agent doctor / toolkit

The deterministic shell around AI coding agents. Agents are nondeterministic; the value is a layer that grounds and verifies them. Build leverage — ground truth, coordination, determinism — not judgment a smarter model already does well.

design build plan rule catalog

kernel

Code index

Content-addressed index of every TypeScript file: parse once, build a symbol graph, dependency edges, and structural fingerprints. Warm and incremental: re-parse one file in ~25µs.

Policy / gate

ACL, architecture layering, protected paths, and region leases. A diff that breaks the rules is denied deterministically — exit non-zero. Same gate for humans, agents, and CI.

Impact test selection

Which tests transitively reach a change, from the import graph. Turns “run 3000 tests” into “run the 12 that matter” — sub-µs once warm.

Semantic merge

AST-level 3-way merge. Two agents adding different functions to the same file merge with zero conflict; only edits to the same declaration conflict. Drops in as a git merge driver.

Context server

A warm kernel answering queries over stdio — plain JSON or MCP. The agent pulls ground truth (does this helper exist? what tests cover this?) instead of grepping.

Orchestrator

The deterministic runner: lease, assemble context, run the agent, gate, then retry with the exact findings. The agent is the only nondeterministic step; everything around it is enforced.

Agent-native VCS

A content-addressed operation log with per-agent/per-task attribution and deterministic undo, first-class leases, and semantic merge — our model on a git-compatible backend.

The loop

orchestrator: decompose, lease disjoint regions
  per task (parallel):
    assemble minimal context (kernel queries)
    run the agent            -- the only nondeterministic step
    gate the diff            -- deny? retry with the exact findings
    run impacted tests       -- fail? retry with the failures
    semantic-merge back      -- conflict? only the real one surfaces
    release lease

A hallucinating agent can't corrupt the repo — every mutation flows through the deterministic gate / test / merge it cannot bypass.

Commands

agent-doctor initscaffold policy, gitignore, MCP config, and the merge driver

agent-doctor gate --base main --actor medeny a diff that breaks policy/ACL/leases

agent-doctor impact --base mainthe tests reaching your diff

agent-doctor merge O A Bsemantic 3-way merge (git driver)

agent-doctor serve --mcpcontext server for your agent harness

agent-doctor orchestrate --executor '…'run a task ledger through the loop

Latency — release, Apple Silicon

project	build	incremental	impact	gate	merge
zod (404 files)	7.8ms	25µs	0.2µs	4µs	19µs
zustand (33 files)	2.5ms	46µs	0.2µs	53µs	21µs

The harness drove out two real regressions, both fixed: impact 1300→0.2µs, gate 1193→4µs. CI gates cold-build µs/file.