bug: TOML-driven agent services lack FACTORY_REPO env and projects/env/state volume mounts — sidecar silently never polls #855

New issue

Closed

opened 2026-04-16 10:23:46 +00:00 by dev-bot · 0 comments

dev-bot commented

2026-04-16 10:23:46 +00:00

Collaborator

Problem

lib/generators.sh (TOML-driven [agents.X] → compose service emission) omits two things the main agents service has, both required for the agent to actually do any work:

No FACTORY_REPO env var. Without it, /entrypoint.sh never executes the "Factory bootstrap: DISINTO_DIR switched to live checkout" branch — the container stays pinned on the baked /home/agent/disinto directory.
No ./projects:/home/agent/disinto/projects:ro (and ./.env, ./state) volume mount. Without it, the baked projects/ dir only contains *.toml.example files (no real disinto.toml).

Combined, the polling loop for toml in ${DISINTO_DIR}/projects/*.toml matches zero files, the loop body never runs, no dev-poll ever spawns, and the container sits in an apparently-healthy sleep loop doing nothing. No error is logged — it just silently does nothing forever.

Confirmed today on disinto-agents-llama: container was up, entrypoint logged "Entering polling loop (interval: 60s, roles: dev)", sleep 60 was running — but no Processing project TOML / Running dev-poll (iteration N) lines ever appeared.

Repro

disinto hire-an-agent dev-qwen dev --local-model <url> --model <name>.
Set COMPOSE_PROFILES=agents-llama in .env (to work around #845).
disinto up.
docker exec disinto-agents-llama ls /home/agent/disinto/projects/ → only *.toml.example files.
docker exec disinto-agents-llama bash -c 'echo $FACTORY_REPO' → empty.
docker exec disinto-agents-llama tail /home/agent/data/agent-entrypoint.log → no Processing project TOML entries.
Poll cadence effectively dead; no work picked up; GPU usage stays at 0.

Why subtle

The container is "running" per docker ps (healthcheck only pgreps entrypoint.sh, which is alive).
Startup log shows all the expected "Entering polling loop" messages.
No error is raised — the bash for loop over a zero-match glob just exits silently.
It looks identical to "idle because no ready issues" until you realise it never polled in the first place.

Dev-qwen ran for days on the legacy ENABLE_LLAMA_AGENT=1 service block (which DID include the projects mount). Switching activation paths via hire-an-agent silently downgraded the sidecar to a zombie.

Fix

In _generate_local_model_services (or whichever helper emits per-agent services), emit:

volumes:
  - ./projects:/home/agent/disinto/projects:ro
  - ./.env:/home/agent/disinto/.env:ro
  - ./state:/home/agent/disinto/state
environment:
  FACTORY_REPO: ${FORGE_REPO:-disinto-admin/disinto}

Exactly matching the main agents service block (which works correctly).

Secondary: add a generate-time (or entrypoint-time) assertion that $DISINTO_DIR/projects/*.toml matches at least one real file, and fail loudly otherwise. The silent-zombie mode is the worst kind of failure.

Acceptance

After hire-an-agent X, the generated service for X has the three volume mounts and FACTORY_REPO env.
disinto up --profile agents-X brings up a sidecar that logs Processing project TOML and Running dev-poll within the first poll interval.
Entrypoint fails loudly (non-zero exit, clear log) if the projects glob matches zero real TOML files.

Related: #845 (profile drift), #846 (activation paths), #847 (secrets), #852 (hyphen env-var), #853 (image ref).

## Problem `lib/generators.sh` (TOML-driven `[agents.X]` → compose service emission) omits two things the main `agents` service has, both required for the agent to actually do any work: 1. **No `FACTORY_REPO` env var.** Without it, `/entrypoint.sh` never executes the "Factory bootstrap: DISINTO_DIR switched to live checkout" branch — the container stays pinned on the baked `/home/agent/disinto` directory. 2. **No `./projects:/home/agent/disinto/projects:ro` (and `./.env`, `./state`) volume mount.** Without it, the baked `projects/` dir only contains `*.toml.example` files (no real `disinto.toml`). Combined, the polling loop `for toml in ${DISINTO_DIR}/projects/*.toml` matches zero files, the loop body never runs, no dev-poll ever spawns, and the container sits in an apparently-healthy sleep loop doing nothing. No error is logged — it just silently does nothing forever. Confirmed today on `disinto-agents-llama`: container was up, entrypoint logged "Entering polling loop (interval: 60s, roles: dev)", `sleep 60` was running — but no `Processing project TOML` / `Running dev-poll (iteration N)` lines ever appeared. ## Repro 1. `disinto hire-an-agent dev-qwen dev --local-model <url> --model <name>`. 2. Set `COMPOSE_PROFILES=agents-llama` in `.env` (to work around #845). 3. `disinto up`. 4. `docker exec disinto-agents-llama ls /home/agent/disinto/projects/` → only `*.toml.example` files. 5. `docker exec disinto-agents-llama bash -c 'echo $FACTORY_REPO'` → empty. 6. `docker exec disinto-agents-llama tail /home/agent/data/agent-entrypoint.log` → no `Processing project TOML` entries. 7. Poll cadence effectively dead; no work picked up; GPU usage stays at 0. ## Why subtle - The container is "running" per `docker ps` (healthcheck only pgreps `entrypoint.sh`, which is alive). - Startup log shows all the expected "Entering polling loop" messages. - No error is raised — the bash `for` loop over a zero-match glob just exits silently. - It looks identical to "idle because no ready issues" until you realise it never polled in the first place. Dev-qwen ran for days on the legacy `ENABLE_LLAMA_AGENT=1` service block (which DID include the projects mount). Switching activation paths via `hire-an-agent` silently downgraded the sidecar to a zombie. ## Fix In `_generate_local_model_services` (or whichever helper emits per-agent services), emit: ```yaml volumes: - ./projects:/home/agent/disinto/projects:ro - ./.env:/home/agent/disinto/.env:ro - ./state:/home/agent/disinto/state environment: FACTORY_REPO: ${FORGE_REPO:-disinto-admin/disinto} ``` Exactly matching the main `agents` service block (which works correctly). Secondary: add a generate-time (or entrypoint-time) assertion that `$DISINTO_DIR/projects/*.toml` matches at least one real file, and fail loudly otherwise. The silent-zombie mode is the worst kind of failure. ## Acceptance - After `hire-an-agent X`, the generated service for `X` has the three volume mounts and `FACTORY_REPO` env. - `disinto up --profile agents-X` brings up a sidecar that logs `Processing project TOML` and `Running dev-poll` within the first poll interval. - Entrypoint fails loudly (non-zero exit, clear log) if the projects glob matches zero real TOML files. Related: #845 (profile drift), #846 (activation paths), #847 (secrets), #852 (hyphen env-var), #853 (image ref).