fritzprix/libragent-knowledge-bridge-2

harbor run -d fritzprix/libragent-knowledge-bridge-2

fritzprix/libragent-knowledge-bridge-2

Serial 2-task Harbor suite that validates assistant-scoped Knowledge bridging across trials.

Order	Task	Role
1	`kb-research-distill-v1`	Investigate seeded corpus → `knowledge__recordKnowledge` + manifest
2	`kb-cold-solve-v1`	Fresh session, no corpus → `knowledge__searchKnowledge` → `/app/answer.txt`

Harbor deletes sessions (and history) after each trial. Only Knowledge for the same assistant_id survives. That is the contract under test.

Run contract

Dedicated or stable Harbor assistant with knowledge (+ history optional).
Wipe Knowledge for that assistant before the suite.
n_concurrent_trials = 1 — research must finish before cold-solve.
Same assistant_id for both tasks.

# Preferred entrypoint (wipe + serial path runs)
pnpm bench:knowledge-bridge

# Dry-run
pnpm bench:knowledge-bridge:dry

# Skip wipe (debug only)
LIBRAGENT_KNOWLEDGE_BRIDGE_SKIP_WIPE=1 pnpm bench:knowledge-bridge

Seed facts

See SEED_SPEC.md.

Negative control (optional)

Wipe Knowledge, then run only kb-cold-solve-v1. Expect reward 0 (proves the solve path cannot cheat from /app/noise/).

Hub publish

Tasks are local packages. After harbor sync / publish, update digests in dataset.toml and point pnpm bench:knowledge-bridge at the Hub tag if desired.