Hiring a second llama-backed dev agent (e.g. `dev-qwen2`) alongside
`dev-qwen` tripped four defects that prevented safe parallel operation.
Gap 1 — hire-agent keyed per-agent token as FORGE_<ROLE>_TOKEN, so two
dev-role agents overwrote each other's token in .env. Re-key by agent
name via `tr 'a-z-' 'A-Z_'`: FORGE_TOKEN_<AGENT_UPPER>.
Gap 2 — hire-agent generated a random FORGE_PASS but never wrote it to
.env. The container's git credential helper needs both token and pass
to push over HTTPS (#361). Persist FORGE_PASS_<AGENT_UPPER> with the
same update-in-place idempotency as the token.
Gap 3 — _generate_local_model_services hardcoded FORGE_TOKEN_LLAMA for
every local-model service, forcing all hired llama agents to share one
Forgejo identity. Derive USER_UPPER from the TOML's `forge_user` field
and emit \${FORGE_TOKEN_<USER_UPPER>:-} per service.
Gap 4 — every local-model service mounted the shared `project-repos`
volume, so concurrent llama devs collided on /_factory worktree and
state/.dev-active. Switch to per-agent `project-repos-<service_name>`
and emit the matching top-level volume. Also escape embedded newlines
in `$all_vols` before the sed insertion so multi-agent volume lists
don't unterminate the substitute command.
.env.example documents the new FORGE_TOKEN_<AGENT> / FORGE_PASS_<AGENT>
naming convention (and preserves the legacy FORGE_TOKEN_LLAMA path used
by the ENABLE_LLAMA_AGENT=1 singleton build).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
--poll-interval was incorrectly written as compact_pct in the project TOML,
misconfiguring CLAUDE_AUTOCOMPACT_PCT_OVERRIDE instead of polling behavior.
Now compact_pct is hardcoded to 60 (the correct default) and poll_interval
is a separate TOML field emitted as POLL_INTERVAL in the compose service.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
