verify-against-intent— agent skill

Preamble (run first)

# Suiperpower telemetry. Routes through the CLI so projects.json + Convex stay in sync.
# Silent on failure so the skill never blocks. Opt out: SUIPERPOWER_TELEMETRY=off.
#
# AGENT NOTE: when this skill finishes, run the matching completion command:
#   suiperpower track verify-against-intent build completed
# Or use "failed" / "aborted" if it ended that way.
command -v suiperpower >/dev/null 2>&1 && suiperpower track verify-against-intent build started >/dev/null 2>&1 &
true

If TEL_PROMPTED is no, before doing real work, ask the user:

Help suiperpower get better. We track which skills get used and how long they take. No code, no file paths, no PII. Change anytime in ~/.suiperpower/config.json.

>

A) Sure, anonymous B) No thanks

Write the answer to ~/.suiperpower/config.json telemetryTier field and create ~/.suiperpower/.telemetry-prompted. Then continue.

What this skill does

Closes the loop on a Sui build session with disk-grounded checks, not session-memory claims. Walks the plan's Object model, capability holders, sponsor surfaces, and upgrade posture against actual Move source, Move.toml, deploy-context.md, and the test suite. Names drift honestly: decorative sponsor integration claimed as load-bearing, capabilities held in the wrong place, public functions without tests, deploy state that does not match the planned network rollout.

Without this gate, agents say "done" against a target that was never written. With it, every criterion gets a pass / fail / partial verdict tied to a concrete artifact.

When to use it

• A build skill (build-with-move, walrus-storage, deepbook-orderbook, scallop-money-market, sui-zk-login, kiosk-marketplace, etc.) just reported done.
• The user says "did we get there", "is this what I asked for", or "review what we built".
• A session is about to be wrapped, before learn captures it.
• The user is preparing to ship: run this before deploy-to-testnet, deploy-to-mainnet, or submit-to-sui-overflow.

When NOT to use it

• .suiperpower/intent.md does not exist. There is nothing to verify against. Use clarify-intent to backfill, then re-run.
• The user wants a security audit, use review-move (different gate, P0 to P3 findings) or ottersec-prep.
• The user wants UX critique, use product-review or roast-my-product.
• The user wants a code-style review. Not this skill.

If you activated this and the user actually wants something else, consult skills/SKILL_ROUTER.md and hand off.

Inputs

• .suiperpower/intent.md (required).
• .suiperpower/build-plan.md if plan-before-code ran.
• .suiperpower/build-context.md, .suiperpower/deploy-context.md if present.
• The current project tree: Move sources, Move.toml, tests/, frontend if any.

If intent.md is missing, stop. Tell the user verification needs a recorded intent. Offer clarify-intent to backfill.

Outputs

A printed report and an append to .suiperpower/build-context.md:

## verify-against-intent, <timestamp>

### Success criteria check
1. <criterion>: pass | fail | partial, evidence: <file:line | testnet tx hash | package id>
2. <criterion>: pass | fail | partial, evidence: <...>
...

### Object model check
- <ObjectName>: planned <ownership, abilities> | actual <ownership, abilities>: match | mismatch
- ...

### Capability holder check
- <CapName>: planned holder <X> | actual at init <Y>: match | mismatch

### Sponsor integration check (load-bearing test)
- <Walrus | DeepBook | Scallop | OZ Sui | zkLogin>: planned surface <...> | actual: <imported only | called but result unused | load-bearing in user flow with evidence <...>>

### Test coverage check
- Public entry points with at least one test: <count> / <total>
- Cap-gated functions with expected-failure tests: <count> / <total>
- Functions missing tests: <list>

### Build check
- `sui move build`: pass | fail (<error summary if fail>)
- `Move.toml` deps pinned (rev or tag): yes | no

### Upgrade and deploy posture check
- Planned network rollout: <devnet -> testnet -> mainnet>
- Actual deploy state: <per-network package id from deploy-context.md, or "not deployed">
- Upgrade authority: planned <X> | actual <Y from on-chain or deploy-context>: match | mismatch

### Drift summary
<one paragraph, honest>

### Follow-ups
- <thing to fix, with the skill that fixes it>
- <thing to document as accepted scope cut>

The skill does not fix the drift. It names it, points to the right skill, hands off.

Workflow

1. Read inputs

• .suiperpower/intent.md. Missing? Stop.
• .suiperpower/build-plan.md if present.
• .suiperpower/build-context.md, .suiperpower/deploy-context.md.
• List Move modules (sources/*.move), tests (tests/*.move), Move.toml, frontend entry points if applicable.

1. Map criteria to Sui-grounded evidence

For each success criterion in intent.md, decide what would prove it:

• On-chain function existing? Check for public entry fun <name> in source.
• Test pass? Check for the named test in tests/.
• Network deploy? Check deploy-context.md for a package id on the right network.
• Frontend behavior? Check that the route calls the named entry point.
• Sponsor user flow? Check that the SDK call sits inside a real user path, not a top-of-file import only.

Do not trust session memory. Read disk.

1. Object model check

For each Object in build-plan.md:

• Find struct <Name> in source. Parse abilities (has key, store, ...).
• Compare to plan. Note mismatch on ownership (e.g. plan said shared, code uses transfer::transfer, that is owned).

1. Capability holder check

For each capability:

• Locate the init function. See where the cap goes (transfer::transfer(cap, ctx.sender()), transfer::share_object(cap), transfer::public_transfer, or burn).
• Compare to plan. Surface mismatch.

1. Sponsor load-bearing check

For each sponsor in build-plan.md:

• Confirm the SDK call or Move call exists in source.
• Confirm the call is reachable from a public entry function (Move side) or a user-triggered handler (frontend side).
• Confirm the result is used (the blob is read into UI, the order is placed against a real pool, the zkLogin proof is verified, etc.).
• If only imported, or the result is unused: mark decorative. This is the most common drift; do not soften it.

1. Test coverage check

• Count public entry functions vs functions covered by at least one test.
• For each cap-gated function, confirm there is an expected-failure test for the unauthorized caller.
• List the gaps.

1. Build check

• Run sui move build. Record pass or fail.
• Read Move.toml. Confirm deps are pinned to a rev or tag, not floating.

1. Upgrade and deploy posture check

• Compare planned network rollout to deploy-context.md. Per network: deployed or not, package id captured or not.
• Compare planned upgrade authority to the upgrade-cap state recorded at publish.

1. Call drift honestly

• For each mismatch above: pass, fail, or partial. Evidence tied to a file path, a tx hash, or a package id.
• Anything in the plan but not built: was it intentionally cut or forgotten? Ask the user.
• Anything built but not in the plan: scope creep, accident, or load-bearing? Ask.
• Anything in intent but not in plan or code: worst kind of drift, surface first.

1. Report and writeback

• Print the report shape above.
• Append the same text to .suiperpower/build-context.md under a new ## verify-against-intent section.

1. Hand off

• Clean report: recommend deploy-to-testnet (or deploy-to-mainnet if testnet already passed) or learn.
• Drift exists: recommend the right skill for the largest gap (build-with-move, review-move, walrus-storage, etc.) via skills/SKILL_ROUTER.md.

Quality gate (anti-slop)

Before reporting done, the skill asks itself:

• Did I read actual Move source and Move.toml, not recall from session memory?
• Did every success criterion get pass / fail / partial with a concrete artifact reference (file:line, tx hash, package id)?
• Did I run the sponsor load-bearing test, not accept "the import is there"?
• Did I check capability holders at init, not just that the cap struct exists?
• Did I name drift honestly when the build skill called itself done?
• Did I avoid fixing the drift in this skill (the job here is calling, not fixing)?
• Did I append the verdict to .suiperpower/build-context.md so the next session sees it?

If any answer is no, the skill keeps working before reporting done.

References

Knowledge docs (load when scope expands):

• core/skills/data/sui-knowledge/03-move-and-objects.md: Object model and ability reference.
• core/skills/data/guides/package-id-capture.md: Where package id and upgrade cap state are recorded.

Use in your agent

• Claude Code: claude "/suiper:verify-against-intent <your message>"
• Codex: codex "/verify-against-intent <your message>"
• Grok Build: run grok, then /verify-against-intent <your message> in the session
• Cursor: paste a message like "did we build what I asked for on Sui", or reference ~/.cursor/rules/verify-against-intent.mdc.

If you activated this and the user actually wants something else, consult skills/SKILL_ROUTER.md and hand off.