Works
Only older checks are saved; no current product test.
Answers apply to each recorded setup, revision and operation. Earlier observations remain in Run details.
Tested on Linux| Platform | Haiku | Luna |
|---|
| Linux | yes | |
|---|
Linux / Haiku 5.5: yes · Oct 17, 2026 · f3fca73aff75 · claude-code / 2.1.300 (Claude Code)
Recorded operation: node bin/medsci-skills.js list
- Needs keys or browser
- None seen
- Setup time
- 6 min 7 s to install, 3 s to first use
- Issue
- none found in recorded check
Safety: no observations were recorded for the selected revision. Functional results do not establish a safety result.
Repair history and test corrections
Current diagnosis: unknown · resolvedKnown concurrently creating Incus VM may temporarily have config=null; admission now waits for its exact ownership tags instead of raising AttributeError. Skill installer remains partial setup; configured-client workflow is outside this smoke.
Repair record fix-0ccfddf00189934f31c7 · current status JSON. Original run outcomes remain unchanged.
Failed run: Aperivue/medsci-skills · Oct 16, 2026 · Our test failed; no repo verdict.
Repair development: Aperivue/medsci-skills · Oct 16, 2026 · Repair work (development) · not an initial-trial result.
Test-repair rerun: Aperivue/medsci-skills · Oct 16, 2026 · Not tested.
- Evidence unavailable in the public snapshot.
Current diagnosis: unknown · resolvedAtomic destroying+revoked cleanup reservations and removed-snapshot reconciliation prevent a concurrent teardown becoming a fresh attempt failure. Known creating config=null waits for exact ownership fields; resource waits defer before measurement. Streamed disk-derived source intake and pinned JSONL inventory replace incidental source/member/stdout ceilings.
Repair record fix-918887ea9513eadbccc3 · current status JSON. Original run outcomes remain unchanged.
Failed run: Aperivue/medsci-skills · Oct 16, 2026 · Our test failed; no repo verdict.
Test-repair rerun: Aperivue/medsci-skills · Oct 16, 2026 · Not tested.
Current diagnosis: unknown · resolvedSource archive and original-copy/JSONL manifest now live on VM disk /var/tmp, not the187484KiB /run tmpfs. Streamed source digest/copy/inventory use actual disk capacity and original remaining attempt deadline. Expired exact-owned teardown now queues new capacity wait; no fresh-target lease failure. Shared resource wait defers before native/model calls.
Repair record fix-c50581db961a9df0aeba · current status JSON. Original run outcomes remain unchanged.
Failed run: Aperivue/medsci-skills · Oct 16, 2026 · Our test failed; no repo verdict.
Test-repair rerun: Aperivue/medsci-skills · Oct 16, 2026 · Not tested.
Current diagnosis: unknown · resolvedColdpass condensed instruction text introduced newline controls rejected by the existing job schema before any model call. Job construction now normalizes instructional whitespace once, including repair context; exact original failure remains immutable. Default/repair schema regression passes.
Repair record fix-f5b823099da63d0e46c4 · current status JSON. Original run outcomes remain unchanged.
Failed run: Aperivue/medsci-skills · Oct 16, 2026 · Our test failed; no repo verdict.
Test-repair rerun: Aperivue/medsci-skills · Oct 16, 2026 · Not tested.
Run details
Haiku 5.5
Native agent · Checked operation passedChecked operation node bin/medsci-skills.js list
Agent-reported scope: node bin/medsci-skills.js list (documented 'medsci-skills list' CLI command, run from the checked-out repo); lists all 54 bundled skills grouped by category
Run details
- run
- Aperivue--medsci-skills-claude-haiku-5-5-1792220074087227142
- date
- Oct 17, 2026
- time to working
- 397.7s
- product hosting / execution
- Local command in the guest
- recorded operation scope
- product operation
- model
- claude-haiku-5-5
- agent / version
- claude-code / 2.1.300 (Claude Code)
- platform
- linux
- reported cost
- $0.007508
- Recorded source
- https://github.com/Aperivue/medsci-skills
- Tool transport
- mcp-assigned-vm
- Test version
- readme-d3659cd3f9210dfa5802d1e6a3b243d7ad5e3fc7dc58d619be334de899f21e50
- revision
- f3fca73aff7538038acf0e8c74aa164a5e3651a4
Whose fault: not assigned · resolved What the agent did
node bin/medsci-skills.js list (documented 'medsci-skills list' CLI command, run from the checked-out repo); lists all 54 bundled skills grouped by category
Read README.md (English) Installation section: requires Node 18+ and Python 3.9+
Found node missing; installed via apt-get: nodejs and npm (node v18.19.1, npm 9.2.0)
python3 3.12.3 already present
Ran node bin/medsci-skills.js --help (exit 0)
Ran node bin/medsci-skills.js list (exit 0, 54 skills in 9 categories)
Wrote .coldpass/probe.json
run JSON · checker evidence · agent report + check · transcript excerpt · error excerpt
Earlier attempts (1)
Haiku 5.5
Native agent · Not testedChecked operation python3 -I /workspace/skills/check-reporting/scripts/prisma_cascade_check.py --round1 /tmp/medsci_example/2_Screening/round1.tsv --round2 /tmp/medsci_example/2_Screening/round2.tsv --round3 /tmp/medsci_example/2_Screening/round3_adjudication.tsv --out /tmp/medsci_probe_out/prisma_cascade.json
Agent-reported scope: Installed MedSci Skills with the README's documented installer (node bin/medsci-skills.js install, the same entry point as npx medsci-skills install), then ran the bundled offline PRISMA cascade check (skills/check-reporting/scripts/prisma_cascade_check.py) on example round1/round2/round3 screening TSVs. It computed the flow counts (10 -> 6 -> 4 -> 3/1) and wrote qc JSON with submission_safe true.
Run details
- run
- Aperivue--medsci-skills-claude-haiku-5-5-1792198335156641103
- date
- Oct 16, 2026
- time to working
- 448.4s
- product hosting / execution
- Local command in the guest
- recorded operation scope
- Covers the Linux terminal install (62/62 skill folders verified in ~/.claude/skills and ~/.agents/skills) and one deterministic stdlib-only script. Not exercised: the agent-side skill workflows (/check-reporting, /orchestrate etc.), which need a running Claude Code/Codex client; the manuscript drift check (no --manuscript supplied); the optional Word/PDF/figure tooling reported as missing by the installer (doctor.py); and the pymupdf4llm/pandas-based scripts. No MCP server is shipped, so no tools/list applies.
- model
- claude-haiku-5-5
- agent / version
- claude-code / 2.1.300 (Claude Code)
- platform
- linux
- reported cost
- $0.013257
- Recorded source
- https://github.com/Aperivue/medsci-skills
- Tool transport
- mcp-assigned-vm
- Test version
- readme-ffc8770b4256cb8b179f7c65f1bbba89cc85d6191e2ebdfe5d161999e983eb8b
- revision
- e6f0589de3863ddabcbeadd1773f18e44452f268
Current diagnosis: unknown · resolvedColdpass condensed instruction text introduced newline controls rejected by the existing job schema before any model call. Job construction now normalizes instructional whitespace once, including repair context; exact original failure remains immutable. Default/repair schema regression passes.
Repair record fix-f5b823099da63d0e46c4 · current status JSON. Original run outcomes remain unchanged.
Failed run: Aperivue/medsci-skills · Oct 16, 2026 · Our test failed; no repo verdict.
Test-repair rerun: Aperivue/medsci-skills · Oct 16, 2026 · Not tested.
Double-checks failed run: Aperivue/medsci-skills · Oct 16, 2026.
Fix: fix-f5b823099da63d0e46c4.
What the agent did
Installed MedSci Skills with the README's documented installer (node bin/medsci-skills.js install, the same entry point as npx medsci-skills install), then ran the bundled offline PRISMA cascade check (skills/check-reporting/scripts/prisma_cascade_check.py) on example round1/round2/round3 screening TSVs. It computed the flow counts (10 -> 6 -> 4 -> 3/1) and wrote qc JSON with submission_safe true.
Read README.md Installation: 'In a terminal' section (Node 18+ and Python 3.9+)
Installed prerequisites with apt: nodejs 18.19.1, npm 9.2.0, git (Ubuntu 24.04); Python 3.12.3 already present
Installation section 'npx medsci-skills install' via local pinned package: node bin/medsci-skills.js install (rc=0)
Ran the cascade operation: python3 -I skills/check-reporting/scripts/prisma_cascade_check.py with example TSVs (rc=0, submission_safe true)
run JSON · checker evidence · agent report unavailable · transcript excerpt unavailable · error excerpt unavailable
No product verdict. Run record.
No product verdict. Run record.
No product verdict. Run record.
No product verdict. Run record.
No product verdict. Run record.
No product verdict. Run record.
No product verdict. Run record.
