Nomarchy/agent/JOURNAL.md

Journal — append-only iteration log

One entry per iteration, newest first. This is the loop's short-term memory: the next session reads the last 3–5 entries to orient. Keep an entry under ~15 lines; durable lessons go to MEMORY.md, not here.

Template:

## YYYY-MM-DD — <task title> (<commit sha, or "no commit">)
- **Task:** which BACKLOG item (or QA sweep / bootstrap / escalation)
- **Did:** what changed, in 2–4 lines
- **Verified:** tier reached (V0/V1/V2) + the actual commands/checks run
- **Pending:** V3 queued? follow-ups filed? anything [stuck]?
- **Next suggestion:** what the following iteration should probably take

---

2026-07-04 — CI is green (run #58) — item 20 main half closed

• Task: monitor follow-up, no code. Run #58 (the .gitea path fix) succeeded — first green CI: the container recipe works on the live runner, eval tier confirmed end to end.
• Did: BACKLOG item 20 reduced to its stretch (KVM runner → enable vm-checks), moved to NEXT. Run #59 (earlyoom push) monitored.
• Next suggestion: NOW#4 (btrfs-assistant override attempt).

2026-07-04 — Memory-pressure protection shipped (loop iteration #3)

• Task: BACKLOG NOW#3 — default-on userspace OOM killer.
• Did: modules/nixos/oom.nix: earlyoom (mkDefault on, desktop notifications, session-plumbing --avoid list, no --prefer) + systemd.oomd.enable = false. Choice rationale: Hyprland session is one cgroup scope → oomd's cgroup-level kill would nuke the desktop; earlyoom is process-level. Found en route: nixpkgs default-enables oomd inert (enable=true, all slice toggles false). README "beyond the surface" note added.
• Verified: V0 (flake check) + V2 — new checks.oom-protection runNixOSTest, run for real: hog (chunked allocator, 686 MB peak in a 1 GB VM) SIGTERM'd by earlyoom in 0.1 s, bystander unit survived, oomd asserted inactive. 25 s test, cheap enough for the suite.
• Pending: nothing hardware-specific (behavior fully VM-provable).
• Next suggestion: NOW#4 (btrfs-assistant override attempt). Also: CI run #58 (the .gitea fix) was running at commit time — check its outcome next tick; this push triggers #59.

2026-07-04 — CI fix: wrong workflow dir (correction to iteration #2)

• Task: Bernardo reported no run despite an active runner + shared his compose file: the server is Gitea (gitea/act_runner), not Forgejo — I'd shipped the workflow to .forgejo/workflows/, which Gitea never reads. The legacy repo's .gitea/workflows/ path was the clue I under-weighted.
• Did: git mv → .gitea/workflows/check.yml; corrected the Forgejo references in the workflow header, TESTING.md §1b, BACKLOG item 20 (now: watch the first run; runner was never the problem); MEMORY.md gains the Gitea-not-Forgejo line.
• Verified: the push of this commit is itself the test — the run must appear; monitoring.
• Pending: first green run (item 20); runner-container memory headroom is the watch-out.

2026-07-04 — CI checks-on-push shipped (loop iteration #2)

• Task: BACKLOG NOW#2 — Forgejo Actions workflow.
• Did: .forgejo/workflows/check.yml (push to main/v1 + dispatch): flake check --no-build, theme-sync py_compile, bash -n over tracked .sh. Scoped to the eval tier deliberately — API probing + archaeology of the legacy .gitea/workflows/check.yml (57 runs on this instance) showed the runner is an act_runner docker container (no KVM/systemd), so the VM suite can't run there; a commented vm-checks job documents the KVM-runner upgrade path. Legacy's hard-won container gotchas (nixbld setup, sandbox=false for Stylix IFD, Nix 2.31.5 pin) carried over verbatim. TESTING.md §1b added.
• Verified: V0 locally (same commands the workflow runs, minus the container Nix install) + yq YAML parse. The real green run needs the runner to be alive — not confirmable from here (runners API is 401 unauthenticated); watching the run after push.
• Pending: new [human] item 20 — confirm/re-register the runner; stretch: a KVM runner unlocks the vm-checks job.
• Next suggestion: NOW#3 (memory-pressure protection).

2026-07-04 — Opt-in auto-commit shipped (loop iteration #1)

• Task: BACKLOG NOW#1 — auto-commit of state mutations (in-flake state Phase 4). First autonomous /loop iteration.
• Did: settings.autoCommit live-read by nomarchy-theme-sync; pathspec-limited commit of theme-state.json on apply/set (fires if flag on before OR after the write → the off-toggle commits too); identity fallback; same-value no-op; bg excluded. Menu: System › Auto-commit (instant, self-gated on .git). Rider: get prints booleans as JSON true/false, fixing the stuck "Auto timezone (on/off)" menu label. BACKLOG numbering switched to stable IDs (no renumbering).
• Verified: V0 (py_compile, flake check) + V1 (HM generation builds; generated menu bash -n) + a 7-assertion sandbox git-repo round-trip (enable/set/same-value/disable/post-disable/apply/identity-fallback).
• Pending: V3 queued in HARDWARE-QUEUE (on-machine toggle + apply).
• Next suggestion: NOW#2 (CI checks-on-push) — but its runner half may need Bernardo to register a Forgejo runner; deliver the workflow regardless.

2026-07-04 — Backlog revision pass (human-requested)

• Task: Bernardo asked for a philosophy-aligned revision of BACKLOG.md ("I'm sure I'm forgetting important items").
• Did: verified guessed gaps against the tree first (GC/optimise/boot limit + all 21 previews already ship — dropped those ideas). Added the confirmed gaps: CI-on-push promoted to NOW#2 (agents push to main now), memory-pressure protection NOW#3 (nothing mitigates OOM; the release bump died to one), viewers+mime defaults #8, update/rollback UX #9, nomarchy-doctor #10, state-file validation #11, screen recording #12, niceties batch #13 (idle-inhibit, low-battery notify, hyprpicker), OCR in LATER; recovery runbook folded into the docs item; zram + default browser added to Decisions.
• Verified: V0 (docs-only).
• Pending: Bernardo to sanity-check the new NOW ordering, esp. CI at #2 and the two new Decisions.
• Next suggestion: unchanged — NOW#1 (opt-in auto-commit).

2026-07-04 — Bootstrap the loop infrastructure (this commit)

• Task: create agent/ (human-requested): LOOP/GOALS/BACKLOG/ CONVENTIONS/MEMORY/HARDWARE-QUEUE/JOURNAL, root CLAUDE.md, README + ROADMAP pointers.
• Did: reworked the forward half of docs/ROADMAP.md into the tiered BACKLOG (5 NOW / 6 NEXT / LATER / Decisions); collected every pending on-hardware check into HARDWARE-QUEUE.md; seeded MEMORY.md from the ROADMAP's decision records.
• Verified: V0 (docs-only change; nix flake check --no-build run to confirm the tree still evaluates).
• Pending: Bernardo to sanity-read GOALS.md pillars + BACKLOG tiering — reorder freely, the order is the instruction.
• Next suggestion: BACKLOG NOW#1 (opt-in auto-commit — Phase 4 of in-flake state, already the agreed next step).