Give lab_setup fragments somewhere to run

[lpezet]

Closes #107. The gap is exactly as reported; two things in the framing are worth correcting, and one detail of the suggested shape doesn't survive contact.

What was actually wrong

lab_setup is the one part of a bank entry template/deployment/ had no mechanism for. ./lab is a build context, the Dockerfile COPYed nothing, and command: sleep infinity wouldn't have run a fragment if one had been there. So github and gcp installed cleanly and did not work.

Correction: this is not a boundary failure

The issue's strongest claim is that gh without git_protocol https "quietly stops being mediated". Under the default it doesn't — it fails closed. The lab network ships internal: true, so there is no default gateway and SSH has no route out at all. The repo already says so in both places that matter: template/deployment/compose.yaml:249 names ssh as the reason to set LAB_INTERNAL=false, and PLAYBOOK.md:275-279 says the symptom of needing that opt-out is DNS failure inside the lab.

The claim holds only under LAB_INTERNAL=false — where egress mediation is knowingly off, check-invariants.sh reports it every run, and gh is not special because everything bypasses.

That changes the severity, not the fix. What's true is that two of four bank entries install non-functional, and nothing says so.

Correction: the mount can't be ./lab

./lab is the build context and now holds entrypoint.sh — a *.sh glob there would pick up the entrypoint and run it inside itself. Fragments live in ./lab/setup.d/, mounted at /etc/agent-setup.d, which also mirrors cred-gateway/gateway.d/.

Executed, not sourced

Both fragments' headers claimed "Sourced by the deployment's setup.sh". They shouldn't be, and don't need to be — every effect either one has is a file it writes (git config --global, the inert ADC document), and neither exports anything.

Sourcing would put each fragment's set -euo pipefail into the entrypoint's shell and let a stray exit kill it. And it would buy nothing even for a fragment that wanted it: docker exec builds its environment from the container's config, not from PID 1, so an export from an entrypoint never reaches an agent that shells in. A fragment needing an agent-visible variable writes /etc/profile.d/, or the entry declares lab_env.

Consequences: headers corrected, and bank/gcp/lab/setup.sh goes 644 → 755 so the two entries stop differing over a bit the entrypoint deliberately doesn't depend on (it invokes bash explicitly, and a mount can flatten the bit anyway).

Shape

• In the image, both lab Dockerfiles, kept identical as the existing copy-diff lint requires. A fragment nothing runs is not an installed provider, and a deployment shouldn't have to supply that mechanism itself.
• One file per provider, lab/setup.d/<name>.sh — the path inside an entry is fixed (lab/setup.sh), so two would collide. This is what makes uninstall rm rather than finding and unpicking lines in a shared file, which is the fragile version of "hooks".
• On start, not on create — adding a provider is a file plus a restart, never a rebuild.
• Read-only mount. A fragment can already do anything the lab can do; persisting a change back into the deployment is a different thing.
• A failing fragment stops the container, rather than coming up half-configured.

Tests

New tests/integration/80-lab-setup.test.sh, 18 assertions. It runs the real entrypoint.sh in the real container shape on a small base — the actual lab image is typescript-node:22 at 1.8GB, which this tier won't pull, and which is why tests/stacks/20-boundary already refuses to start the lab at all. 00-config-lint asserts the part that test can't see: that both Dockerfiles genuinely COPY the script and set ENTRYPOINT/CMD, so the mechanism can't be deleted from the image with every behavioural test still green.

One case worth calling out, because it's the path most deployments take: an empty setup.d/. printf '%s\n' dir/*.sh under nullglob still emits one empty line, so an empty directory read as a single fragment named "" and refused to start. Caught while testing, fixed, and covered.

Local runs of all four CI jobs: integration 17 suites green, stacks 10 and stacks 20 green.

Targets release/1.14.0 as a feature.

🤖 Generated with Claude Code

https://claude.ai/code/session_01HZotg9g1oPFfkJjGGQLRhJ