DeployGuard Uses an Incident Memory to Answer 'Have We Seen This Deploy Fail Before?'

PranaviYalala · reddit · 2026-09-30

A developer built DeployGuard after realizing most bad deploys weren't new: the same failure had often happened months earlier, with evidence buried in CI logs nobody would search. DeployGuard watches connected GitHub repos, assigns each push a LOW/MEDIUM/HIGH risk call based on the repo's own history, and cites past deployments—advising only, never blocking or rolling back.

The heart of the project is incident memory, and the author shares five design lessons:

Separate facts from memory: PostgreSQL answers "what happened"; Hindsight (a searchable memory layer) answers "have we seen something like this".
Record only what you observed: keep the failed job and log tail, leave root cause empty—never let a model guess, because "a plausible guess stored as fact poisons every later answer".
Write unknowns explicitly: each memory literally says "Root cause: not known".
Write memories as sentences, not JSON: semantic recall works far better with natural language containing failure stage, changed files, and log tail.
Make every write idempotent: GitHub retries and CI re-runs shouldn't duplicate incidents.

The payoff: asking "have we seen a database connection timeout before?" recalls the past incident whose log said ETIMEDOUT, even without matching keywords. Code is open source on GitHub.

Original post →

DeployGuard Uses an Incident Memory to Answer 'Have We Seen This Deploy Fail Before?'

More from coding & agent