Experimental · local-first · open source

Catch the turn
before it drifts.

Spotter supervises coding agents while they work. It watches the trajectory—not only the final diff—so weak assumptions, repeated failures, and unsafe actions can be caught before their cost compounds.

Prototype status: deterministic gates are active. Semantic advisories remain off by default; recovery controls are not shipped.

01 / OBSERVE
02 / DETECT
03 / ACT
01 / WHY

The cost of late feedback

A bad trajectory rarely begins with a bad diff.

A weak assumption shapes the next search. The search shapes the next edit. By the time review catches the result, the agent may have spent minutes, tokens, and repository churn defending the wrong path.

Post-hoc review asks whether the output is acceptable. Runtime supervision asks whether the work is still heading somewhere worth going—and whether the next action is safe.

02 / MODEL

Supervision loop

Observe the work.
Change the outcome.

Spotter keeps deterministic enforcement fast and local while slower semantic review stays off the synchronous tool path.

  1. 01

    Observe

    Collect actions, evidence, edits, validation, and runtime identity into a durable trajectory.

  2. 02

    Detect

    Find repeated failures, stalled exploration, scope growth, missing validation, and stale hypotheses.

  3. 03

    Intervene

    Block deterministic violations now; evaluate careful semantic guidance before broader use.

  4. 04

    Recover

    Preserve snapshots and replay material today while live interrupt and restart remain target capabilities.

Intervention ladder

Escalate only as evidence warrants.

  1. BLOCKactive · deterministic
  2. VERIFYexperimental · opt-in
  3. NUDGEexperimental · opt-in
  4. INTERRUPTtarget · not shipped
  5. RESTARTtarget · not shipped
03 / ARCHITECTURE

Current and target

A prototype with a deliberate runtime boundary.

CURRENT

Hook-backed supervision

Codex HooksHook process ↔ spotterdJournal · Gate · Review

Hook processes write journals, snapshots, and periodic reviews; bounded daemon IPC serves deterministic gates. Replay, diagnostics, and signal-driven review also exist.

TARGET

Standalone runtime control

Codex App ServerspotterdObserve · Steer · Recover

App Server ingestion and control foundations are partial. Broad live advisories, interrupt, and verified restart remain gated by evidence and hardening.

Read the architecture contract →
04 / STATUS

Honest by design

What works today.

This is a stable summary, not a duplicate changelog. The repository status document is the source of truth for exact capability boundaries.

Implemented

  • Durable trajectory journals
  • Bounded deterministic gates
  • Snapshots, forks, and replay
  • Runtime health diagnostics

Partial / shadow

  • App Server observation
  • Event-driven detection
  • Semantic review
  • Opt-in live advisories

Not live yet

  • Broad VERIFY / NUDGE
  • Live INTERRUPT
  • Verified RESTART
  • Proven outcome benefit
Check the canonical current status →
05 / ROADMAP

Evidence before escalation

Build trust in named outcomes.

  1. 01Runtime
  2. 02Observe
  3. 03Detect
  4. 04Intervene
  5. 05Recover
  6. 06Harden

Each outcome has an evidence gate. Implementing a mechanism is not the same as proving it improves results.

Explore the roadmap and evidence gates →

Get started

Run Spotter locally.

Install the CLI and daemon from the official Homebrew tap. Setup is explicit, transactional, and does not silently change Codex during package installation.

terminal
$ brew install spotter-agent/spotter/spotter
$ spotter setup codex --dry-run
$ spotter setup codex
$ spotter doctor