Harborworks manifest-composition experiment

Date: 2026-07-13
Status: Deterministic mechanics and a full-union materialization run are complete; manifest contract and named-harness evaluation remain open.

Question

Can multiple context repositories be presented as one navigable surface without destructive flattening, and does repository-level scope filtering add useful selection power before artifact-level harness discovery?

Scope

It does not test an LLM, prompt token use, semantic retrieval, malicious instructions, real Git resolution, signatures, or authority-system exports.

Inputs and provenance

The nine Harborworks repositories are deliberately synthetic investigation inputs created on 2026-07-13. They do not represent an external organization or authoritative domain knowledge.

The manifests use JSON-compatible YAML and simulated revision strings. This avoids third-party parser and Git-network dependencies; it is not a manifest format decision.

Raw catalogs and materialized overlays remain local under .cache/results/harborworks-composition/run-001/.

The Run 002 harness surface is generated locally under .cache/harborworks-operations-workbench/.

Fixture topology

The corpus contains one always-available core, three function contexts, four domain contexts, and one composing workbench:

Every repository exports two knowledge files, six shallow skills, and one dependency-free CLI. Each skill instructs the consumer to read its owning knowledge_base/overview.md and invoke its src CLI. Domain facts are not embedded in skill instructions.

The resolved workbench graph has 9 repositories and 27 unique import edges. Repeated transitive imports exercise diamond resolution. Five repository- relative paths collide intentionally:

knowledge_base/glossary.md across 9 repositories;

knowledge_base/overview.md across 9 repositories;

skills/summarize-status/SKILL.md across 5 repositories;

src/cli.py across 5 repositories;

src/cli.mjs across 4 repositories.

Method

The scratch implementation at .cache/src/harborworks-composition/compose.py:

recursively resolves local imports and verifies declared identities and revisions;

applies the root capability ceiling and fails on missing required capabilities;

enumerates only declared export directories;

assigns each artifact an identity of (context id, revision, repository-relative path) plus a SHA-256 digest;

constructs a logical catalog and a deliberately naive physical overlay;

compares full, repository-filtered, and artifact-filtered surfaces; and

invokes every Python and Node CLI as a smoke test.

The filter query selected two functions and three domains:

{
  "functions": ["analysis", "briefing"],
  "domains": [
    "software-delivery",
    "data-governance",
    "customer-operations"
  ]
}

Findings

Destructive flattening loses information in this corpus. The logical catalog retained 81 identities. A naive relative-path overlay produced 54 files after 27 overwrites.

A flat experience does not require flat identity. Namespaced catalog identities retained all colliding artifacts together with revision, digest, scope, policy, and source path.

Repository prefiltering helped but was not sufficient. It excluded the remediation function and facilities domain repositories, reducing 54 skills to 42. Six remediation skills still remained inside the selected domain repositories; artifact filtering reduced the set to the 36-skill golden result.

Scope dimensions need explicit applicability semantics. Function contexts required a domain wildcard, while core and workbench contexts required an always-available marker. A simple list intersection is not expressive enough without those concepts.

Conflicts should be visible. The revision and policy scenarios both failed closed with specific diagnostics.

The tested order property is narrow. Reversing import declaration order for the same root, graph, revisions, and policies produced the same 72 imported artifact identities. This does not prove that changing the root or resolving version ranges is order-independent.

A functional physical handoff is possible for the fixture. All 81 artifacts materialized without destination collisions; 54 skills were discoverable through one root and all rewritten links and CLI paths worked.

Relocation is not free. The materializer applied 270 controlled rewrite occurrences across 54 skills. Unknown relative references remain a compatibility problem rather than something the experiment solved.

Interpretation

The evidence supports a namespace-preserving logical union as the semantic model. Physical trees may still be generated for filesystem-only consumers, but they need namespace-aware paths or an explicit resolution ledger and must not define composition semantics.

Repository scopes look useful as optional coarse discovery hints. They should not replace artifact metadata or harness-level selection: a repository is a governance and release envelope, not necessarily a homogeneous retrieval unit.

The experiment also suggests retaining content digests alongside the proposed identity tuple. A digest exposes mutation under a reused revision label even though real immutable Git revisions would normally prevent it.

Limitations and next work

Simulated revisions are not Git commit resolution or a lockfile.

The root-only capability check is intentionally shallow; transitive policy intersection and artifact-class policies remain unspecified.

Directory exports need trials with include/exclude patterns and generated indexes.

Deterministic label filtering is not evidence of LLM discovery accuracy, task success, or prompt savings.

No poisoning, authorization-bypass, or provenance-in-answer benchmark has been run.

Run 002 used a harness-neutral filesystem discovery simulation, not a named agentic harness.

The scratch composer should remain in .cache/src/. The next decision is whether to extend the experiment with a harness/model selection trial or first settle the identity, export, locking, and policy semantics. Only after those semantics are accepted should a durable specification be written and code be promoted into committed src/.