Hiro development journal

Historical GitHub discovery query and expansion boundaries

2026.09.20 at 09:11:35 America/Los_Angeles Read-only code and frozen-evidence review complete Machine-readable JSON

Executive summary

Identified the exact GitHub retrieval contract used September 15–19: one fixed public repository query, Python language qualifier, update-descending order, first page of at most 20 records, and a two-hour source interval.

All 50 frozen GitHub fetch records carry the SHA-256 of the same exact query; all 319 cycle packets carry the same source-policy hash.

The production discovery path has adaptive scheduling and cross-source evidence aggregation but no operational feedback path that expands sources or search vocabulary.

No Hiro code, policy, runtime state, or discovery run was modified or executed. Official GitHub documentation was consulted only to explain query syntax.

Work completed

Historical binding

Completed

Compared the source-policy, external-source adapter, discovery entrypoint and actionable-discovery files at repository revisions immediately before September 15 and September 20 Pacific. The relevant blobs match across the interval.
Historical revisions inspected: 3513da2084df74b3616f23f7d743740f1b41f60b and 387a4b00c38ca5ab9358a4e69c2980c7fa719b86.
Exact GitHub query: agentic AI evaluation reliability memory tools language:Python
Query SHA-256: 94d946edc2918f0182fcce0187de2a000724e803b290ac8873aac2396684d9ac. All 50 recorded GitHub fetches match; they contain 251 repository-item observations, not necessarily distinct repositories.

Retrieval eligibility

Completed

Repository search uses the supplied lexical query and language qualifier; the adapter does not search issues, pull requests, discussions, code or topic pages.
Without an in qualifier GitHub searches repository names, descriptions and topics, not README text. The adapter consumes bounded repository name, description and topics only.
No pagination, link-following, entity-follow-up search or query reformulation exists in this active adapter.
A new repository name need not be known in advance, but any returned repository must satisfy the fixed query and appear within the selected page. No retrospective assertion is made about a particular repository without its historical metadata and search-index state.

Expansion mechanisms

Completed

The policy validator requires the exact source-name set, approved endpoint URLs and exact GitHub/arXiv query strings. A different GitHub query raises GitHub discovery query is not approved.
Novelty deduplication, reservoir refill, new-evidence timing and cross-source support influence existing observations and priorities; none rewrites queries or adds sources.
The separate Phase 3C actionable-discovery tool has four preconfigured arXiv queries and can complete selected GitHub records with README text. It still uses the same fixed GitHub query; this is not autonomous vocabulary expansion.
General assistant web-search tools and research proposals do not constitute a connected autonomous source-expansion loop.

Decisions and reasoning

Separate retrieval eligibility from later relevance filtering.
Do not infer historical GitHub repository metadata or claim that not knowing a name alone made discovery impossible.
Provide no implementation changes or replay.

Validation and evidence

Check	Status	Result
Historical code and frozen-query binding	passed	Relevant boundary blobs match; 319 policy bindings and 50 GitHub query bindings agree with the inspected policy.
Hiro tests and discovery rerun	not_run	Read-only source and artifact inspection only.
Journal tests/build	passed	npm run test:hiro followed by npm run build passed; 233 entries validated, including timestamp, alias and noindex checks, with TypeScript and Vite complete.

Current state

Confirmed fixed repository-query and source-allowlist boundaries. No adaptive external search-space expansion found in the production discovery path.
Frozen packets do not preserve a complete historical GitHub search index or every upstream repository metadata record, so particular-repository historical eligibility remains conditional.

Next steps

None performed or proposed in this diagnostic task.