EDO core rules for every phase
Apply these with the rules in SKILL.md in every EDO phase.
Working rules
- Use the user's identity. Use the profile established under Find the user's credentials. Verify account and role/user before account work and before every AWS mutation; stop on an unexpected change. A refreshed session for the same role is expected. Pass the agreed profile and Region explicitly. Do not switch profiles without agreement, read credential files, print credential variables, mint credentials, assume another role, or sign in/refresh for the user. They authenticate their usual way when needed.
- Separate discovery from changes. Read-only checks, local dependency
setup, and local review files may proceed without an AWS approval prompt.
Explain the checks and their results; an unavailable or denied check is
unknown, not passed. Do not claim the user's credentials are read-only.
Run each read-only check as a single command without pipes, so the
user's pre-approvals match it; limit output with
--query/--output. Write a blueprint only from values the user gave or accepted in answer to the discovery questions; a request for a new blueprint starts those questions. A question about a file never changes it. - Execute the agreed plan. Approval covers its named AWS changes, including setup, uploads, permissions, and launch. Run them without repeated approval prompts. Verify the reviewed files still match; changed scope/target (including a different bucket, key, or Region) or recovery after failure needs agreement. A status request or saved record grants no mutation authority. Respect harness permissions; for a blocked step, provide its resolved command and verify user-run results.
- Use existing commands, not new code. Use released EDO tools and documented APIs with values filled in for the user's shell. Do not write helper programs/scripts, modify EDO code, or bypass its validators. Blueprints, parameter files, plans, and progress records are data files.
- Make choices and evidence clear. Distinguish requested values,
observations, and proposed defaults. Explain consequential choices.
Do not invent workload sizing, prices, licensing interpretation, or
acceptance of service terms. Link official pricing pages; never call
pricing APIs, compute costs, or choose options by price. Use only the
starting configuration in
discover.md as defaults, labeled as such,
never other example values. In any blueprint draft, keep
evs.terms_accepted: falseuntil the user confirms the EVS terms as described in plan.md. - Confirm destructive scope. Identify the exact environment and affected resources. Teardown removes and verifies the environment first, then separately approves infrastructure removal. Never stop or terminate EVS hosts through EC2 or suppress deletion prompts without an explicit request.
Find the user's credentials
Use a profile the user names in this conversation or the deployment
record names; AWS_PROFILE, environment credentials, and the default
chain are active credentials, not a named profile. Otherwise find
candidates yourself: the active credentials only (a profile's own name is not
a hint). Before asking, run one aws sts get-caller-identity on the active
credentials, only when the request needs AWS and the session has AWS access.
If the session has no AWS access, make no AWS call: say so and ask the
credentials question. Make no other AWS call and no profile listing or config read
before the answer, including with --profile default. The question offers the
active credentials and lets the user name another profile; list none.
Use a candidate without asking only when the user's request or files point to its account or profile and no other candidate could match; a candidate whose check failed (for example, no network) is not ruled out. State the profile, account, principal, and why. Otherwise, before any other check, ask the credentials question in formats.md; working active credentials are an option to offer, not a choice to make. If the chosen identity's first checks are denied or find nothing for EVS, report that and ask instead of continuing. If several fit or none work, list only the working ones with account and principal, summarize failures in one line by cause, and ask which to use; ask nothing that discovery would answer. Never choose by name alone. Probe only with that identity call, never read credential files, and never sign in or refresh for the user.
Name the cause instead of "no credentials": expired or invalid token means
the user re-authenticates that profile; NoCredentials with profiles
present means select a profile; no profiles visible when the user expects
some means the shell may be sandboxed from their AWS configuration; an
unreachable endpoint or unresolved host means the agent's sandbox may
block network access, which the user allows in their agent's settings;
AccessDenied later means that check is unknown.
Use the matching EDO release
EDO's README owns technical prerequisites, inputs, defaults, commands,
validation, timings, and lifecycle. Read only relevant sections of
Deploy/EVS-Deployment-Orchestrator/README.md; do not copy its procedures
or dependency/version tables into the skill.
Use the requested release, otherwise the available template's named runner
ref. If neither exists, discover and propose a stable published release.
Verify the public ref; never silently choose main or a prerelease. Never
bypass a release blocker with testing-only template parameters or packaged
local code; that needs the user's explicit choice and stays a plan warning.
Use one matching checkout for README, blueprint reference, validators, and
template; verify runner and blueprint compatibility. Preserve the user's
clone and use a separate checkout if needed. Record ref and resolved commit;
stop launch on a mismatch or missing documentation. Cite EDO files at that ref.
EDO's public repository is https://github.com/aws/solutions-for-amazon-evs.
With no user checkout, check the release ref there with git ls-remote. Clone
it, or compare a local checkout with it, only when the user asks; don't ask the
user to fetch or provide a checkout. Without a matching checkout, read the
release's files at its tag instead of cloning:
https://raw.githubusercontent.com/aws/solutions-for-amazon-evs/<tag>/Deploy/EVS-Deployment-Orchestrator/<file>
for README.md, the template, the reference blueprint, tools/validate_blueprints.py,
and spec-generator/constants.py (which the validator imports). Save them in
the deployment folder in the same layout and run the validator there. List
any check whose files you couldn't read as not checked under Warnings.
For an existing deployment, establish account, Region, stack/environment, deployed release, and current state from live read-only evidence. Do not assume a local checkout describes it. Resolve an ambiguous target before changing it.
Carry context between phases
Keep review files together: deployment-plan.md, the working blueprint,
launch parameters, and deployment-state.json. Use one folder per
deployment, evs-deployments/<environment-name>/ in the current
workspace, keep any isolated Python environment there too, and tell the
user its path. Don't write elsewhere in the EDO checkout or edit its files. The record stores
paths and checksums of reviewed files, identity/location/release, check
evidence and blockers, current-conversation approval scope, action results/resource IDs,
and latest observations. Update it after material changes; avoid duplicating
the full plan or blueprint.
Reconcile it with the conversation, current files, identity, and live state before continuing. Re-check evidence affected by changed inputs or stale account state. Resolve identifiers returned by approved actions only through the bindings described in the plan. Never repeat a create/upload solely because a local record is missing. After an ambiguous write, inspect live state before considering a retry; record an unresolved outcome as unknown.