Compaction suite v0.5: cascade + mutchain (eligibility) and codesweep (quality). Anti-grep pressure via large sequential write windows.

Stratified 9-task smoke suite from Terminal-Bench 2.1 for LibrAgent harness eval. Difficulty mix: 2 easy / 4 medium / 3 hard across 7 categories (SE, debugging, sysadmin, security, data-science, scientific-computing, data-querying). Disjoint from NovitaAI/tb21-file-recovery. See README for sampling rationale and per-task notes.

10-task LibrAgent harness: 5 Terminal-Bench vision/OCR/video + 2 Harbor Index (batch OCR + GAIA chess exact-match) + 3 custom web visual-bug tasks (browser__). All verifiers are deterministic (pytest / exact string) — no JUDGE_MODELS / LLM-as-judge. Not a full TB leaderboard.

2-task serial suite: research+distill into Knowledge, then cold-session solve via searchKnowledge. Validates assistant-scoped knowledge bridging across Harbor trials (history does not survive trials).

12-task LibrAgent browser dual-mode harness: 6 scenarios × sidecar|userChrome. In-container static server published to host. userChrome tasks require Extension Bridge Connected (skip when absent). Deterministic pytest verifiers.

7-task LibrAgent suite for explicit Org teamwork: scaffold/createOrg/spawn/checkSession across hub-spoke, sequential, swarm, rework, and bootstrap-trap missions. Verifies ORG_HARNESS_REPORT.json + per-task product evidence.

LibrAgent iteration suite: 8 easy SWE-bench Verified tasks (one per repo, <15 min fix). Not a Verified solve-rate estimate.

Compaction suite v0.5: cascade + mutchain (eligibility) and codesweep (quality). Anti-grep pressure via large sequential write windows.

Stratified 9-task smoke suite from Terminal-Bench 2.1 for LibrAgent harness eval. Difficulty mix: 2 easy / 4 medium / 3 hard across 7 categories (SE, debugging, sysadmin, security, data-science, scientific-computing, data-querying). Disjoint from NovitaAI/tb21-file-recovery. See README for sampling rationale and per-task notes.

10-task LibrAgent harness: 5 Terminal-Bench vision/OCR/video + 2 Harbor Index (batch OCR + GAIA chess exact-match) + 3 custom web visual-bug tasks (browser__). All verifiers are deterministic (pytest / exact string) — no JUDGE_MODELS / LLM-as-judge. Not a full TB leaderboard.

2-task serial suite: research+distill into Knowledge, then cold-session solve via searchKnowledge. Validates assistant-scoped knowledge bridging across Harbor trials (history does not survive trials).

12-task LibrAgent browser dual-mode harness: 6 scenarios × sidecar|userChrome. In-container static server published to host. userChrome tasks require Extension Bridge Connected (skip when absent). Deterministic pytest verifiers.

7-task LibrAgent suite for explicit Org teamwork: scaffold/createOrg/spawn/checkSession across hub-spoke, sequential, swarm, rework, and bootstrap-trap missions. Verifies ORG_HARNESS_REPORT.json + per-task product evidence.

LibrAgent iteration suite: 8 easy SWE-bench Verified tasks (one per repo, <15 min fix). Not a Verified solve-rate estimate.