lux/docs/plans/workflow-system.md

# Workflow system execution

Source: [`../workflow-discovery.md`](../workflow-discovery.md)

Apply the global and repository-local `AGENTS.md` files and the ponytail skill
at full intensity throughout. Preserve user work, prefer native capabilities,
and add no abstraction without a demonstrated repeated need.

## Non-negotiable activation boundary

Only the user may switch or activate configurations. The agent must never run
`nixos-rebuild switch`, a Home Manager activation or switch command, or any
equivalent operation that changes the active generation. Building and
evaluating configurations is allowed; after a successful build, give the user
the exact activation command and wait for confirmation.

On 2026-09-05 the agent violated this boundary by directly running a built Home
Manager activation package. It restarted the Kitty service and closed the
user's terminal session. Do not repeat this, including for changes described as
user-level, harmless, temporary, or needed only for live verification.

## Execution discipline

Advance every safe independent lane before asking the user to act. Batch
declarative changes behind one user-controlled activation and consolidate
empirical, credential, hardware, and approval gates into one checklist only
after no autonomous work remains. A single pilot reaching its user gate does
not make it the global next step while another pilot can still progress.

## Current checkpoint

- Tranche: autonomous consolidation across repositories, tools, and pilots
- State: autonomous implementation exhausted; consolidated gates await the user
- Branch: `chore/workflow-system`
- Baseline: `nix flake check --no-build --quiet` and a no-link Zenith system
  build passed on 2026-09-05 after the consolidated workstation retirement and
  `rclone` preparation; only the expected dirty-tree and unknown `modules`
  warnings were emitted. The generated project menu finds 30 live repositories
  without `.archive` or `.worktrees` entries, and the generated daily-note
  configuration passes an isolated same-day idempotence test. The managed Actual
  and WhatsApp entries are active in the current user session, Evolution's
  source registry activates over D-Bus, and the VPN reconnects successfully.
  The user activated the terminal root fix: `xdg-terminal-exec` now selects
  Kitty correctly and Yazi
  launches from Walker. After restarting the stale Elephant and Walker
  processes, repository launch also succeeds and exiting Neovim leaves a shell
  in the repository. The daily-note launcher and migrated notebook contents
  also pass their live sanity check. The attempted Elephant menu restart trigger
  stopped its dependent Walker during activation without bringing Walker back,
  so it has been removed in favor of applying rare menu changes at login. After
  activating that removal and relogging, Walker opens normally and repository
  selection starts Neovim in Oil at the project root. Four fresh local clones
  pass their flake checks, expose their tracked command catalogs, and remain
  clean after shell entry. The dashboard contract additionally passes its full
  check and a bounded two-process development smoke test. The dashboard and
  traceability PostgreSQL stacks also ran concurrently while both repositories'
  checks passed; all test processes and containers were then stopped.
- Next: present one consolidated user batch covering the built activation and
  the independent empirical, credential, hardware, and approval gates below.

## Progress

| Batch | State | Evidence or blocker |
| --- | --- | --- |
| Bootstrap and repository ownership | Waiting for user | Repository and dashboard reconciliation complete; credential cleanup remains |
| Declarative workstation foundation | Complete | All-host builds pass with one Nixpkgs pin, `just`, and an explicit Podman Compose provider |
| Repository contracts and workbench | Waiting externally | Four clean-checkout contracts and read-only Kubernetes pass; data-science overlaps uncommitted work, while API, database, report, and secret inputs remain externally gated |
| Desktop navigation and interaction | Active | Walker project and terminal actions plus stable Actual/WhatsApp launchers work; Outlook and two-output checks remain |
| Agent workflow | Waiting for user | The `nixos-maintainer` skill validates and matches its live owner; external policy blocks nested Codex evaluation without approval, and Herdr awaits its first live check |
| Notes | Complete | The daily launcher and migrated notebook pass live checks; the approved rollback copy is recoverable from Trash |
| PIM | Active | Native NixOS Evolution 3.60.2 plus Microsoft 365 integration is live beside Thunderbird; account and daily-use gates await credentials |
| Transcription | Pending | Requires a real Fossify recording-transfer pilot before implementation moves from Zenith to Orion |
| Conditional retirement and completion | Active | Superseded HTTP/database/note tools are removed with earlier retirements; Bruno, DBeaver, image editors, and PIM clients remain behind separate gates |

## Repository checkpoints

| Repository | Branch | Commit | Purpose |
| --- | --- | --- | --- |
| NixOS configuration | `chore/workflow-system` | `41bc61b` | Discovery, paths, Git identity, and Codex defaults |
| NixOS configuration | `chore/workflow-system` | `69552b7` | Simplify flake inputs and declare workstation lifecycle tools |
| NixOS configuration | `chore/workflow-system` | `0c121a6` | Declare Garden tooling and the local workbench |
| NixOS configuration | `chore/workflow-system` | `2fb6061` | Package Garden completion declaratively |
| NixOS configuration | `chore/workflow-system` | `ad01b9b` | Enable NetworkManager's native OpenVPN plugin |
| NixOS configuration | `chore/workflow-system` | `c9ee688` | Advance VPN, desktop, agent, note, and PIM pilots |
| NixOS configuration | `chore/workflow-system` | `cfa5d6d` | Correct PWA ownership and Evolution service activation |
| NixOS configuration | `chore/workflow-system` | `c2f48df` | Restore terminal-backed Walker actions |
| NixOS configuration | `chore/workflow-system` | `6631889` | Open Walker projects at their repository root |
| NixOS configuration | `chore/workflow-system` | `ca65328` | Remove the unsafe Elephant restart trigger |
| NixOS configuration | `chore/workflow-system` | `d085512` | Advance from Walker to the agent pilot |
| NixOS configuration | `chore/workflow-system` | `ee64a7b` | Retire superseded tools and prepare the S3 pilot |
| `yookr-dashboard-new` | `feat/gogv2` | `934605d` | Remove obsolete dashboard converter |
| `yookr-dashboard-new` | `feat/gogv2` | `8e6e4bc` | Establish its project contract and native lifecycle |
| `rules-service` | `chore/workflow-contract` | `555d9cd` | Establish the first common repository contract |
| `yookr-api` | `chore/workflow-contract` | `a76b2b7` | Establish the HTTP and OpenAPI contract |
| `traceability-data-model` | `chore/workflow-contract` | `33cb7ff` | Establish the database and schema contract |
| `traceability-data-model` | `chore/workflow-contract` | `774e710` | Remove Mathesar and simplify the local stack to PostgreSQL |
| `yookr-workbench` | `main` | `badc61c` | Create the workbench and isolate its Garden session |

## Deferred or externally blocked actions

None of these block work on independent pilots. Surface the exact applicable
step to the user only when it blocks the current task.

- Activate the consolidated Zenith build with `cd ~/.config/nixos && nh os
  switch`. This removes the retired tools and installs `rclone`; only the user
  may run it.
- In a separate terminal, enter this repository and run `herdr`, start `codex`,
  and ask it to read `docs/plans/workflow-system.md` without editing and report
  the exact next implementation step. Confirm Herdr shows the agent's working
  and waiting/done states, detach with `Ctrl+B` then `q`, reattach with `herdr`,
  and confirm the pane survived with its output intact.
- Run the bounded Pinta trial using a recent real edit: clipboard/open,
  selection, crop/resize/canvas, text/drawing/shapes, layers/transparency,
  basic adjustments, and PNG/JPEG export. Tell the agent whether it fully
  replaces GIMP; one of the two will then be removed.
- Install Outlook from its signed-in page as a Vivaldi web app, then tell the
  agent so its generated desktop entry can be identified and replaced by a
  stable Home Manager owner.
- During a docked Zenith session, tell the agent the external display is
  attached so its stable identity and the two-output `Super+A` pilot can be
  completed. The current session exposes only `eDP-1`.
- After activation, launch Evolution and add accounts without removing any
  existing client or changing authority: personal and old-personal through
  their existing IMAP/SMTP service, university and work through the
  `Microsoft 365` account type, and Radicale through native CalDAV, CardDAV,
  and VTODO. Complete tenant consent/sign-in when prompted, then tell the agent
  which accounts connected; the full mail/calendar/contact/task gate remains
  a daily-use pilot rather than a one-time setup claim.
- On Android, complete the DAVx5/Radicale and Tasks.org pilots described in the
  discovery plan, including correct default accounts, recurrence, reminders,
  and Microsoft To Do within its documented limits.
- Make one Fossify recording on the Pixel and let Syncthing transfer it. Do not
  rename or move it manually; the agent must first observe the real write and
  settlement behavior before implementing Orion automation.
- Reauthenticate the GitHub CLI with `gh auth login -h github.com`; its current
  `kirigar` token is invalid. The agent can then complete the read-only PR and
  isolated review-worktree gate without more direction.
- Explicitly authorize the five ephemeral, read-only `codex exec` skill-scope
  checks if sending this private repository to another authenticated Codex
  session is acceptable. External policy correctly rejected them without that
  approval; static skill validation already passes.
- Rotate the exposed Context7 key and provide `CONTEXT7_API_KEY` through a
  mode-0600 machine-local environment file on each workstation before removing
  the static user-config header.
- Relocate the mode-0644 `/home/kiri/work/SECRETS` file to an approved private
  owner and restrict it; its contents have not been inspected.
- Move the ignored `rules-service/.env` to
  `~/.config/yookr/rules-service.env` with mode 0600; its contents have not been
  inspected or copied.
- Move the ignored `yookr-dashboard-new/.env` to its approved machine-local
  mode-0600 owner, then update `.envrc`; the non-secret repository contract is
  already complete.
- Commit or otherwise hand off the extensive dirty `yookr-data-science` work
  before its flake, live-fetch defaults, local credentials, and missing project
  contract are changed. Those files overlap the active work directly.
- Explicitly approve moving the stale `~/.config/codex` tree to Trash if its old
  authentication, conversations, databases, and plugins are no longer needed.
- Explicitly approve removing the orphaned Podman volume
  `traceability-data-model_mathesar-secrets` and the old `mathesar_django`
  database if no Mathesar rollback is wanted. They are disconnected from the
  repository but preserved because deletion would be irreversible.
- Supply one real Yookr S3 task plus a least-privileged remote and authorize one
  bounded upload. `rclone` is prepared, but no remote or wrapper is guessed.
- Restore access to the locked `yookrdev/constant-store` commit or replace that
  retired dependency deliberately, then rerun `just check` in `yookr-api`.
- Deploy a GET image that exposes `/full-api-docs/openapi.json`, then rerun
  `just openapi-refresh test`; the current test deployment returns 404.
- Populate the ignored API HTTP credentials and perform the planned login plus
  harmless administrator read; no credential-bearing request has been made.

## Deferred Polaris replication

Do not access or test Polaris during this implementation. Treat a Zenith pass
as representative, but retain these machine-local steps for later execution on
Polaris:

1. Activate this NixOS configuration so `networkmanager-openvpn` is installed.
2. Run `chmod 600 ~/sync/yookr-vpn.ovpn`.
3. Run `nmcli connection import type openvpn file ~/sync/yookr-vpn.ovpn`.
4. Rename the imported `yookr-vpn` connection to `Yookr VPN`, disable
   autoconnect, and add the required legacy cipher:
   `nmcli connection modify yookr-vpn connection.id "Yookr VPN" connection.autoconnect no +vpn.data data-ciphers=AES-256-CBC`.
5. Connect with `nmcli connection up "Yookr VPN"`.

## Decisions and discoveries

- Use `/home/kiri/work/repos/yookr-dashboard-new` as the reconciliation base;
  its `feat/gogv2` head is 17 commits newer than the sync copy and matches its
  remote-tracking branch.
- The user confirmed the legacy BSON migration is no longer needed and the
  former work copy is authoritative. The sync copy was explicitly deleted;
  its removal may propagate through Syncthing.
- Active repositories now live under `/home/kiri/dev/work`; inactive material
  is preserved intact under `.archive`, and the registered dashboard sensor
  worktree is under `.worktrees`.
- Removed the obsolete dashboard converter app and Python dependency. The
  dashboard flake evaluates, while the pre-existing feature work remains
  uncommitted and unstaged.
- Pruned the absent `feat/echarts-domain` worktree metadata; its branch and
  commit remain. The registered sensor worktree is valid at its new path.
- Shared Codex defaults and MCP endpoints are generated at
  `/etc/codex/config.toml`; Context7 reads its header from the environment. The
  unused generic `~/.config/mcp/mcp.json` is absent.
- Removed redundant legacy-path trust entries from the current Codex config;
  the existing trusted home scope covers the moved repositories.
- Bruno's transitional OpenAPI reference now follows the archived management
  repository until the HTTP pilot retires or retains Bruno.
- Post-move repository IDs, remotes, dirty counts, and the registered sensor
  worktree match the pre-move inventory. The Nix flake still evaluates.
- No-link system builds for Zenith, Polaris, and Orion pass with the migrated
  paths and declarative Codex defaults.
- Removed the unused stable Nixpkgs input and made Noctalia follow the root
  package-set pin. Added `just` to the workstation tools and selected the
  store-backed `podman-compose` provider through native `containers.conf`.
- The resulting flake check and no-link builds for all three hosts pass.
- The verified-empty legacy directories `~/work/.data/mongodb`, `~/work/.data`,
  and `~/work/repos` are removed. The stale `~/.config/codex` tree is preserved
  because it contains old authentication and conversation state; moving that
  recoverably requires explicit user approval.
- `rules-service` now has a non-mutating direnv shell, minimal Nix toolchain and
  formatting check, native Cargo recipes, factual repository instructions,
  one `.http` request collection, explicit test/live deployment paths, and a
  corrected test-oriented environment example. The full locked Cargo check
  passes; Clippy reports 18 pre-existing advisory warnings.
- Private Cargo Git dependencies remain in the ordinary developer SSH
  boundary. They are not vendored or bridged into sandboxed Nix builds.
- `yookr-api` now owns its reviewed OpenAPI 3.1 contract, local/test/live HTTP
  metadata, curated login/health/admin-read requests, a non-mutating shell, and
  one explicit snapshot-and-diff recipe. `oasdiff` 1.31.0 is pinned with source
  and vendor hashes because Nixpkgs does not provide it.
- The API flake evaluation and built OpenAPI validation pass. Cargo stops before
  compilation because its locked `constant-store` repository is unavailable
  through both anonymous HTTPS and the working SSH identity. The test service
  is reachable, but its currently deployed GET image returns 404 for the full
  OpenAPI route already present in local source and ingress configuration.
- `traceability-data-model` preserves the existing domain/schema redesign while
  replacing automatic schema recreation with explicit `just reset`. Ordinary
  `just dev` retains its Compose volumes, and the flake check loads fixtures in
  text-ID mode plus a fresh schema in UUID-ID mode. Both the built check and the
  pinned `podman-compose` configuration validation pass.
- Mathesar is removed from `traceability-data-model`; the development stack now
  owns only PostgreSQL plus the explicit schema-reset job. The existing named
  volume and its historical local role are preserved to avoid stranding data.
  `just reset`, `just dev`, the PostGIS fixture query, Compose validation, and
  the flake check pass after the simplification.
- `yookr-dashboard-new` now has the same minimal project contract: a
  non-mutating flake shell, `just dev` over its existing Process Compose graph,
  `just check`, factual `AGENTS.md` and README guidance, a flake formatter and
  formatting check, and a complete public environment example. Its full type,
  lint, and format checks pass; a bounded smoke test reached ready MongoDB and
  Vite processes and stopped cleanly. The commit contains only contract files;
  the user's existing dashboard source work remains unstaged and untouched.
- Fresh local clones of `rules-service`, `yookr-api`,
  `traceability-data-model`, and `yookr-dashboard-new` all pass `nix flake
  check`, expose their tracked `just` catalogs, and remain clean after entering
  their development shells. Credential-bearing and externally blocked checks
  remain separate gates.
- `yookr-data-science` still lacks a committed project contract, its untracked
  flake mutates the checkout during shell entry, and its explicit fetch pipeline
  still targets live services. Those exact files overlap extensive uncommitted
  user work, so a partial agent edit would be unsafe and misleading; wait for a
  commit or deliberate handoff before changing them coherently.
- Gardenctl 2.14.0 and Gardenlogin 0.10.0 are pinned from their official release
  assets; the hashes match the existing binaries byte-for-byte. All three host
  builds pass, the OIDC kubectl plugin is declarative, and Zsh consumes the
  native `gardenctl rc zsh` output with its completion packaged declaratively.
- The local-only `yookr-workbench` has no remote and initially owns only
  `just k8s test`. Each invocation creates an isolated Garden session, rejects
  every environment except test, and opens K9s read-only. Its flake check and
  command catalog pass, and the live rejection path exits before Garden access.
- The user activated and rebooted into the Garden tooling, then confirmed that
  `just k8s test` opens the test shoot in read-only K9s. The initial shell error
  came from Garden writing completion into an absent Flatpak `$fpath` entry; the
  initial target warning came from the fresh session inheriting its parent's
  `KUBECONFIG`. Completion is now store-owned, and the child recipe initializes
  Garden's session-local `KUBECONFIG` before targeting instead of unsetting it.
  The former mutable Garden binaries and generic MCP output are already absent,
  so their stale cleanup gate is closed without further user work.
- NetworkManager's native OpenVPN plugin is declared on Zenith and Polaris;
  no profile is embedded in Nix or imported automatically.
- Zenith's initial NetworkManager activation never established a tunnel:
  OpenVPN 2.6 preserved the profile's legacy `cipher=AES-256-CBC` setting but
  excluded it from cipher negotiation. Adding `data-ciphers=AES-256-CBC` to
  the imported connection fixes the negotiation; NetworkManager then installs
  the pushed tunnel address, default route, private routes, and DNS servers.
  Company database access, disconnect/reconnect, suspend, and reboot pass on
  Zenith, so OpenVPN 3 is retired.
- For the entire workflow implementation, do not access or test Polaris.
  Zenith results are accepted for both workstations; preserve exact deferred
  Polaris replication steps instead.
- OpenVPN 3 is removed from the configuration. A full Zenith system build
  passes, and the first password-authorized switch removed it from the live
  generation. NetworkManager reconnected the VPN successfully afterward.
- Walker's native cached Lua menu scans the four identity roots and refreshes
  when their top level changes. Its default action pipes the selected path to
  one Nix-owned `open-project` command. Elephant supplies that path without a
  terminating newline, so the launcher now accepts a successful partial read
  instead of exiting under `errexit`. It opens Neovim through the configured
  `xdg-terminal-exec` owner and leaves an interactive shell after Neovim exits.
  Elephant reads menu actions only at startup, so rare declarative menu changes
  take effect at the next login. The generated Walker configuration extends,
  rather than copies or replaces, the pinned default provider list and prefixes.
- The shared terminal-launch failure came from the Kitty package rewrite:
  unanchored `sed` changed both `Exec=kitty` and `TryExec=kitty` to include
  `-1`. Arguments are invalid in `TryExec`, causing `xdg-terminal-exec` to
  reject the configured terminal. The rewrite now changes only `Exec`; live
  resolver output selects `kitty.desktop` and retains `kitty -1`, and Yazi
  launches successfully from Walker.
- Restarting the pre-change Elephant and Walker processes loaded the corrected
  project action. The user confirmed that repository launch works and exiting
  Neovim leaves a usable shell in that repository. Passing the project root to
  Neovim now opens the configured Oil default directory explorer immediately.
- A custom restart trigger for Elephant proved unsafe: Walker requires Elephant,
  so activation stopped Walker with its backend but restarted only Elephant.
  The trigger is removed; relogging is the deliberately simple refresh path
  unless real usage proves that inconvenient. The user activated the removal,
  relogged, and confirmed both Walker and the project-root Oil action work.
- Removed the decided Niri arrow/Page Up/Down/numbered duplicates, direct quit
  and monitor-power bindings, extra screenshot modes, first/last and centering
  actions, repeated resize families, and their four shortcut-only scripts. The
  existing `Super+A` helper remains until a two-output session can validate its
  bounded name-based replacement.
- Home Manager now owns the existing Vivaldi launcher filenames for Actual and
  WhatsApp at the higher-precedence user-data location, so both invoke Vivaldi
  8.1.4087.70 with the `Default` profile and stable app ID. Direct activation
  replaced both generated files with managed links and retained the originals
  as adjacent `.bak` files. The entries use native forced ownership because
  Vivaldi may regenerate those exact paths. Outlook is not installed yet, so no
  ID was guessed.
- The first `nixos-maintainer` skill is instruction-only, Nix-owned, and passes
  the skill-creator validator. It encodes repository-specific ownership,
  research, formatting, no-switch, and affected-host validation behavior
  without duplicating generic agent guidance or adding scripts.
- Removed Zed, the inactive Helix modules and note integration, Helix's cache
  settings and complete lock subgraph, and inactive Foot modules. Neovim and
  Kitty remain the sole declared editor and terminal. A repository-wide search
  found no configured npm caller, so global Node.js and Bun are also removed;
  project shells retain language runtimes where needed.
- Postman, HTTPie, Posting, and Resterm are removed after their valuable request
  definitions moved to tracked HTTP owners; Bruno remains through its own
  pilot. Rainfrog, dblab, Harlequin and its profile, Tabularis, DbGate, pgAdmin,
  MongoDB Compass, pgModeler, tbls, global `usql`, and global MariaDB are
  removed; Dadbod and wrapper-local clients remain, while DBeaver stays through
  its gate. Obsidian, Planify, and Calcure are removed after note and task
  authority was established. Pinta/GIMP and the Evolution-era PIM clients remain
  independently gated.
- `rclone` is installed as the only preparation for the selected S3 pilot. No
  remote, wrapper, or interactive browser is added before a real task and
  least-privileged credentials exist.
- GitHub PR review is blocked by the invalid `kirigar` CLI token, not repository
  state. The agent can complete the read-only review and isolated worktree gate
  autonomously after reauthentication.
- The live `nixos-maintainer` skill matches its Nix source and passes static
  validation. Five ephemeral read-only behavioral checks were deliberately not
  run because external policy requires explicit approval before another
  authenticated Codex session receives this private repository.
- In zsh, never use lowercase `path` as a loop or local variable: it is tied to
  `PATH`. A diagnostic command temporarily hid normal executables by doing so;
  the declarative session environment was correct and needed no patch.
- Pinta is added beside GIMP only for its bounded trial; neither outcome is
  preselected before real editing evidence.
- `sync/notes` is now the sole declarative notebook owner, with `daily`,
  `meetings`, and `transcripts` directories, a native `zk daily` alias, and a
  `Super+N` dedicated Kitty/Neovim launcher. The five unique authored work notes
  were copied byte-for-byte, and both distinct todo lists were merged without
  dropping an item. Syncthing returned to idle with no needed files, pull
  errors, or new conflicts. The user confirmed the daily launcher and migrated
  contents, then explicitly approved moving the intact `sync/work/notes`
  rollback copy to desktop Trash. The pre-merge todo copies in `/tmp` may now
  expire normally.
- Evolution 3.60.2 is composed through Nixpkgs with `evolution-ews` 3.60.2,
  whose pinned source contains the separate native Microsoft 365 mail,
  calendar, contacts, recurrence, and OAuth backends. Live inspection caught
  that overriding `evolutionWithPlugins.plugins` replaced its default Evolution
  package, and that a Home Manager package alone did not register the data
  server on D-Bus. The pilot now uses NixOS's native `programs.evolution`
  integration with the EWS plugin beside Thunderbird. The follow-up activation
  succeeded, its source-registry service activates through D-Bus, no account is
  preconfigured, and no authority is changed.
- Named database connections and a Marimo-to-HTML report remain workbench
  evaluation evidence, not scaffolding requirements. Add them only for a real
  cross-system investigation that gives the queries and report durable value.
- Run independent pilots in parallel, but gate every retirement separately.