Handover: ADF gate hardening + repo-local autonomous agents

Date: 2026-06-04 Focus: Gitea epic #1910 follow-ups -- harden the ADF pre-merge gate and enable ADF to autonomously maintain the polyrepos via repo-local agent config. Orchestrator: bigbox adf-orchestrator.service -- healthy/active on a binary built from terraphim-agents main (includes #2174, #2199, #2203).

1. Progress summary

Shipped, merged, gated, deployed

| Issue | What | Where | |-------|------|-------| | #2174 | AutoMerge re-fetches required commit statuses + re-runs reconcile_pr_gate immediately before merge_pr (closes a same-SHA status-regression race). Closed. | terraphim-agents PR #5 | | #2175 | Verdict-agent gate replicated to ALL 6 polyrepos: extra_projects (multi-project agent registration) + multi-repo spawn/status routing (re-point def.project to the PR's project; resolve owner/repo from project gitea) + synchronized PR-action webhook fix. Branch protection on all 6 now requires native-ci + adf/pr-reviewer + adf/validation + adf/verification. Closed. | terraphim-agents PRs #6, #8, #11 | | #2199 | Review/CI agents (pr-reviewer/validator/verifier, build-runner) no longer auto-commit their working tree (they post a status, not code) -- gated on commit_status_post. | terraphim-agents PR #12 | | #2203 | Repo-local .terraphim/adf.toml agents loaded into the live fleet via OrchestratorConfig.project_sources; last_cron_fire re-keyed to (project, name) for per-repo cron. | terraphim-agents PR #13 | | docs | adf rebuild/redeploy procedure added to docs/adf/operations.md. | terraphim-ai PR #2201 (merged to main) |

Canary PROVEN end-to-end (#2203)

On terraphim-agents: issue #15 -> repo-local implementation-swarm fired on its :30 cron (per-project) -> branched task/15-impl off origin/main -> implemented the exact test asked for -> opened PR #16 on terraphim-agents -> 4-check verdict gate green -> merged (a5cc3c6) -> #15 auto-closed. ADF can now autonomously maintain a polyrepo from that repo's own config.

Working

• All 6 polyrepos enforce the 4-check verdict gate; coordinator auto-merges only at 5/5 confidence (else human review -- verified on terraphim-ai PR #2201).
• Repo-local agent loading is live (terraphim-agents .terraphim/adf.toml parsed into the fleet when its project_source is present).

Paused / blocked

• terraphim-agents implementation-swarm is PAUSED: its [[project_sources]] entry was removed from /opt/ai-dark-factory/orchestrator.toml and the orchestrator restarted, per "stop for now". It will not fire. terraphim-agents/.terraphim/adf.toml remains in the repo (dormant). terraphim-agents #1 (dead-code, 24 suppressions) is open + unblocked + untouched -- it was the agent's next pick.

2. Technical context

• Working checkout: /tmp/agents-fix (terraphim-agents). Research/design docs in /tmp/agents-fix/.docs/{research,design}-implementation-swarm-portability.md, .docs/*-2116-premerge-gate.md.
• Repos: code = terraphim/terraphim-agents (the terraphim_orchestrator crate, builds the adf binary); issue tracker + docs = terraphim/terraphim-ai. Both on git.terraphim.cloud.
• Key code (terraphim_orchestrator/src): config.rs (ProjectSource + merge_project_sources; gitea_owner_repo_for_project; extra_projects validation), agent_registry.rs (extra_projects registration), pr_handlers_impl.rs (ReviewPr dispatch repo-routing + status posts), reconcile_impl.rs (#2199 auto-commit gate; terminal status routing), scheduling_impl.rs
  • lib.rs (last_cron_fire (project,name)), webhook.rs (synchronized action), auto_merge_impl.rs (#2174 re-check), scheduler.rs.
• Deploy procedure: docs/adf/operations.md -> "Rebuilding and redeploying the adf binary" (build from a clean worktree at origin/main with shared CARGO_TARGET_DIR; adf --check before restart; atomic-rename the busy binary; rollback .bak). Backups on bigbox: /usr/local/bin/adf.bak-*, /opt/ai-dark-factory/orchestrator.toml.bak-*.

3. Outstanding

1. Resume / expand repo-local agents (#2203 rollout) -- canary done; extend to the other 5 polyrepos: per repo add a [[project_sources]] (id + root) to orchestrator.toml + commit a .terraphim/adf.toml defining its implementation-swarm, with STAGGERED schedules (avoid native-runner starvation #2185). The project (gitea/token) must stay central in conf.d/<repo>.toml; do NOT redeclare the shared verdict agents (they are central via extra_projects) or the dup-check refuses startup.
2. #2185 -- native runner: stuck push-runs, double-fetch with 2 instances, terraphim-ai starving the shared org runner. P2.
3. #2076 -- Firecracker route; also the hardening for untrusted repo-local sources.
4. Robustness: consider making merge_project_sources fail-open (skip+warn on a duplicate) rather than refuse startup -- the #2203 deploy crash-looped on a stale terraphim-ai project_source (its agents are central; removed it).
5. Carried backlog (re-triage via gtr ready): dead-code epic #2072, #2077, #2078, #2176, ~34 conflicted fleet PRs.

4. Resume recipe

Re-enable the terraphim-agents autonomous implementer:

ssh bigbox
F=/opt/ai-dark-factory/orchestrator.toml
sudo cp -p "$F" "$F.bak-resume"
printf '\n[[project_sources]]\nid = "terraphim-agents"\nroot = "/data/projects/terraphim/terraphim-agents"\n' | sudo tee -a "$F"
cd /opt/ai-dark-factory
GITEA_TOKEN=$(systemctl cat adf-orchestrator.service | grep -oP 'Environment=GITEA_TOKEN=\K\S+') \
  /usr/local/bin/adf --check orchestrator.toml   # expect routing table, terraphim-agents implementation-swarm
sudo systemctl restart adf-orchestrator.service

It will fire at :30 and pick the top unblocked terraphim-agents issue (currently #1 dead-code). Gate protects any merge.

5. Memory written this session

project_adf_rebuild_deploy_procedure, project_adf_agent_task_semantics (cli_tool runs the task as a PROMPT, not bash; repo-local loading rules), project_native_gitea_runner_quirks, feedback_fmt_check_before_native_ci_push. Index in ~/.claude/projects/.../memory/MEMORY.md.