aegis-boundary — Aegis Boundary Skill
Advisory review: does this work stay within Aegis product boundaries — in
what it sees, does, outputs, and claims?
What this skill is NOT (read first, non-negotiable)
This skill does not replace the aegis-trust SDK or Aegis Core.
- It does not make authorization decisions for production systems.
Runtime enforcement belongs to the SDK (LITE, cooperative hygiene) and to
Aegis Core (FULL, non-cooperative enforcement with key separation).
- It does not produce receipts, tamper-evident audit, or
decision_ids.
Its output is advisory review, labeled source: aegis-boundary-skill —
never source: CORE (that label is reserved for gateway-constructed
decisions).
- Product, roadmap, pricing, and publication decisions require human approval.
This skill can only return REVIEW / STOPPED and route there.
Routing — what this skill does NOT do
Stay scoped to boundary review. Route other needs to the right place:
- Release / ship-readiness scoring → your team's release process and readiness
checks, not this skill.
- Adversarial fact-verification of a specific strong claim → a dedicated
claim-verification step.
- A production access decision ("may this agent read X now?") → that is SDK /
Core runtime territory, out of scope here.
Review procedure
Step 1 — classify the surface under review
Pick every class that applies:
- A. code change touching an Aegis integration (SDK usage, policy files,
MCP proxy wiring, gateway clients)
- B. document / claim change (README, docs, site, sales, pitch, customer
answer) that says what Aegis does
- C. release / publish preparation of an artifact that makes claims about
Aegis
- D. agent task plan — an agent is about to read data / run tools / emit
output through Aegis surfaces
Step 2 — machine checks (decision-capping)
Resolve the script against this skill's own directory (it does not depend on
the current working directory):
python3 "<skill-dir>/scripts/check_overclaims.py" <changed .md files...>
# In a Claude Code plugin this is "${CLAUDE_PLUGIN_ROOT}/skills/aegis-boundary/scripts/check_overclaims.py";
# when run from the skill directory itself, "scripts/check_overclaims.py" also works.
Scans for the three known overclaim families defined in
references/boundary_canon.md §5-A (C-1, C-2, C-3). Exit codes cap this
review's decision: any FAIL means
the final decision can never be READY. They cap nothing at runtime. Do not
reinterpret, soften, or override a machine FAIL with narrative judgment.
Known limitation: the patterns match English phrasings only. For non-English
claims, apply Step 3 review against the reframes in
references/boundary_canon.md — a lint PASS on non-English text is evidence
of nothing.
Step 3 — boundary review against the bundled canon
Read references/boundary_canon.md (vendored, parity-tracked — see
docs/PARITY.md in the repository) and check, for the classes from Step 1:
| Class |
Check |
| A, B |
Names used correctly (aegis-trust = SDK; Aegis Core = gateway + crypto core; the package name aegis-sdk is reserved and never used) |
| A, B |
Responsibility split — SDK is cooperative hygiene; only Gateway+Core enforce against a non-cooperating agent. Never describe the former as the latter |
| B |
LITE/FULL claim separation (inviolable table in the canon) — "Enforced by Aegis Core", "tamper-evident", "policy authorized" are FULL-only vocabulary |
| B |
Claim discipline — classify every capability statement as implemented / customer-side / gap+roadmap / never-claim before writing it |
| A, B, C, D |
Known pitfalls — references/pitfalls.md |
| D |
What may the agent see / do / output: a declared purpose is a label, not authorization; fail-closed beats silent degrade |
Step 4 — decide and emit
Fill templates/boundary_review.md. Decision rules:
- STOPPED — any machine check FAIL; or the change introduces a never-claim;
or it attributes FULL-only vocabulary to LITE; or it requires a
product/roadmap/pricing/publication decision (route to a human).
- REVIEW — responsibility boundaries blur but intent seems legitimate; a
conditional claim lacks its mandatory qualification; naming is off-canon;
implemented-vs-roadmap status is uncertain; or evidence is missing.
- READY — machine checks pass AND no canon conflict found. Say so briefly.
Output discipline
- Quiet when clean: READY output is a few lines. Spend words only on REVIEW /
STOPPED (what was caught, why, what is allowed instead, how to proceed).
- Every Blocked item cites its canon section or pitfall number. No uncited
blocks.
- Never assert time-dependent facts (versions, counts, customers, prices) from
this skill's text — point at the live source.
Adaptation
The rules above are fixed; everything else adapts. If the situation is not
covered, do not guess a rule: return REVIEW, state what could not be checked,
and name what the canon would need to cover. If a bundled reference file is
missing, that is a STOPPED finding, not a license to improvise.
1---2name: aegis-boundary3description: |4 Aegis Boundary Skill — review whether AI-agent work stays within Aegis product5 boundaries in what it sees, does, outputs, and claims.6 ADVISORY review only: it does not authorize, enforce, or produce receipts —7 runtime enforcement belongs to the aegis-trust SDK (LITE) and Aegis Core (FULL).8 Reviews whether an agent task, code change, document, or public claim preserves9 Aegis boundaries (SDK/Core responsibility split, LITE/FULL claim separation,10 honesty guardrails, never-claims, naming) and returns READY / REVIEW / STOPPED11 with machine-check evidence.12 Use when: working with aegis-trust or Aegis Core integrations; editing README,13 website, sales, or customer-facing claims about Aegis; changing what an AI agent14 is allowed to see, do, or output through Aegis; before publishing any artifact15 that makes claims about Aegis; or when asked "boundary review", "claim check",16 "境界確認", "claim parity", "can we say this publicly?".17 NOT for: production authorization decisions (SDK/Core runtime territory),18 projects unrelated to Aegis.19allowed-tools:20 - Bash21 - Read22 - Grep23metadata:24 source: https://skillmd.com/skills/nemotek-inc/aegis-boundary25---2627# aegis-boundary — Aegis Boundary Skill2829> Advisory review: does this work stay within Aegis product boundaries — in30> what it sees, does, outputs, and claims?3132## What this skill is NOT (read first, non-negotiable)3334This skill does **not** replace the aegis-trust SDK or Aegis Core.3536- It does **not** make authorization decisions for production systems.37 Runtime enforcement belongs to the SDK (LITE, cooperative hygiene) and to38 Aegis Core (FULL, non-cooperative enforcement with key separation).39- It does **not** produce receipts, tamper-evident audit, or `decision_id`s.40 Its output is **advisory review**, labeled `source: aegis-boundary-skill` —41 never `source: CORE` (that label is reserved for gateway-constructed42 decisions).43- Product, roadmap, pricing, and publication decisions require human approval.44 This skill can only return REVIEW / STOPPED and route there.4546## Routing — what this skill does NOT do4748Stay scoped to boundary review. Route other needs to the right place:4950- Release / ship-readiness scoring → your team's release process and readiness51 checks, not this skill.52- Adversarial fact-verification of a specific strong claim → a dedicated53 claim-verification step.54- A production access decision ("may this agent read X now?") → that is SDK /55 Core runtime territory, out of scope here.5657## Review procedure5859### Step 1 — classify the surface under review6061Pick every class that applies:6263- **A. code change** touching an Aegis integration (SDK usage, policy files,64 MCP proxy wiring, gateway clients)65- **B. document / claim change** (README, docs, site, sales, pitch, customer66 answer) that says what Aegis does67- **C. release / publish preparation** of an artifact that makes claims about68 Aegis69- **D. agent task plan** — an agent is about to read data / run tools / emit70 output through Aegis surfaces7172### Step 2 — machine checks (decision-capping)7374Resolve the script against this skill's own directory (it does not depend on75the current working directory):7677```bash78python3 "<skill-dir>/scripts/check_overclaims.py" <changed .md files...>79# In a Claude Code plugin this is "${CLAUDE_PLUGIN_ROOT}/skills/aegis-boundary/scripts/check_overclaims.py";80# when run from the skill directory itself, "scripts/check_overclaims.py" also works.81```8283Scans for the three known overclaim families defined in84`references/boundary_canon.md` §5-A (C-1, C-2, C-3). **Exit codes cap this85review's decision: any FAIL means86the final decision can never be READY.** They cap nothing at runtime. Do not87reinterpret, soften, or override a machine FAIL with narrative judgment.8889Known limitation: the patterns match English phrasings only. For non-English90claims, apply Step 3 review against the reframes in91`references/boundary_canon.md` — a lint PASS on non-English text is evidence92of nothing.9394### Step 3 — boundary review against the bundled canon9596Read `references/boundary_canon.md` (vendored, parity-tracked — see97`docs/PARITY.md` in the repository) and check, for the classes from Step 1:9899| Class | Check |100|---|---|101| A, B | Names used correctly (aegis-trust = SDK; Aegis Core = gateway + crypto core; the package name `aegis-sdk` is reserved and never used) |102| A, B | Responsibility split — SDK is cooperative hygiene; only Gateway+Core enforce against a non-cooperating agent. Never describe the former as the latter |103| B | LITE/FULL claim separation (inviolable table in the canon) — "Enforced by Aegis Core", "tamper-evident", "policy authorized" are FULL-only vocabulary |104| B | Claim discipline — classify every capability statement as implemented / customer-side / gap+roadmap / never-claim before writing it |105| A, B, C, D | Known pitfalls — `references/pitfalls.md` |106| D | What may the agent see / do / output: a declared purpose is a label, not authorization; fail-closed beats silent degrade |107108### Step 4 — decide and emit109110Fill `templates/boundary_review.md`. Decision rules:111112- **STOPPED** — any machine check FAIL; or the change introduces a never-claim;113 or it attributes FULL-only vocabulary to LITE; or it requires a114 product/roadmap/pricing/publication decision (route to a human).115- **REVIEW** — responsibility boundaries blur but intent seems legitimate; a116 conditional claim lacks its mandatory qualification; naming is off-canon;117 implemented-vs-roadmap status is uncertain; or evidence is missing.118- **READY** — machine checks pass AND no canon conflict found. Say so briefly.119120## Output discipline121122- Quiet when clean: READY output is a few lines. Spend words only on REVIEW /123 STOPPED (what was caught, why, what is allowed instead, how to proceed).124- Every Blocked item cites its canon section or pitfall number. No uncited125 blocks.126- Never assert time-dependent facts (versions, counts, customers, prices) from127 this skill's text — point at the live source.128129## Adaptation130131The rules above are fixed; everything else adapts. If the situation is not132covered, do not guess a rule: return REVIEW, state what could not be checked,133and name what the canon would need to cover. If a bundled reference file is134missing, that is a STOPPED finding, not a license to improvise.