Everything here is open source. I build local-first tools for AI-agent governance, software-repo sensemaking, and evidence-backed health workflows. The health side spans trainer intelligence, athlete recovery, nutrition experiments, return-to-play tracking, and academic load planning; the repo tooling side is about making AI-assisted work easier to inspect, constrain, and revive.

Pasadena, CA · @conalhck

A layered review stack for AI-agent work: one substrate, five PR-time detectors, one live runtime monitor, one meta-reviewer.

Substrate

• agent-gov-core — canonical Finding schema, mergeFindings, JSONC/TOML/MCP/shell/transcript parsers. Zero runtime deps.

PR-time detectors — each runs standalone as a CLI or GitHub Action.

• ScopeTrail — diff of agent config files (.claude/settings.json, .mcp.json, Codex sandbox).
• PolicyMesh — contradictions across MCP, Claude, Cursor, Codex, Aider configs.
• CapabilityEcho — new network, subprocess, eval, lifecycle, or workflow-permission signals on the added diff lines.
• TaskBound — scope creep: PR diff vs stated task.
• SessionTrail — risky runtime behavior in Cursor/Claude/Codex transcripts (credential reads, curl|sh, unknown MCP servers, scope escapes).

Live runtime monitor

• AgentPulse — terminal dashboard that classifies live agent sessions (converging / exploring / stuck / done / drifting / idle). Deterministic, no LLM.

Meta-reviewer

• GovVerdict — ingests JSON reports from the five PR-time detectors, dedupes by fingerprint, renders one consolidated PR review.

Demo

• agent-gov-demo — a rogue PR (#1) that trips all five detectors at once. The PR is deliberately titled "fix: typo in README" — TaskBound is meant to catch that.

Example workflow: agent-gov-review.yml.

Standalone tools for understanding codebases, reviewing risky changes, and bringing old repos back into motion.

• overreach — fast capability scanner for diffs, files, and repos. Flags outbound network calls, subprocess spawns, sensitive-file reads, curl | sh, disabled TLS, and hardcoded secrets before they merge.
• Project Autopsy — evidence-backed autopsy reports for stale software repositories: score, verdict, findings, stall hypotheses, revival tasks, and source evidence from the tree, manifests, docs, commits, and dependency signals.
• RepoBrief — orientation layer for unfamiliar repos. Turns a public GitHub URL or local path into an architecture map, key-file guide, risk summary, hotspot list, and where-to-start onboarding path.

Evidence-backed health and performance tools. The shared pattern is conservative, explainable decision support: clear inputs, visible rules, downgraded confidence when data is thin, and no medical diagnosis or automatic clearance.

• fit-ontology — the trainer-facing side. Client-intelligence ontology that unifies wearables, intake, and ACSM guidelines into one queryable model with an explainable rules layer. Engine v2 produces a weekly training recommendation per client, traceable back to the exact metric rows that fired each rule.
• recovery-trail — the athlete-facing companion. Drop an Apple Health export in your browser and get a two-week recovery briefing — a heatmap across HRV, RHR, sleep, and load, an ACSM-aligned training verdict, and the exact rules that fired with the raw slope numbers behind each one. Same engine v2 logic, ported to TypeScript. 100% client-side; the file never leaves the tab. Live demo.
• Nutrition Experiment Lab — private lab notebook for n-of-1 nutrition experiments. Pick one bounded question, log daily outcomes, and get a transparent readout of what changed, adherence quality, confounders, confidence, and what to test next.
• Injury Return-To-Play Tracker — clinician- and coach-friendly return-to-play workflow for injured athletes: phase progress, symptom trends, functional test evidence, workload tolerance, reports, and explicit human clearance decisions.
• Academic Load + Burnout Monitor — student command center for courses, assignments, check-ins, study sessions, and planning blocks. Produces explainable workload signals and recovery-aware next actions without diagnosing burnout or any mental health condition.