agent-purgeOne persona, many coding agents

A local prompt and persona console · zero production dependencies · every write is byte-reversible

One configuration, many entry points

CLI and Web share the same persona library and armed state — single-shot launches, persistent configs, in-session /purge, and project rule files.

Per-harness autonomy

Harnesses follow global settings by default, or get their own persona and injection mode. The console only offers channels that were measured to work.

Hot-switch with honest boundaries

Harnesses with per-turn callbacks apply changes on the next turn; static config entry points require a new session, and the UI says so.

Byte-reversible

Every write stashes the original bytes and restores them exactly on uninstall. Project rules get diff preview and SHA conflict detection.

Local first

No external backend, no config uploads. Data lives in ~/.agent-purge/ and survives upgrades.

Measured, not claimed

Purge effect, built-in safety clauses and residue per harness are backed by isolated-HOME captures, and tests fail when numbers contradict measurements.

The problem

Every coding agent ships its own system prompt: product identity, behavioural rules, refusal clauses, tool instructions. To work your way, you rewrite a different config for each host — and a bad edit is hard to undo.

agent-purge collects those entry points into one place: one persona, injected into the hosts you pick; one command, restored exactly.

Two ways to drive it

ConsoleCLI

bash

agent-purge web

bash

agent-purge codex --mode replace
agent-purge use concise
agent-purge arm pi --slash

Read the measurements first

Purge capability differs a lot per host, and this project writes the difference down:

Full replace: Codex, OpenCode, DSH, Claude, Qwen, Qoder, Kimi, Kiro, CodeBuddy, Grok, omp — the host exposes a full system-prompt entry point.
Partial: Claude (a 62-character identity sentence remains), Kilo (soul() prepends an identity).
Append only: Antigravity, Crush, Continue, Auggie, Goose, zcode — no full entry point; the native identity stays in the prompt.

Per-harness evidence, measured sizes and built-in safety clause data: Support matrix.

Console

agent-purge console

agent-purge project rules