fix(agent): make agent /tmp tmpfs size configurable via AGENT_TMP_SIZE (#1231)

[dolho]

Problem

Agent containers mount /tmp as a hardcoded 100 MB noexec,nosuid RAM-backed tmpfs (capabilities.py). It fills easily — e.g. leftover gh CLI install artifacts (~38 MB) plus normal scratch. Once full, every /tmp write fails with No space left on device, including git's commit scratch, so autonomous scheduled runs "complete" but silently fail to persist (commit/push). Because the size was a literal, operators couldn't tune it without a code change + base-image rebuild. Fleet-wide (base image).

#1098 already redirects $TMPDIR-honoring tools off /tmp, but install scripts that hardcode /tmp (the gh case) bypass it and still exhaust the cap.

Fix

• capabilities.py — AGENT_TMPFS_MOUNT size is now read from AGENT_TMP_SIZE (env on the backend service, which builds the agent mount spec), default 512m, validated ^\d+[mg]$ with empty/invalid → default. noexec,nosuid stay hardcoded — only size is configurable, and it stays bounded (tmpfs counts against the agent memory cgroup). Single source of truth → create (crud.py) and recreate (lifecycle.py) can't drift.
• Wire AGENT_TMP_SIZE=${AGENT_TMP_SIZE:-512m} on the backend service in both docker-compose.yml and docker-compose.prod.yml; document in .env.example.
• architecture.md Container Security — note the now-configurable size.

AGENT_TMP_SIZE	result
unset / "" / 512 / 512Mi / 0.5g	→ default 512m
256m, 2g, 1G	→ used (1G→1g)

Acceptance criteria

• size from AGENT_TMP_SIZE, default 512m, single source of truth (capabilities.py)
• noexec,nosuid remain hardcoded — only size configurable
• value validated (^\d+[mg]$); empty/invalid → default (never broken/unbounded)
• documented in .env.example + both compose files
• Container Security section of architecture.md updated
• deploy-time: verify inside a freshly recreated agent with df -h /tmp (mount specs are creation-time — existing agents pick up the new size on recreate, not restart)

Verification

tests/unit/test_1231_agent_tmp_size.py — default / valid m,g / case-fold / invalid→default / security-flags-never-dropped. Full unit run 22 passed (new + existing test_capability_set). Both compose files parse with AGENT_TMP_SIZE present; py_compile clean.

Out of scope (separate abilities-repo issue): the agent-side git-sync hook that reports success while silently failing to persist. Full /tmp is only the trigger.

Related to #1231

🤖 Generated with Claude Code

[github-actions]

⚠️ Nightly unit-suite check skipped — merge conflict against dev.

Resolve by running git merge dev locally and pushing the result. The next nightly run will re-test once the conflict is gone.