YAML Compatibility

Agent CI aims to run real GitHub Actions workflows locally. The table below shows current support against the official workflow syntax.

✅ = Supported   ⚠️ = Partial   ❌ = Not supported   🟡 = Ignored (no-op)   🚫 = Not planned

Workflow-Level Keys

Key	Status	Notes
name	✅
run-name	🟡	Parsed but not displayed anywhere
on (push, pull_request)	✅	Branch and path filters are evaluated when using --all
on.<event>.branches / branches-ignore	✅	Branch-name glob filters; --all only runs workflows whose filter matches the current branch (main for non-PR runs).
on.<event>.paths / paths-ignore	✅	Path-pattern filters. With --all, a workflow is skipped if none of the changed files match paths: or all of them match paths-ignore:.
on.<event>.tags / tags-ignore	❌	Tag-pattern filters for push events are not evaluated — --all treats a tagged run as an untagged push. Workflows gated only by tag patterns will run when you don't expect them to.
on.<event>.types	🟡	Activity-type filters (e.g. pull_request.types: [opened, synchronize]) are parsed but not applied — --all runs the workflow regardless of event sub-type.
on.workflow_dispatch.inputs	🟡	workflow_dispatch itself is not simulated, so its declared inputs are parsed but unreachable — the CLI cannot inject them.
on.workflow_call.inputs.* (type / required / default)	✅	Reusable workflows resolve caller-side with: values against the callee's inputs: declarations, honoring default: when a caller omits an input.
on.workflow_call.outputs.*.value	✅	Reusable-workflow callees can expose outputs.<name>.value: ${{ jobs.<id>.outputs.<name> }}, and callers read them the same way as a normal needs.*.outputs.*.
on (schedule, workflow_dispatch)	🟡	Accepted without error, but Agent CI does not simulate event triggers — workflows must be run manually
on (workflow_call)	✅	Local reusable workflows (uses: ./.github/workflows/...) are inlined into the caller's dependency graph. Remote refs are fetched from GitHub (requires --github-token or AGENT_CI_GITHUB_TOKEN). inputs:/outputs: passing and nested reusable workflows are supported
on (other events)	🟡	Covers every GitHub-Actions webhook event Agent CI does not simulate (issues, issue_comment, label, pull_request_review, release, registry_package, status, deployment, deployment_status, fork, watch, star, discussion, milestone, project, repository_dispatch, check_run, check_suite, workflow_run, etc.). The parser accepts the event name but the run is not triggered — the workflow only ever runs when invoked manually via pnpm agent-ci run.
env	✅	Workflow-level env is propagated to all steps
defaults.run.shell	✅	Non-bash shells (sh, python, pwsh) are invoked via a parse-time heredoc wrap because the runner ignores inputs.shell for script steps. bash is the default.
defaults.run.working-directory	✅	Passed through to the runner
permissions	🟡	Accepted but not enforced — the mock GITHUB_TOKEN has full access
concurrency	🚫	Concurrency groups are a GitHub-side queuing and cancellation mechanism. Agent CI has no persistent server to track group state across runs, so this cannot be implemented locally

Job-Level Keys

Key	Status	Notes
jobs.<id>	✅	Multiple jobs in a single workflow
jobs.<id>.name	✅
jobs.<id>.needs	✅	Jobs are sorted topologically into dependency waves
jobs.<id>.if	✅	success(), failure(), always(), cancelled(), ==/!=, &&/||, !, needs.*.outputs.*, needs.*.result
jobs.<id>.runs-on	⚠️	ubuntu-* / linux (and unspecified self-hosted / custom labels) run in a Linux container — the default image is minimal, see runner-image.md to add system tools. macos-* runs in a real macOS VM on Apple Silicon hosts with tart + sshpass installed; other hosts skip the job with a reason. windows-* is not yet supported and skips on every host. When a label declares a larger runner spec (e.g. ubuntu-latest-8-cores) than the local machine provides, the job is tagged degraded and a warning is printed before it starts — execution is not blocked, only annotated.
jobs.<id>.environment	🟡	Accepted but not enforced — environment protection rules are GitHub-side only. Object form (environment.name / environment.url) is also accepted and ignored.
jobs.<id>.permissions	🟡	Accepted but not enforced — job-level permission scopes have no effect because the mock GITHUB_TOKEN has full access to the local API emulation.
jobs.<id>.env	✅
jobs.<id>.defaults.run	✅	shell and working-directory
jobs.<id>.outputs	✅	Resolved after each job completes and accumulated across dependency waves
jobs.<id>.timeout-minutes	❌	Not implemented. Agent CI's pause-on-failure model is the intended way to handle long-running steps — a hard timeout would destroy the container state that makes local debugging possible
jobs.<id>.continue-on-error	❌	Not implemented. Agent CI pauses on failure so you can inspect and fix the container in place; continue-on-error would skip past failures and discard that debugging opportunity
jobs.<id>.concurrency	🚫	See workflow-level concurrency above
jobs.<id>.container	⚠️	Short and long form; image, env, ports, volumes honored. Within options:, only --env/-e and --label/-l are forwarded to the runner container; other Docker flags (--privileged, --user, --network, --cap-add, --workdir, etc.) are silently ignored because they tend to clash with agent-ci's own container orchestration.
jobs.<id>.container.credentials	❌	Registry username / password credentials for pulling a private container image are not parsed. Private images fail to pull unless the host's docker already has credentials for the registry (e.g. docker login was run out-of-band).
jobs.<id>.services	✅	Sidecar containers with image, env, ports, and options
jobs.<id>.services.*.credentials	❌	Per-service registry credentials are not parsed — see jobs.<id>.container.credentials. Private service images fail to pull unless the host's docker has credentials out-of-band.
jobs.<id>.uses (reusable workflows)	✅	Local refs (./) are expanded inline. Remote refs are fetched from GitHub (requires --github-token or AGENT_CI_GITHUB_TOKEN; public repos may work without auth). with: (inputs) and secrets: pass-through are supported
jobs.<id>.secrets	🚫	Agent CI cannot access GitHub's secret storage. Secrets are resolved from a .env.agent-ci file at the project root or from shell environment variables as fallback. --github-token / AGENT_CI_GITHUB_TOKEN auto-populates secrets.GITHUB_TOKEN. All are injected as ${{ secrets.* }} expressions

Strategy / Matrix

Key	Status	Notes
strategy.matrix	✅	Cartesian product of all array values is fully expanded
strategy.matrix.include	❌	Not implemented. The matrix parser only processes array-valued keys; include entries (which are objects) are silently dropped. Adding support would require post-processing the Cartesian product
strategy.matrix.exclude	❌	Not implemented — same reason as include. exclude entries are objects and are dropped by the array-only parser
strategy.fail-fast	✅	Setting fail-fast: false allows remaining matrix jobs to continue after a failure
strategy.max-parallel	❌	Not implemented. Parallelism is controlled by Agent CI's host-level concurrency limiter (based on CPU count), not per-workflow job limits

Step-Level Keys

Key	Status	Notes
steps[*].id	✅
steps[*].name	✅	Expression expansion in names
steps[*].if	⚠️	The condition is passed to the official runner binary, which evaluates it at runtime. Limitation: steps.*.outputs.cache-hit and similar outputs resolve to an empty string at parse time because prior steps have not yet run when the workflow is parsed
steps[*].run	✅	Multiline shell scripts with ${{ }} expression expansion
steps[*].uses	✅	Public actions are downloaded via the GitHub API
steps[*].uses (local ./)	✅	Local actions (e.g. uses: ./.github/actions/my-action) are resolved from the workspace
steps[*].uses (docker://…)	❌	Docker-image action references (uses: docker://alpine:3.19) are not resolved — agent-ci treats every uses: as a repository action. The step fails with an action-not-found error rather than running the image.
steps[*].with	✅	Expression expansion in values
steps[*].env	✅	Expression expansion in values
steps[*].working-directory	✅
steps[*].shell	✅	See defaults.run.shell above — the same heredoc-wrap mechanism applies.
steps[*].continue-on-error	❌	Not implemented — see jobs.<id>.continue-on-error above for the reasoning
steps[*].timeout-minutes	❌	Not implemented — see jobs.<id>.timeout-minutes above for the reasoning

Expressions (${{ }})

Key	Status	Notes
hashFiles(...)	✅	SHA-256 of matching files; supports multiple glob patterns. Descends into dotted directories (e.g. .github/) when the pattern asks for them.
format(...)	✅	Template substitution with recursive expression expansion
matrix.*	✅
secrets.*	✅	Resolved from .env.agent-ci at the project root, with shell environment variables as fallback. --github-token auto-populates secrets.GITHUB_TOKEN
vars.*	✅	Repository / organization / environment variables. Resolved from --var NAME=VALUE CLI flags or JSON supplied via --var-file <path> / --var-file -; unresolved vars.* references fail the run at pre-flight with a list of every missing name.
env.*	✅	The merged step environment (workflow-level → job-level → step-level, with step taking precedence) is available as env.* in with: inputs, run: scripts, name:, and env: value expressions. Not yet supported in if: conditions (the expression is evaluated by the runner, not the parser). Note: env.* references inside an env: block (one env var referencing another declared in the same block) are not evaluated — only the already-computed merged value is exposed.
inputs.*	✅	Reusable-workflow and (nominally) workflow_dispatch input values. Inside a called workflow, inputs.<name> resolves to the caller's with: value or the declared default:.
runner.os	✅	Linux for container-based jobs, macOS for jobs routed to the macOS VM
runner.arch	✅	X64 for container-based jobs, ARM64 for macOS VM jobs (Apple Silicon)
runner.name, runner.temp, runner.tool_cache, runner.debug, runner.environment	❌	Only runner.os and runner.arch are populated; the rest of the runner.* context resolves to an empty string. Workflows that read runner.tool_cache (common in tool-setup actions) will see an empty value — the actions themselves typically fall back to the RUNNER_TOOL_CACHE env var which the runner does set.
github.sha	✅	Real commit SHA from HEAD. When the working tree is dirty, a synthetic tree SHA is computed so that expressions that hash-seed from github.sha still differ between working-tree states.
github.repository, github.repository_owner	✅	Derived from the git remote origin URL (owner/repo and owner).
github.ref, github.ref_name, github.head_ref, github.base_ref, github.actor, github.run_id, github.run_number, and other github.* fields	⚠️	Returned as static defaults: ref_name / head_ref default to main, base_ref is empty, actor is the repo owner, run_id / run_number are 1, api_url / server_url point at Agent CI's local API emulation. Anything not in this row or one of the ones above resolves to an empty string — that covers action_path, workflow_ref, workflow_sha, triggering_actor, retention_days, secret_source, token, graphql_url, job, run_attempt, ref_protected, ref_type, env, path, and the rest of the context documented by GitHub.
github.event.*	⚠️	All event payload fields return empty strings — no real webhook event is triggered locally
strategy.job-total, strategy.job-index	✅
steps.*.outputs.*	⚠️	Always resolves to an empty string at parse time — the producing step has not run yet and the runner does not re-evaluate ${{ }} inside run: script bodies. Use needs.*.outputs.* for cross-job values instead.
steps.*.conclusion / steps.*.outcome	❌	Step result states are not exposed — neither conclusion nor outcome resolves against a meaningful value. Workflows that branch on prior-step status via these context variables will always evaluate as if the step didn't produce a result.
job.* (job.status, job.container.id, job.services.*)	❌	The per-job runtime context is not populated. Expressions that read job.status etc. resolve to empty strings.
* object-filter operator	❌	Array-of-object filtering like github.event.commits.*.author.name is not implemented — the entire expression resolves to an empty string.
needs.*.outputs.*	✅	Resolved after dependency jobs complete. The needs context is built from actual job outputs and passed into subsequent job evaluation
Boolean/comparison operators	✅	==, !=, <, >, <=, >=, &&, ||, !, parentheses. Case-insensitive string comparison. Numeric coercion follows GitHub Actions semantics: empty string and null become 0, numeric-looking strings are parsed, anything else becomes NaN and compares false.
toJSON, fromJSON	✅	toJSON emits 2-space-indented JSON matching GitHub Actions' pretty-printed output, so hashFiles/diff uses of toJSON(x) are reproducible.
contains(search, item)	✅	Case-insensitive substring for strings; element membership for JSON arrays (e.g. contains(fromJSON('["a","b"]'), 'b')). The * object-filter form (contains(labels.*.name, 'bug')) is not supported — see the * object-filter row.
startsWith(searchString, searchValue)	✅	Case-insensitive.
endsWith(searchString, searchValue)	✅	Case-insensitive.
join	✅	Joins JSON arrays with a separator (default , )
success(), failure(), always(), cancelled()	⚠️	success() / failure() / always() are evaluated against the accumulated job status in dependency waves. cancelled() always returns false locally — there is no cancellation signal for Agent CI to observe, so steps gated by if: cancelled() never run.
Type literals (true, false, null, numbers)	✅	Boolean, null, and numeric literals supported in expressions

GitHub API Features (DTU Mock)

Key	Status	Notes
Action downloads	✅	Action tarballs are resolved and downloaded from github.com
actions/cache	⚠️	Cache is stored on the local filesystem via bind-mount, giving ~0 ms round-trip on cache hits. Scoping is by cache key only — GitHub's ref-based isolation (a branch cannot read another branch's cache, default branch is a fallback) is not implemented. If two workflows on different refs share a key they share the cache.
actions/checkout	✅	The workspace is rsynced into the container with clean: false to preserve local changes
actions/setup-node, actions/setup-python, etc.	✅	Tool setup actions run natively inside the runner container
actions/upload-artifact / download-artifact	✅	Artifacts are stored on the local filesystem
GITHUB_TOKEN	⚠️	Not injected automatically. Pass --github-token or set AGENT_CI_GITHUB_TOKEN to make it available as secrets.GITHUB_TOKEN. GitHub API calls from the runner are answered locally by Agent CI's API emulation layer. OIDC ID-token issuance (permissions.id-token: write → federated cloud auth) is not implemented — actions that exchange the OIDC token with AWS/GCP/Vault will fail.
Workflow commands (::set-output::, ::error::, etc.)	✅	Handled by the official runner binary