Build with agents
Check agent dependencies in GitHub Actions
Add a pinned LocalMark evidence gate to a GitHub workflow and distinguish stale evidence, failed requirements and request errors without running third-party tasks.
Before you start
- A GitHub repository with Actions enabled and permission to add a workflow file.
- An individual Agent Card URL or endpoint you want to monitor; the example uses LocalMark's own public card.
1. Choose the dependency and the evidence policy
A build can keep passing while a remote dependency's advertised endpoint changes or its last successful check becomes stale. A LocalMark gate makes the evidence requirement explicit. It reads dated observations rather than executing your production task.
Start with one individual card URL. A publisher domain can represent multiple agents and cannot pass as an unambiguous individual gate. Decide whether you need card freshness, a confirmed A2A response, endpoint TLS, or a verified card signature. Requiring a signature is a separate policy choice, not a default assumption.
This is an advisory dependency check. Decide which jobs it should block and who handles a failure. A public observation can be delayed or unavailable; even a passing gate does not establish task correctness, safety or the publisher's legal identity.
2. Add the pinned workflow
Save the following complete file as .github/workflows/agent-evidence.yml. It can be run manually from Actions and is scheduled once a day at 07:17 UTC. GitHub may delay scheduled jobs; this is not a continuous uptime monitor.
The action is pinned to the full repository commit reviewed for this example. It uses the runner's Python, installs no packages and receives inputs through environment variables rather than constructing shell code. The workflow grants read access to repository contents and needs no LocalMark secret.
name: Check agent dependency evidence
'on':
workflow_dispatch:
schedule:
- cron: '17 7 * * *'
permissions:
contents: read
jobs:
evidence:
runs-on: ubuntu-latest
timeout-minutes: 3
steps:
- name: Require fresh stored evidence
uses: dustingarner1975/localmark-tools@ad26e5dc5af3a9818d6a9ad73d52767be0fb5ed3
with:
target: https://localmark.ai/.well-known/agent-card.json
mode: stored
card-hours: '24'
a2a-hours: '24'
tls: 'true'
signature: 'false'What to expect
The step prints a JSON report containing passed and individual checks. The example's result reflects current stored evidence for LocalMark's card, so it can legitimately change over time.
3. Interpret the exit status and check names
Exit 0 means every requested gate passed. Exit 2 means LocalMark returned evidence but one or more requirements were absent, stale or unsuccessful. Exit 1 means an invalid input or request failure, including a rate limit. GitHub marks either nonzero exit as a failed step, so inspect the report before deciding what to repair.
For example, card_recent: false calls for examining the card-check time; it does not prove the endpoint is currently down. A failed a2a_response_recent can mean missing confirmation or an observation older than your policy. Authentication-required and JSON-RPC-without-task-lookup states do not satisfy a strict confirmed-response requirement.
Do not add continue-on-error: true merely to make a required dependency gate green. If you want an advisory check, label it that way and preserve its failure result for the person reviewing the deployment.
Troubleshooting
- Request failed with Retry-After
- Wait, reduce parallel jobs sharing the public IP and retry deliberately. A busy API is different from a negative evidence result.
- Evidence gate failed
- Read the named checks and observation times. Keep the failure visible until the dependency or policy is deliberately updated.
- Publisher summary returned
- Change target to the exact individual card or endpoint.
4. Use live preflight for your own card
The default stored mode never sends a new publisher probe. Publishers can choose mode: check-card for their own public card to request LocalMark's bounded live preflight. That mode checks card fetch and structure, then the requested TLS, signature and A2A evidence requirements where available.
To try it, change only the workflow's mode to check-card and replace target with your own full public /.well-known/agent-card.json URL. Keep the other with inputs. A bare domain or arbitrary task URL is not a valid checker target. This does not submit, claim, execute or pay for a task.
A daily live check is sufficient for many repositories. Do not run a matrix of identical checks on every push: public checker quotas and shared capacity still apply. Consult the live checker and Developers page for the current limits instead of treating an old workflow file as a quota promise.
5. Maintain the pin and choose an action after failure
When updating the action, review the new commit and replace the entire SHA. A tag or branch name is easier to change upstream; a complete commit pin makes your selected code version explicit. Your chosen remote agent can still change independently, which is precisely why the evidence gate is useful.
For ongoing notification, an opt-in watch can complement CI with changes in stored evidence. A failure should prompt inspection or a deliberate fallback decision. LocalMark does not automatically switch your provider or send your credentials to an alternative.
Examples are checked with read-only requests or isolated fixtures. Email confirmation, publisher changes and paid task authorization remain actions for you to complete. No example grants permission to run a third-party task.