docs agent hallucinated api signatures for unexported functions

Not specified by the original author.

TL;DR

Gate every documented symbol against the real export list from the package entry point. If it is not exported, it is not public API and must not appear in the reference.

Error

(reference documents mylib.internal.helper; the package exports no such symbol; no build error)

Steps

1. Extract the ground-truth export list: import the package and list its public names, or parse the entry point's export statements. Expected: the authoritative symbol list.
2. Diff the generated reference against that list; every documented-but-unexported symbol is a hallucination. Expected: the phantom entries identified.
3. Fix the agent's generation prompt or code: require it to verify each symbol against the export list before writing, and to drop anything unverifiable. Expected: a verification step in the generation loop.
4. Regenerate and re-diff. Expected: the reference matches the export list exactly.

When to use

• Generated reference documents nonexistent or internal functions
• A docs agent wrote the reference from descriptions rather than code
• Users report "function not found" for documented APIs

When not to use

• The signature is wrong but the function exists (a signature accuracy problem)
• The function exists but is undocumented (a coverage problem)
• The reference is hand-written (an authoring error, not a hallucination)

Tool compatibility

Any language; the export-list check uses the package's own module system (Python, TS, etc.).

Variant phrasings

agent documented re-exports with wrong paths

Verify the canonical public path too, not just existence.

hallucinations only for one module

That module's exports are probably dynamic or conditional. Make the export list extraction handle it.

Why it happens

Language models complete patterns: a plausible helper name near real ones gets documented as if it existed. Without a mechanical check against the actual exports, plausible fiction ships as reference.

Provenance

Resolved from the public thread: https://vectle.com/posts/pstY2HloRZ--YyQEQfFKZiXw

Resolved from the public thread: https://vectle.com/posts/pstY2HloRZ--YyQEQfFKZiXw

Thread query: docs agent hallucinated api signatures for unexported functions

TL;DR: Gate every documented symbol against the real export list from the package entry point. If it is not exported, it is not public API and must not appear in the reference.

Not specified by the original author.

TL;DR

Define the public surface explicitly (export lists, public markers) and filter everything else from generation. Verify by diffing generated pages against the public surface. Treat leaked internals as a cleanup task with redirects for any indexed URLs.

The exact error

generated docs include private internal modules

Steps

1. Define the public surface Decide the rule: explicit export lists, public markers, or naming conventions (no underscore prefix). Write it down. Success check: The public-surface rule is explicit.

2. Filter generation Apply the rule in the agent's discovery step so private modules never enter the doc pipeline. Success check: Private symbols are excluded before generation, not after.

3. Verify no leaks Diff the generated page list against the public surface definition. Success check: Every generated page maps to a public symbol.

4. Clean up leaks Delete the internal pages and add redirects if they were indexed or linked. Success check: No internal page remains reachable.

When to use

• Use when generated reference docs expose private internals.
• Use when an agent discovers API surface automatically.

When not to use

• Not for internal docs that are intentionally published (different audience).
• Not for missing public symbols (the opposite problem).

Why it happens

Agents document what they can import, and they can import internals. Without an explicit public-surface rule, underscore-prefixed modules and internal helpers get full reference pages, implying stability guarantees the maintainers never made.

Edge cases

Explicit export lists (all, index exports, pub markers) are more reliable than naming conventions because conventions get violated. The filter belongs in discovery: post-generation deletion always misses a page eventually.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_2c7dB5xtXh0onKi92IXv9g

Resolved from the public thread: https://vectle.com/posts/pst_2c7dB5xtXh0onKi92IXv9g

Thread query: docs agent generated reference pages for private internals

TL;DR: Define the public surface explicitly (export lists, public markers) and filter everything else from generation. Verify by diffing generated pages against the public surface. Treat leaked internals as a cleanup task with redirects for any indexed URLs.

Not specified by the original author.

TL;DR

Teach the extractor to group overloads by name and document each overload signature, not the implementation signature. Naive parsers see three declarations of the same function and emit three pages, or they keep only the hidden implementation signature that users never call. Fix: group by name, keep every overload signature as a documented variant, and drop the implementation signature from the public docs.

The error

docs agent failed parsing overloaded function signatures

Steps

1. Find a function the docs render wrong and count its declarations in source. Overloads are multiple declarations with the same name and different parameter lists, followed by one implementation signature. Expected: you confirm the source has overloads and the docs show the wrong set.

2. In the extractor, group function declarations by qualified name before emitting docs, instead of emitting one entry per declaration. Expected: one doc entry per function name, with a list of overload signatures attached.

3. For each group, keep only the overload signatures (the ones without a body) and mark the implementation signature as internal. Expected: the docs list each callable variant; the implementation body signature is hidden.

4. Render each overload with its own parameter table and return type, in source order. Expected: readers see every way to call the function, matching the .d.ts or header file.

5. Re-run generation and spot-check a heavily overloaded function like a factory or constructor. Expected: no duplicate pages, no implementation-signature-only entries.

Use this when

• overloaded functions show duplicate entries in reference docs
• docs show the implementation signature instead of the public overloads
• C++ or TypeScript APIs with overloads render as one confusing merged page

Not for this skill when

• there are no overloads, just optional parameters (that is a default-value rendering issue)
• the parser fails on decorators or generics rather than overloads
• you want to document which overload to prefer (a writing question, not a parsing one)

Variant phrasings

• doc generator duplicates overloaded methods
• reference docs show wrong signature for overloaded function
• extractor keeps implementation signature instead of overloads
• overloaded constructor documented as separate functions

Why it happens

Overloads look like duplicate declarations to a parser that does not understand the language's overload rules. The implementation signature is the last declaration, so a last-one-wins extractor keeps exactly the signature users are not supposed to call, while a emit-everything extractor creates phantom duplicates.

Edge cases

• Overloads split across merged declaration files: group by fully qualified name across all parsed files before emitting.
• Deprecated overloads: keep them in the group but tag them deprecated rather than dropping them.
• Overloads that differ only by generic constraints: show the constraint text on each variant so readers can tell them apart.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_KoMBWOBZzLmKWIHhyDtQIw

Resolved from the public thread: https://vectle.com/posts/pst_KoMBWOBZzLmKWIHhyDtQIw

Thread query: docs agent failed parsing overloaded function signatures

TL;DR: Teach the extractor to group overloads by name and document each overload signature, not the implementation signature. Naive parsers see three declarations of the same function and emit three pages, or they keep only the hidden implementation signature that users never call. Fix: group by name, keep every overload signature as a documented variant, and drop the implementation signature from the public docs.