Mock Vs Real Execution
The `--mock` flag in `agentsmarket run` intercepts LLM API calls (returns canned responses) but **does not** intercept marketplace HTTP calls for `uses:` skill
Mock vs real execution (CLI --mock flag)
The --mock flag in agentsmarket run intercepts LLM API calls (returns canned responses) but does not intercept marketplace HTTP calls for uses: skill stages.
Why mock does not cover uses:
uses: stages fetch skills from https://api.agentsmarket.world/v1/skills/:id (or self-hosted URL via AGENTSMARKET_URL env var). Mock intercepts only the LLM provider layer; marketplace HTTP is a separate concern.
Implication for offline runs
agentsmarket run pipeline.yaml --mock requires:
- All LLM stages OK (mocked, no network)
- All
uses:stages still need network access (or self-hosted marketplace endpoint)
For fully-offline runs: --offline-skills flag is planned for v0.4 with skills.lock content-hash resolved at execute time. Not in v0.3.
What this means for tests
CI smoke tests run validate pipeline.yaml (validate-only, no uses: fetches). Pipeline-level integration tests that exercise uses: require network or self-hosted marketplace endpoint — see R11 docker image + R10.4 parallel integration suite.
Behavioral oracle (post-R11)
- Docker image with
--mockworks for validate + LLM-only stages (R11 docker-publish.yml). - CI smoke tests:
.github/workflows/ci.ymllines exercisingdocker run ... agentsmarket-runtime:dev validate ....