Experimental Flags

Wenlan's main branch often carries opt-in work before a release promotes it. A flag on main means the code path exists for testing; it does not mean normal users receive that behavior after npx setup.

When writing docs, bug reports, or benchmark notes, describe the flag, the branch or release, and whether it is enabled by default.

Recent main-branch work explores how to retrieve and assemble smaller, more useful context bundles: graph activation, temporal filtering, FTS hardening, session diversity, salience priors, episode and fact channels, PRF, k-hop graph traversal, global preludes, iterative retrieval, LLM routing, and read-time context compression.

These flags can change ranking, context composition, token cost, or evaluation labels. Do not compare results across flag sets without naming the active flags.

Other experiments target write-time and background maintenance: temporal grounding, entity near-dedup, page rewrite shrink guards, dual-pool dedup/contradiction resolution, soft eviction, and debounced reflection.

These are especially sensitive because they affect what becomes durable context. Use them on isolated development data unless a release note says the behavior is ready for normal installs.

Every flag that changes retrieval, compression, graph expansion, or maintenance can also change benchmark meaning. A result is only useful if it records the fixture, schema, model, run count, and active flags.

If a flag is eval-only or not wired into the normal daemon path yet, say so plainly. Context compression is an example of a main-branch experiment that should not be described as default chat-context behavior until release notes promote it.

Do not test experimental maintenance flags against your normal product data unless you mean to. Run a source checkout on a separate port and data directory so experiments cannot rewrite the memory store you rely on daily.

After a flag test, capture the exact command, environment, and git commit. That makes the result reproducible and keeps later docs from turning a local experiment into a product claim.