06-26"
updated: "2026-06-27"
state_hub_workstream_id: "e4400d9c-021a-4e44-8e9b-719f94a9561a"
---

# Discovery connectors

Grow the configuration-surface registry from **automated discovery** instead of
only hand authoring — by reusing `repo-scoping`'s scanner → candidate → approval
machinery rather than building bespoke scanners.

This is **Phase 2** of [`specs/ArchitectureBlueprint.md`](../specs/ArchitectureBlueprint.md)
§6 and realizes PRD FR-8 ([`specs/ProductRequirementsDocument.md`](../specs/ProductRequirementsDocument.md)).
Connectors are **read-only and stateless**: they emit *candidate* surface entries
for human/agent PR review and never write any source system or auto-merge
([`docs/ecosystem-boundaries.md`](../docs/ecosystem-boundaries.md) §2.4).

**Depends on:** ATLAS-WP-0002 (surface schema, registry layout, CI gate). The
candidate → approval → registry-truth state machine and source-linked evidence
model are reused from `repo-scoping`, not reinvented.

**Exit condition:** at least one connector produces candidate surfaces that enter
via PR, validate against the schema, and can be promoted to entries; stale/unowned
surfaces are reported. No connector mutates a source system.

**Sequencing:** T01 (contract) first; T02–T04 (connectors) depend on T01 and may
proceed in parallel; T05 (stale/unowned) depends on having connector-produced data.

## Define the connector contract and candidate workflow

```task
id: ATLAS-WP-0003-T01
status: done
priority: high
state_hub_task_id: "e7b03e49-7e49-4629-ada1-facdf596569b"
```

Result 2026-06-27: Added `tools/connector_base.py` (validates + writes candidates,
refuses to overwrite promoted entries, never stores values) and the contract in
`docs/discovery-connectors.md` (read-only/stateless, candidate->PR->promote,
`status: candidate` + provenance). Added `candidate` to the schema status enum;
candidates/ is gitignored and excluded from the gate.

Specify the read-only connector contract and the candidate lifecycle. Define the
candidate entry format (a surface entry with `status: candidate` + provenance) and
its location (`registry/surfaces/candidates/`), and the `connector → candidate YAML
→ PR → validate → promote/merge` flow. Reuse `repo-scoping`'s
`Scope→…→Evidence→Fact` provenance and candidate→approval model
(`~/repo-scoping/docs/scope-md-spec.md`, `src/repo_scoping/candidate_graph/`).
Document in `docs/discovery-connectors.md`.

- **Acceptance:** a documented contract (inputs, candidate output shape, promotion
  rules); connectors are specified as stateless and non-mutating; the candidate
  schema reuses the surface-entry schema with a `candidate` status + provenance.

## repo-scoping fact ingestion connector

```task
id: ATLAS-WP-0003-T02
status: done
priority: high
state_hub_task_id: "2447547b-1776-4225-af4f-f73680ccb2df"
```

Result 2026-06-27: Added `tools/connector_reposcoping.py` (+ make
connect-reposcoping). Consumes repo-scoping facts (--facts file or
REPO_SCOPING_URL), filters config facts, emits schema-valid candidates; degrades
gracefully when the API is down. Verified on synthetic facts (2 config candidates,
non-config skipped).

Build the connector that consumes `repo-scoping` observed facts/evidence as input
and emits candidate configuration surfaces, adding only the config-kind and layer
classification on top (ecosystem-boundaries §2.4 option a). Map repo-scoping facts
about config files/env/params to `surface.*` candidates with `kind`, `scope`, and
`sources[]` populated.

- **Acceptance:** running the connector against a registered repo's repo-scoping
  facts emits valid candidate surfaces (schema-valid) with source links; zero writes
  to repo-scoping or the scanned repo.

## git-config deterministic scanner

```task
id: ATLAS-WP-0003-T03
status: done
priority: medium
state_hub_task_id: "ddfb8eaf-46b4-4b15-9719-b167538c15fb"
```

Result 2026-06-27: Added `tools/connector_gitconfig.py` (+ make connect-gitconfig).
Deterministic scan for *.env.example / values*.yaml / config*.yaml; real .env ->
secret-ref (no value read). Verified on ~/state-hub: 4 real candidates including a
Helm values.yaml and a secret-ref .env.

Build a deterministic scanner over repository config surfaces — env files, YAML/TOML
config, Kubernetes ConfigMap/Secret *references*, and Helm `values*.yaml` overlays —
emitting candidate entries with inferred `kind` and layer `role` per source. Secret
references become `secret-ref` candidates (reference only, never values).

- **Acceptance:** scanning a sample repo emits candidates classified by kind, with
  Helm overlays mapped to layer roles; secret references carry no values.

## feature-control flag connector

```task
id: ATLAS-WP-0003-T04
status: done
priority: medium
state_hub_task_id: "9e2f5893-7b98-4ca6-89d7-94d093d6bd4b"
```

Result 2026-06-27: Added `tools/connector_featurecontrol.py` (+ make
connect-featurecontrol). Emits `feature-flag` surfaces linking to feature-control
keys (role: feature-control-key, openfeature endpoint) with no eval logic (FR-12);
degrades when no key registry exists. Verified on synthetic keys.

Build a connector that inventories `feature-control` keys and emits `feature-flag`
surfaces that **link** to the feature-control key (`sources[].role:
feature-control-key`) and contain no evaluation logic (PRD FR-12; reinforces the
delegation boundary). Surface stale flags as a signal.

- **Acceptance:** feature-control keys appear as `feature-flag` candidate surfaces
  linking to the authoritative key; config-atlas holds no flag-evaluation logic.

## Stale and unowned surface detection

```task
id: ATLAS-WP-0003-T05
status: done
priority: medium
state_hub_task_id: "ddcf070c-a863-47df-8c99-61c1980a8d18"
```

Result 2026-06-27: Added `tools/registry_health.py` (+ make registry-health).
Reports unowned (missing/unresolvable owner vs reuse-surface roster as domain-tree
stand-in) and stale (evidence.last_seen) surfaces. Verified: 4 promoted surfaces,
all owned and fresh.

Add a report that flags surfaces with no resolvable `owner` (against domain-tree)
and surfaces whose sources were not seen in the latest scan (stale/drift signal),
using `evidence.last_seen`. Wire it into the validation tooling (`tools/`).

- **Acceptance:** the report lists unowned and stale surfaces; an unowned or
  unseen-since-N-days surface is surfaced for review.

---

After workplan or registry file updates, sync from `~/state-hub`:

```bash
make fix-consistency REPO=config-atlas
```
</content>