Agent Dojo — Discipline Capsule v3.8

Publisher: Agent Dojo · aiagentdojo.com · [email protected]
This page is the canonical integrity reference for the capsule package.

Package

What it installs — and all it installs

1. A delete-guard hook (active from install): mechanically blocks hard-delete commands; forces a recoverable trash path or explicit human confirmation.
2. Two shadow guards (NEW in v3.8), installed watching-only: a secret-egress guard (flags secret-shaped strings — API keys, tokens, private keys — entering send-shaped commands or finished replies; logs are structurally redacted and never contain the caught value) and a done-claim guard (flags "it's done/live/deployed" claims made without verification evidence; its date-claim leg — pure calendar arithmetic — is active day one and corrects wrong weekday+date pairs on the spot). In shadow mode they cannot block anything: they log to ~/.agent-dojo/guards/ and wait. Arming is owner-only: a read-only status tool (~/.agent-dojo/guard-status.py) shows each guard's burn-in record and prints the exact one-line arm command when — and only when — that guard's own log on YOUR machine is clean. Fail-open by design: malformed input or missing files never break a working agent.
3. Always-on operating rules written into the agent's instruction file between clear markers: field-earned disciplines that only narrow behavior, a self-introducing transparency rule (first 3 activations each, tracked in a local counter the hook itself reports), one-time onboarding duties (briefing offer, first-24-hours work, Day-1 report — owner approval required for any setup), and week-one equipping duties that restrict how the agent may talk about access: PROPOSE-only, one access at a time, owner-named platforms only, revoke path stated in every proposal, owner's no or silence final (one day-3 reminder maximum). The owner initiates and executes every grant; a local grant-cards reference (rules/grant-cards.md, installed to ~/.agent-dojo/) gives the owner ready-to-click least-privilege steps.
4. Optionally, in the same run only: if a paid license key + auth code + owner name are supplied in the same install.sh invocation, one licensed fetch runs immediately (to capsule-api.aiagentdojo.com, sending only those three values) and writes the licensed Power files to ~/.agent-dojo/powers/ as plain local files. Without those three values, no network call happens at all — and nothing is left anywhere describing one as pending.

Restrict-or-disclose pledge: every rule and mechanism in this package either narrows what your agent will do, or requires it to disclose that an installed rule acted (each discipline announces itself its first 3 activations — tracked in a local counter file the delete-guard itself reports — then acts silently; the behavior never fades). It grants no permission, names no trusted party, relaxes no built-in safety behavior (the rules state explicitly: stricter built-in rules win), and never touches secrets — the two new guards detect secret shapes without opening any credential store, and their logs never contain a caught value. The package invites your agent to verify all of this itself before installing — see VERIFY.md inside.

Note on rules/agent-dojo-rules.md: this file is a fixed placeholder in the downloadable package on purpose — it contains no real rule text, and never will. The actual operating disciplines are delivered gated to your paid license, via a single license-gated fetch performed by install.sh itself (step 6) in the same run, using the license key + auth code + owner name from your purchase email — supplied together with the install request, never in a separate later message. That is the one and only network call anywhere in this install; the placeholder carries no standing instruction to fetch, trust, or install anything at any future time. This keeps the real training content out of any public download while leaving nothing pending in your agent's memory.

Verify

curl -fsSL https://aiagentdojo.com/capsule/agent-dojo-capsule-v3.8.tar.gz -o agent-dojo-capsule-v3.8.tar.gz
echo "c91d8f0b0e5b2576373267e4b31f19c274842aaf0ba82f96915e8395116f2642  agent-dojo-capsule-v3.8.tar.gz" | sha256sum -c
tar -xzf agent-dojo-capsule-v3.8.tar.gz && cat agent-dojo-capsule-v3.8/VERIFY.md

Install

Restriction-only (no license):

bash agent-dojo-capsule-v3.8/install.sh

With a paid license — key, auth code, and owner name in the SAME command, so the whole install (local files + licensed fetch) happens in one sitting:

AGENT_DOJO_LICENSE_KEY=<key> AGENT_DOJO_AUTH_CODE=<code> AGENT_DOJO_OWNER_NAME="<your name>" \
  bash agent-dojo-capsule-v3.8/install.sh

Re-running is safe and is also the upgrade path (guard modes you have set are never overwritten — an upgrade never arms or un-arms anything). Uninstall completely: bash uninstall.sh — one command, ships in the package (and a copy is installed to ~/.agent-dojo/uninstall.sh). It prints every path before removing it, edits settings.json surgically (everything else kept), and moves ~/.agent-dojo/ to a timestamped backup folder instead of hard-deleting — even the uninstall is reversible. Zero network calls; hold it to the same audit as install.sh.

Earned in our field. Installed in your agent.

Previous versions

agent-dojo-capsule-v3.7.tar.gz · SHA-256 a80ae4018edf2604fb0e1309700426b605058b6c17de997f053b5c957868ddcd — superseded by v3.8 (adds the two shadow guards — secret-egress and done-claim/date-claim — plus the read-only guard-status tool; shadow-first, owner-armed, fully removed by the same one-command uninstall). Same-machine upgrade: run the v3.8 install over it — introduction counters and any guard modes you set are preserved.

agent-dojo-capsule-v3.6.tar.gz · SHA-256 2fe3f4bc9956a4fdbdf7a07aa4131ad8dcad7ff884387345815e515a6dc196ba — superseded by v3.7 (adds a tested one-command uninstall: uninstall.sh ships in the package and is copied to ~/.agent-dojo/uninstall.sh at install time; no behavior changes to the hook, rules, or licensed fetch). Same-machine upgrade: run the current (v3.8) install over it — introduction counters are preserved.

agent-dojo-capsule-v3.5.tar.gz · SHA-256 b343c343b27ce86e66442757bfcd046c2f427c3de5ad871baccf092f5db2aba1 — superseded by v3.6 (the license-gated fetch now happens inside the same install.sh run as the local install, from credentials supplied in the same request; no forward-looking "fetch when asked later" language anywhere in the installed rules block). Same-machine upgrade: run the current (v3.8) install over it — introduction counters are preserved.

agent-dojo-capsule-v3.tar.gz · SHA-256 bcc4e0c54516744bc7c99d57eb384a59cd1c8e9671ef2c896f067f87d3997093 — superseded by v3.5 (delete-guard hardening: catches deletes wrapped in bash -c/python one-liners; briefing questions mirror the owner intake form word-for-word; equipping duties + grant cards added).