CtrlK
BlogDocsLog inGet started
Tessl Logo
PluginPublic

ainativedev/aidevcon-2026-nyc

AI Native DevCon 2026 New York — practical agent skills grounded in conference talks

Quality

92%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide
SecuritybySnyk

Passed

No findings from the security scan

Overview
Quality
Evals
Security
Files

contract-receipt-template.mdskills/talk-meppiel-versioned-agent-contracts/references/

Contract and receipt design worksheet

This worksheet is Tessl's recommended application of the talk. Field names and controls are design prompts, not an official agent-contract, APM, in-toto or other executable schema. Fill unknowns with explicit placeholders. Do not create fake hashes, signatures, run results or compliance claims.

1. Factory identity and graph

FieldDraft value
Factory purpose and ownerTo define
Factory definition version and immutable digestTo resolve from the actual artifact
Contract graph and handoffsFor each edge: output artifact → next contract input
Input repository revision / assignment artifactTo pin for this run
Approved harness and runtime configurationTo verify against the chosen implementation
Resolved dependency inventoryNames, versions and digests of skills/tools; link to lock record

Version the process, not just its generated code. The graph may include loops or other compositions appropriate to the task; the talk's linear demo is one example.

2. Contract definition

Stable role: What this unit of work is responsible for.

Per-run assignment: The specific change or question it is being asked to handle.

Inputs: Required intent/specification, repository state and other artifacts, with their identity and expected form.

Dependencies: Resolved skills, tools and other resources required to perform the role.

Budget: Cost, time or execution limits selected for this project; the talk does not mandate numerical thresholds.

Deliverables: Named output artifacts and the downstream consumer's expectations.

Required checks:

CheckExact definition/versionRelevant inputs and outputExecutor and protection boundaryRequired result and evidence
To specifyImmutable referenceWhat it actually verifiesWho runs it; why the implementing agent cannot replace itResult plus verifiable evidence

Use deterministic checks where appropriate, following the speaker's Q&A. A deterministic formatting check can establish format compliance; it cannot by itself establish that business behavior is correct. Select tests for the actual task.

3. Run receipt design

EvidenceWhat needs to be bound together
Output identityThe exact diff/artifact submitted for admission
Factory provenanceFactory version/digest and the contract steps that produced the output
Dependency inventoryResolved dependencies used in this run, rather than declarations alone
Check integrityExact check definitions and their results, tied to the relevant output
Run contextInputs and relevant environment identity needed to interpret or reuse results; added implementation recommendation
AttestationSigner identity, evidence covered and verification method supported by the chosen implementation

Keep the agent's execution authority separate from the authority that validates and signs its evidence. This is a recommended trust boundary; the talk's descriptions of the signer's deployment location conflict. Document who can change the checks, submit results, access signing credentials, and update policy. Verify those permissions rather than inferring isolation from machine names.

4. Admission policy and failure behavior

Specify the organization's rules for accepted factory identities, permitted dependencies, trusted signers, required checks, and the binding between evidence and the output under review.

ConditionRecommended gate response
Missing or invalid signature / untrusted signerDo not count the receipt as trusted evidence; route for investigation
Receipt refers to a different output or check definitionReject the evidence for this output
Required check missing, failed or unverifiableDo not admit as a passing run
Denied dependency or unapproved factory identityBlock admission under the specified policy
Evidence valid and all policy conditions satisfiedMark this policy satisfied; apply any remaining release requirements

These are suggested policy behaviors, not authorization to merge or deploy. A policy pass establishes satisfaction of that policy, not universal correctness or regulatory compliance. Preserve testing, scanning or review required by the actual project.

5. Change and incident exercises

  • A check changes: Publish a new factory identity, validate the new check, and update admission requirements. Test that an old receipt cannot satisfy the new requirement simply because its old checks passed.
  • A skill is compromised: Locate affected resolved versions in receipts/inventories, deny them through policy, identify affected outputs and rerun after remediation.
  • An agent edits a check: Demonstrate that the changed definition is detected or cannot be used to sign an accepted result.
  • Evidence is replayed against another diff: Verify that the output binding rejects it. This is an added validation recommendation.
  • Checks move earlier in the workflow: Specify when signed evidence can be reused and which input or environment changes invalidate it. Avoid treating the proposal to reduce duplicate CI work as permission to eliminate verification.

6. Audit summary

FindingObserved evidenceConsequenceProposed controlValidation
Supported / missing / unverifiedExact supplied artifact or absent evidenceWhat remains uncertainSmallest useful changeConcrete test or inspection

Close with unresolved assumptions, the next design decision, and the evidence that would support it.

skills

talk-meppiel-versioned-agent-contracts

tile.json