Senior QA Engineer · AI Agents
I build devin-* — twenty local-first tools that turn Devin's own session data into backups, search, metrics, memory and QA. No cloud, no telemetry, nothing to sign up for.
Senior QA Engineer, post-graduated in Software Testing and Software Engineering. I work at the intersection of quality engineering and developer tooling — and I spend my open-source time building instruments for the thing most agent platforms leave opaque: their own local data.
The devin-* ecosystem is that work in public: twenty MIT-licensed, local-first utilities for the Devin CLI — diagnostics, history export, backup, metrics, memory, evaluation and a calibrated decision layer that runs on Devin's own model. Every tool is read-only by default, documented in English and Portuguese, and works with no VM, tunnel or extra services.
Understand the system — Devin's store internals documented: schema detection, typed parsers, fixtures.
01 · understand · ★1 · v0.1.0Verify the agent — claims vs. reality, reported as PASS / PARTIAL / UNVERIFIED.
02 · verify · ★1Measure the agent — replay recorded sessions against rubric graders, deterministically.
03 · measure · ★1Control the agent — policy-gated ACP client with allow / deny / ask rules.
04 · control · ★1Judge the decision — calibrated answer layer; ECE 0.170 → 0.071 through Devin's own model via ACP.
05 · judge · ★1 · v0.1.0brew doctor for a Devin install — stores, locks, config, disk → concrete fixes.
diagnostics · ★1 · v0.1.0Session history → Markdown, JSON, CSV, Obsidian-ready.
export · ★1 · v0.1.0Project manager over sessions — rollups, milestones, status reports.
reporting · ★1Snapshot, verify, restore, rotate — with schema-version manifests.
backup · ★1Secret/PII redaction that understands tool-call semantics, in-place in SQLite.
security · ★1Cost, tokens, sessions per project/model/day — local only.
analytics · ★1Knowledge graph — sessions ↔ projects ↔ files ↔ tools.
graph · ★1FTS5 full-text search across every session.
search · ★1Anti-poisoning memory — provenance, versioning, quarantine.
memory · ★1Session lifecycle — export-then-delete, tiered, dry-run first.
automation · ★1Background-worker fan-out — deterministic planner, caps, no nesting.
orchestration · ★1Maintainer hub — registry, scaffolder, weekly reports.
meta · ★1Live Devin activity as an animated SVG circuit board — sessions, subagents, tools.
dashboard · ★1 · v0.1.0Optional add-on — Ollama ↔ OpenAI bridge for self-hosted model operators.
optional · ★1Curated list of Devin resources and the devin-* open-source tooling ecosystem
new · ★1In local SQLite stores under ~/.local/share/devin/cli/ plus ~/.config/Devin/ on Linux, or %APPDATA%\devin\ and %APPDATA%\Devin\ on Windows. devin-internals-spec documents the schema and detection logic.
devin-history reads the sessions DB and exports Markdown, JSON or CSV — including Obsidian-ready notes.
devin-backup snapshots the stores with schema-version manifests, then verifies, restores and rotates them.
devin-redact understands tool-call semantics, not just plaintext patterns — redacting in place inside SQLite, with a .bak safety copy.
Yes — the whole catalog needs only the Devin CLI and Python ≥ 3.10: no VM, no tunnel, no model server, no Slack, no Obsidian. The only Node.js dependency is devin-bridge, an ACP client for the Devin CLI itself. An external runtime (hooks, scheduler, Slack) can recreate proactivity locally, but four losses are irreducible: wake-while-idle when the machine is suspended, remote reach without a notification channel, heavyweight offload to a VM, and multi-machine dashboard topologies without a tunnel.
No. Everything is local-only, read-only by default, and MIT-licensed. Nothing phones home; there is nothing to sign up for.
synced from github · 2026-10-03 18:07Z