How task-aware selection works

Fuzit's current selector is deterministic.

It does not require a generative model to decide which files are relevant.

1. Normalize task terms

The task is normalized through:

• Unicode normalization;
• camelCase splitting;
• path separator normalization;
• punctuation normalization;
• lowercase conversion;
• concept normalization.

Examples:

authentication → auth
tests → test
implementation → implementation

Authentication tasks can also add related concepts:

security
identity
jwt
session
login

2. Categorize source material

Files are classified as:

documentation
historical
implementation
test
other

Historical material receives a negative selection signal.

3. Calculate evidence

Current evidence components include:

lexical
pathTopic
exact
generated
historical
authority
git
test
dependency
manifest
doc
security

4. Apply profile weights

Each profile gives different importance to those signals.

5. Sort deterministically

Primary order:

higher score first

Tie-break:

stable repository path ordering

6. Add anchors

Tasks mentioning implementation, tests, or architecture can receive category-specific anchors so that one evidence category does not completely dominate the result.

7. Enforce the budget

Each candidate has a token estimate derived from its content size.

Selection stops when the configured token budget would be exceeded.

Candidates can also be excluded because of:

• category diversity limits;
• irrelevance;
• insufficient topical evidence;
• historical/generated penalties.

------------------------------------------------------------------------