neatlogs-hack

Star0
View all files

Activity

Contributors

README

Blast Radius

What does a vulnerable dependency put at risk, in business terms, and can an agent fix it safely? Team Redemption (solo) · neatHack 2026 · built with neatlogs, Entire and cfo.ai

When a package is vulnerable, the people who decide what to do (product, finance, support, security) don't see packages. They need what breaks, who owns it, what it costs, by when. Blast Radius maps code → packages → features → revenue and owner, then an agent upgrades the packages, patches code that breaks, runs the tests, and re-checks with an independent scan.

Results (honest numbers)

Same task, 3 runs each, tests pass 3/3 on both (cost is our estimate from token counts, not billing):

Naive agentImproved agentChange
LLM steps19.74.3-78%
Tool calls36.35.3-85%
Est. cost / run$0.199$0.041-79%
Time48.4s18.5s-62%

On 4 real open-source repos (rewound to just before maintainers merged a Dependabot fix) the agent matched or beat the human fix and tests still passed. Only 6 of 51 old repos could even be rebuilt, which is a finding in itself. Details and limits: OSS_REPLAY.md.

How the three required tools are used

ToolIn the productIn building itEvidence
neatlogsAudit trail of every fix (what the agent read, changed, spent)The first trace exposed wasted re-reads; fixing it gave the table above. Also holds a real failed-action → recovery tracehackathon.md (trace IDs), social/media/*trace*.jpg
EntireGraph finds the files and covering test for a package (agent --mode entire). Checkpoints tie each commit to the session that made itEvery commit has a checkpointCheckpoint · ID 01M4JYYTA9J94PDTA97JDYS36X
cfo.aiPricing and runway from the measured cost per fixShowed why the cost cut matters at real-repo scaledocs/cfo-plan.md, social/media/cfo-*.jpg

Run it

cp .env.example .env            # LITELLM_API_KEY, NEATLOGS_API_KEY (agent only); the dashboard needs no keys
uv venv && uv pip install -r requirements.txt

.venv/bin/uvicorn web.app:app --port 8765     # dashboard at /, onboarding at /onboard
demo/shop/scripts/reset.sh                    # fresh seeded demo shop (vulnerable pins)
.venv/bin/python -m agent.run --label try --mode entire        # fix it (naive | improved | entire)
.venv/bin/python -m oss.replay data/oss_candidates.json Snehasis5/ci-cd-demo   # replay a real repo

The agent runner uses a shared, budgeted model key, so it is not deployed publicly; the deployed dashboard is read-only.

Layout

agent/ scan (OSV), graph, fixer loop, tools · web/ dashboard + onboarding · oss/ real-repo replay · demo/shop/ seeded app · video/ Remotion demo · data/runs/ raw run results · docs/ plans

Limitations

Demo business data is hand-written (business.yaml); real onboarding asks the buyer to confirm it. Python requirements.txt only. Small samples (n=3 per agent, 4 real repos, some with a single test). Costs are estimates. Onboarding is a prototype with no sign-up or billing.