5.9. Developer source questions

Agent Master can answer implementation questions such as “Where is _dispatch_code implemented in agent_master?” or “Show the unit tests for test_exact_edit_bypasses_extraction_and_retrieval”. These use a separate, read-only developer search. Ordinary science questions retain the existing science RAG corpus and ranking.

Developer answers include source paths and line numbers. Offline mode shows the matching excerpts; a configured model can synthesize an explanation from the same evidence. Tests are included only when explicitly requested and are labelled as expected-behavior examples, not public API documentation. Missing evidence produces a limitation rather than a guessed API.

The search reads the current checkout on every request, so the next search discovers saved source edits without an index rebuild. Each evidence record includes a file SHA-256, read timestamp, checkout revision and package-version information. A checkout/runtime mismatch is reported. A Git revision alone does not describe uncommitted files; the per-file hashes identify the evidence actually read. Concurrent edits can be flagged as changes during search.

5.9.1. Search boundaries

Only pycsamt, docs/source, docs/examples, examples and assistant_recipes under the checkout are eligible. Source text is inspected statically: modules are never imported or executed by these tools. Environment files, sensitive filenames, hidden directories, caches, generated assets and unrelated user files are excluded. Embedded credential assignments are omitted from excerpts. Retrieved text is reference data and cannot authorize actions.

Search is lexical and bounded to 6000 candidate files, 32 MB of inspected text and a ten-second scan budget; file enumeration has a five-second budget. Individual files must be at most 1 MB. At most six results are returned, and direct source reads are limited to 80 lines / 4000 characters. A truncated search reports its limit; narrow the question to a symbol or path when needed. These bounds can omit relevant evidence, and absence is not proof that an implementation does not exist. A model-generated explanation still needs review and does not imply execution or successful tests.

5.9.2. Python interface

from pycsamt.assistant.tools.repository import RepositoryTools

repo = RepositoryTools()
result = repo.search("Where is _dispatch_code implemented?")
signature = repo.inspect_symbol(
    "pycsamt/app/agent_master/callbacks/chat.py", "_dispatch_code"
)
excerpt = repo.read_source(
    "pycsamt/agents/_generation.py", start=1, lines=30
)

PackageQAAgent.execute also accepts scope="developer" to explicitly select this scope. use_rag=False disables developer source access as well as science retrieval. Repository questions do not edit source files.