@zerwiz/ymir 0.1.27 → 0.1.28

package/docs/fixes/runtime/unversioned-2026-09-17-the-fork-stands-the-machine.md

@@ -0,0 +1,12 @@
+## runtime · unversioned · 2026-09-17 — the fork stands the machine
+
+### Why
+- **`config/app-repos.yaml` (public)** — the five app repos, the fork's road:
+  a fresh clone with no hoard registry falls back to this map in the installer's
+  `apps` step, so `git clone` + `bin/ymir-install.sh` pull the four surfaces as
+  they would on the author's seat. Verified on a clean clone (`ymir-plan.sh`:
+  hlidskjalf/odrerir/sessrumnir/smidja all `DO`). The hoard registry stays
+  authoritative when present.
+
+### Files
+- *(carried from the frozen CHANGELOG.md)*

package/docs/fixes/runtime/unversioned-2026-09-17-the-fragment-convention-is-enforced-where-github-runs-it.md

@@ -0,0 +1,26 @@
+## runtime · unversioned · 2026-09-17 — the fragment convention is enforced where GitHub runs it
+
+### Why
+- **The gap, found by doing it.** The pre-push hook folds fragments before a
+  push — but a **GitHub merge bypasses the local hook entirely**. PR #43 merged
+  and left its own fragment unfolded on `main`, so the ledger silently fell
+  behind what the fragments already told. `bin/changelog-assemble.sh --check`
+  catches exactly that state; nothing was calling it.
+- **The gate, at the layer that runs.** `.github/workflows/ymir-changelog-fragments.yml`:
+  - **on `push` to main** — `bin/changelog-assemble.sh --check` fails when any
+    fragment is unfolded, so the ledger can never fall behind a merge.
+  - **on a PR** — a change touching code or docs must add a fragment, edit
+    `CHANGELOG.md`, or carry the `no-changelog` label. The label records the
+    deliberate decision that no entry belongs, rather than letting it be
+    forgotten.
+- **The pattern is the standard one.** This is the well-trodden "news fragments"
+  design (towncrier — Twisted, pytest, pip, attrs; changesets; CPython). The
+  zeroc-ice proposal states the root cause exactly: *"PRs edit one shared file.
+  Better merge discipline doesn't fix it; not editing the shared file does."*
+- **`merge=union` was considered and rejected on evidence.** The Rigor ADR-105
+  tried it and measured it: *"GitHub's PR-mergeability and merge computation
+  ignore `.gitattributes` merge drivers, union included."* It works against local
+ 
+
+### Files
+- *(carried from the frozen CHANGELOG.md)*

package/docs/fixes/runtime/unversioned-2026-09-17-the-layout-swept-every-reader-taught-the-new-doors.md

@@ -0,0 +1,24 @@
+## runtime · unversioned · 2026-09-17 — the layout swept: every reader taught the new doors
+
+### Why
+- **bin/nornir-job-observer.sh** — masterplan, runes ledger, and the well now
+  resolve to the hoard homes (`hodd/docs/masterplan.md`,
+  `hodd/memory/runes_audit.md`, `hodd/memory/well`), overridable via
+  `BROKK_MASTERPLAN`/`BROKK_WELL_DIR`.
+- **bin/saga-session-start.sh** — the Sága digest's open-orders read follows
+  the home (`hodd/docs/masterplan.md`).
+- **bin/nornir-job-daily-briefing.sh** — message/comment paths updated to the
+  same doors.
+- **bin/docs-guard.sh** — block message names both new homes (hodd/docs,
+  memory/plans).
+- **.agents/agents/{brokk,forseti-reviewer,mimir-planner}.md** — the plan
+  shelf, masterplan, and append-only log pointers moved to their hoard
+  homes (`memory/plans/<domain>/`, `hodd/docs/*`).
+- **.agents/skills/{tyr-check,ymir-host,ratatoskr-a2a}/SKILL.md** — the same
+  pointers, plus the Ratatoskr spec → `memory/plans/mesh/25-ratatoskr-a2a.md`
+  and `hodd/docs/a2a-runs.md`.
+- **docs/workspaces.md** — status note: workspaces settled under
+  `hodd/workspaces/` (2026-09-17).
+
+### Files
+- *(carried from the frozen CHANGELOG.md)*

package/docs/fixes/runtime/unversioned-2026-09-17-the-model-that-raised-itself-and-the-workhorse-that-repl.md

@@ -0,0 +1,26 @@
+## runtime · unversioned · 2026-09-17 — the model that raised itself, and the workhorse that replaced it
+
+### Why
+- **A model raised itself at login and killed the session.** `qwen3-6-35b-a3b`
+  carried `[Install] WantedBy=default.target`, so systemd started it at
+  12:36:33; with `--no-mmap` on a unified-memory APU the weights were
+  unreclaimable, memory and swap drained, and the kernel's global OOM killer took
+  `gnome-shell` — and the Ymir stack with it. `local-models.md` §4 gains
+  **three traps** from this (anonymous load on unified memory; backend chosen by
+  reputation rather than measurement; reasoning configured through the chat
+  template), and a new **§12** states the rule: a model service unit must never
+  raise itself.
+- **`§11 — Reasoning is a resource you must configure`** is new. Unbounded
+  reasoning fails *totally* — no answer at all — not merely worse. The same model
+  on the same request produced clean, degenerate (repetition), or unterminated
+  output depending only on how reasoning was configured. Prefer the engine's
+  native budget flag; the chat-template keyword is a different mechanism.
+- **Seated a new local workhorse** (a 176B-parameter MoE, Vulkan backend, MTP
+  speculative decoding, file-backed load). Measured before seating: prefill and
+  decode both at parity with the reference stack for this silicon. The per-host
+  record belongs in the hoard, not here — §6.
+- **Ran a 32-run A/B on guidance during reasoning** (counsel at the moment of
+  failure vs counsel u
+
+### Files
+- *(carried from the frozen CHANGELOG.md)*

package/docs/fixes/runtime/unversioned-2026-09-17-the-package-shipped-the-memory-well-one-version-now-excl.md

@@ -0,0 +1,26 @@
+## runtime · unversioned · 2026-09-17 — the package shipped the memory well (one version, now excluded)
+
+### Why
+- **The exposure.** `@zerwiz/ymir@0.1.5` carries five files that are nobody's but
+  the operator's: `.agents/memory/kaia.engram`, its `-shm` and `-wal` sidecars, and
+  `.agents/memory/well/{episodes,workspace}.jsonl`. A public npm tarball contained
+  the live memory well — the store and the episodes. Verified by fetching the
+  published artefact and listing it; `0.1.0`–`0.1.4` are clean, and this branch's
+  build is clean.
+- **The cause, and it is a trap worth naming.** `.gitignore` excludes
+  `.agents/memory/kaia.engram*` and `.agents/memory/well/*.jsonl`, and the repo is
+  clean — but **npm does not consult `.gitignore` when `files[]` names a whole
+  directory.** `files: [".agents/"]` packs that subtree *including* the ignored
+  files. Nothing warned: the leak travelled in the artefact, not the repo.
+- **The fix.** The manifest excludes them explicitly, because a `files[]` list
+  cannot rely on the ignore file it overrides:
+  `!.agents/memory/kaia.engram*`, `!.agents/memory/well/*.jsonl`,
+  `!.agents/memory/*.db`, plus `!**/__pycache__/` and `!**/*.pyc` for the compiled
+  junk that was travelling the same way. Verified: `npm pack --dry-run` carries
+  864 files and **no** memory store — the memory README and the ledger's scaffold
+  header are all that remain, as they should be.
+- **The rule for every future package:** a `files[]` entry that names a directory
+  overrides `.giti
+
+### Files
+- *(carried from the frozen CHANGELOG.md)*

package/docs/fixes/runtime/unversioned-2026-09-17-the-phantom-optional-pins-cut.md

@@ -0,0 +1,9 @@
+## runtime · unversioned · 2026-09-17 — the phantom optional pins, cut
+
+### Why
+- The platform's stale \`optionalDependencies\` (sessrumnir pinned at the broken
+  0.1.9) shadowed the required \`^0.1.10\` — npm never fetched the good seat.
+  Removed; the required deps now resolve to latest (0.1.18).
+
+### Files
+- *(carried from the frozen CHANGELOG.md)*

package/docs/fixes/runtime/unversioned-2026-09-17-the-sweep-and-the-door-that-mends-itself.md

@@ -0,0 +1,26 @@
+## runtime · unversioned · 2026-09-17 — the sweep, and the door that mends itself
+
+### Why
+- **A retired name is swept from the launcher.** `ymir-visualizer.desktop` and its
+  icon pointed at a glyph nobody ships — a blank square for anyone who knew the old
+  name. `design-icon.sh install` removes retired entries now (`ymir-visualizer`,
+  and `ymir-hlidskjalf-mobile` where that app is not installed).
+- **`ymir hlidskjalf` mends the Electron runtime itself.** npm gates install
+  scripts by default, so a fresh install can build the web app perfectly and have
+  no window. The door now runs the repair it used to *describe*
+  (`npm install-scripts approve electron` + `npm rebuild electron`), and only then
+  refuses — with both the manual command and the honest alternative
+  (`ymir install --no-desktop`).
+- **Two faults are provably in the app repos, not here** (recorded so nobody hunts
+  them in Ymir again):
+  - **the fourth hall** — the landing's list is a hardcoded SPA component
+    (`HallsSwitcher`; the bundle says *"three halls of Ymir — one seat, three
+    roofs"*), not data from the gate API. Óðrerir must be added in
+    `zerwiz/hlidskjalf` and the SPA rebuilt.
+  - **the ember glow** — the seat ships `emberBootTheme` in its built renderer but
+    no themed cloth beside it; the house's framework-neutral
+    `midgard/design-system/ember.js` exists precisely so every surface can share
+    one fire, and the seat must be told to light it (`zerwiz/sessrumnir`).
+
+### Files
+- *(carried from the frozen CHANGELOG.md)*

package/docs/fixes/sessrumnir/unversioned-2026-09-12-the-editor-keeps-its-own-seat.md

@@ -0,0 +1,16 @@
+## sessrumnir · unversioned · 2026-09-12 — The editor keeps its own seat
+
+### Why
+- **`open-editor.ts` seats the editor in Þjazi:** inside herdr, `/edit` and
+  `ctrl+shift+e` now open the Allfather's editor in its **own tab** of the home
+  workspace (`herdr tab create --focus` + `herdr pane run`, label
+  `ymir:edit[:<base>]`, recorded in `state/herdr-seats` for `close-all`) instead
+  of suspending the pi session — the agent no longer dies when the editor
+  opens. Outside herdr the old roads remain: GUI editors launch detached,
+  terminal editors suspend the TUI and resume on exit.
+- **`open-editor.ts` shows hidden files:** the picker and `/edit`
+  tab-completion now merge the git listing with a bounded walk, so gitignored
+  dotfiles (`.env.local`, `.env.realm`, …) are openable too.
+
+### Files
+- *(carried from the frozen CHANGELOG.md)*

package/docs/fixes/sessrumnir/unversioned-2026-09-13-sessr-mnir-the-seat-hall.md

@@ -0,0 +1,28 @@
+## sessrumnir · unversioned · 2026-09-13 — Sessrúmnir, the seat-hall
+
+### Why
+- **Sessrúmnir adopted:** the Apache-2.0 **pi-desktop** GUI
+  (`FaqFirebase/pi-desktop`, v0.1.7-alpha) is vendored at `apps/sessrumnir/`
+  and rebranded **Sessrúmnir** — the seat-hall where the Allfather converses
+  with the machine. The external engine keeps its product name; the GUI the
+  user sees is Sessrúmnir, themed with the way-of palette (sky `#38bdf8` on
+  deep navy) as the new default built-in theme. Window titles, launcher,
+  package identity, tray/autostart strings, data-dir name, and the Pi logo
+  (now a Sowilo rune bolt) are rebranded; upstream `LICENSE` is preserved and
+  `NOTICE` records the adoption.
+- **Forge scripts:** `bin/sessrumnir-ensure.sh` (status/ensure/install — deps
+  are never committed, installed on first run with the electron-binary heal
+  `scripts/electron.sh` already uses) and `bin/sessrumnir.sh`
+  (start/status/stop with a workspace argument).
+- **Install wiring:** `bin/ymir-install.sh` gains a `sessrumnir` step (after
+  `hermes`) and lists Sessrúmnir in `workspace/INSTALL.md`; the
+  `installation.md` asset's step table, adopted-engines table, and Hermes-style
+  section are updated in the same change.
+- **Registry & audit:** the Hoard registry (`hodd/identity/projects.yaml`)
+  gains the `sessrumnir` project (upstream remote); the adoption is inscribed
+  in Runes.
+- **Not yet installed:** `apps/sessrumnir/node_modules` is absent — run
+  `bin/ses
+
+### Files
+- *(carried from the frozen CHANGELOG.md)*

package/docs/fixes/sessrumnir/unversioned-2026-09-13-the-carved-cloth-reaches-the-seat-hall-sessr-mnir.md

@@ -0,0 +1,27 @@
+## sessrumnir · unversioned · 2026-09-13 — the carved cloth reaches the seat-hall (Sessrúmnir)
+
+### Why
+- **The landing's cloth is now the seat's own look.** Every semantic token in
+  `apps/sessrumnir/src/renderer/src/index.css` `@theme` was mapped to its carved
+  twin from the landing page's `:root` (`CodeP/ymir-homepage/src/lore.html`):
+  stone `#0e0c09` (app) with panels `#151209`/`#1a1610`/`#14100b`, bone text
+  `#cfc3a9`/`#9a8f75`/`#6b6250` (faint/ghost as bone washes), bronze
+  `#c9973f` accent and `#7d5f2a` brass rule for `border-strong`, blood
+  `#c2584a`/`#7c3a30` for error, steel `#96a0a8` for success. Semantic names are
+  unchanged — only the values moved.
+- **Fensalir** (the weaving halls) is registered as a built-in theme
+  (`themes/fensalir.json`, id `fensalir`, appended to `BUILTIN_THEME_IDS` — an
+  additive change that keeps every persisted id resolving). The seat's own
+  built-in `sessrumnir` theme was carved to the same cloth, so the default look
+  of an existing profile is the cloth on restart. A new test holds the two
+  together (`the cloth is one`) and another holds the CSS `@theme` base to the
+  theme file (`never drift`).
+- **Cloth is the default, not a cage:** the ThemeEngine's ordering is untouched —
+  every other built-in theme, any user theme, and `high-contrast` still win when
+  chosen.
+- **Type:** Cormorant (display), Newsreader (body), IBM Plex Mono (mono) bundled
+  from `@fontsource*` (offline, `font-src 'self'`), with Inter/JetBrains Mono
+  kept as
+
+### Files
+- *(carried from the frozen CHANGELOG.md)*

package/docs/fixes/sessrumnir/unversioned-2026-09-13-the-new-tree-the-mesh-the-seat-hall.md

@@ -0,0 +1,25 @@
+## sessrumnir · unversioned · 2026-09-13 — the new tree, the mesh, the seat-hall
+
+### Why
+- **Everything in the new tree:** the Sessrúmnir work plus the orphaned fixes
+  were migrated from the `ymir-old` working tree into this canonical tree —
+  `apps/sessrumnir/` (deps vendored locally, never committed), the two
+  `bin/sessrumnir-*.sh` scripts, `step_sessrumnir` in `bin/ymir-install.sh`,
+  the Hoard registry entry, the `installation.md` asset, and the
+  `syn-watch-arm.js` continuity fix (re-arm when `.supervision-armed` exists
+  even with no task metadata). The superseded self-service register-gate work
+  was **not** ported — this tree's committed invite-gate is canonical.
+- **Full install ran green** in the new tree: every step OK, invite minted
+  (`YMIR-DBN4-FXEX`), all five services up, `bin/ymir-validate.sh` 11/11 PASS.
+- **Ratatoskr mesh healed:** the `a2abridge` engine (MIT,
+  `vbcherepanov/a2abridge` v3.0.0) is installed at `~/.a2abridge/bin/`, its
+  directory daemon runs on `127.0.0.1:7777`, and `bin/ratatoskr.sh
+  status|doctor` pass. Forged `bin/a2abridge-ensure.sh` (installs the engine,
+  patches the systemd unit's unwritable `/var/log` log paths to journal,
+  keeps the directory up) and wired it into migration `0002-a2a-mcp.sh`, so
+  every install/update heals the mesh instead of leaving a dangling MCP path.
+- **MCP bridges fixed:** `engram-mcp` failed at startup (its Python env
+  lacked the `mcp` SDK; mcp 2.x renamed FastMCP, so it is pinned to `mcp<2`
+
+### Files
+- *(carried from the frozen CHANGELOG.md)*

package/docs/fixes/sessrumnir/unversioned-2026-09-13-the-realm-seat-carved-wayof.md

@@ -0,0 +1,24 @@
+## sessrumnir · unversioned · 2026-09-13 — The realm seat, carved (wayof)
+
+### Why
+- **Problem:** the realm tree was bare (only the shipped example), no tenant
+  was seated, and the session digest resolved the Allfather's realm to
+  `default` — his company and private holdings unreachable by the map.
+- **Seat:** `svartalfaheim/wayof/` carved from the example — `.env.realm`,
+  `SECRETS.md`, `AGENTS.md` (persona), `domains/`, `projects/`, the five
+  workspace branches, and the 9 company cards seeded from the identity hoard
+  (askr · brokkforge · dvalin · mannheim · muninn · runestone · utgard · wayof ·
+  ymirlabs), realm field unified to `wayof`, repo paths corrected to this
+  machine. `data/realm.md` now pins `wayof` (was drifting between `way-of`,
+  `wayof`, and `default`; hoard cards + persona updated to match).
+- **Hood:** `svartalfaheim/wayof/HOOD.md` is the map of the hoard and the
+  seat; `bin/saga-session-start.sh` stage 6 (context digest) now prints
+  `--- hood ---` whole from the realm seat, so a session opens knowing the
+  operator's holdings. Living overviews: `workspace/company/wayof-overview.md`
+  and `workspace/company/aigf.md` (hodd stays the private store; the realm
+  points at it).
+- **Verified:** `bash -n` green; realm-lib resolves `wayof`; `realm env:
+  present`.
+
+### Files
+- *(carried from the frozen CHANGELOG.md)*

package/docs/fixes/sessrumnir/unversioned-2026-09-16-apodex-seat-verified-the-weave-mended.md

@@ -0,0 +1,27 @@
+## sessrumnir · unversioned · 2026-09-16 — Apodex seat verified; the Weave mended
+
+### Why
+- **Correction (cites the entry below, never rewrites it).** The earlier
+  entry said *"LM Studio is retired; its :1234 port belongs to Apodex now."*
+  That was false as carved. LM Studio still holds `:1234` and is what serves
+  the Apodex GGUF; Apodex took the **model seat**, not the server.
+- **Live verification.** `bin/apodex-smoke-test.sh` → PASS (valid tool-capable
+  response, exit 0). `bin/huginn-research-worker.sh` with **bare defaults** →
+  `status: completed`, verdict written, observed into Mimirsbrunn. The Apodex
+  seat (`http://127.0.0.1:1234/v1`) is live.
+- **One model id, everywhere.** The seat serves **`apodex-1.0-mini`**; the
+  worker, smoke test, `.env.example`, `opencode.json`, `config/agents.yaml.example`
+  and Huginn's profile now all say so. Previously three names drifted
+  (`apodex/Apodex-1.0-mini-Q4_K_M`, `apodex/apodex-mini-q4`, the served id) and
+  the default worker call failed with `No models loaded`.
+- **Four defects mended.** `bin/huginn-research-worker.sh` made executable
+  (was 644 — bare invocation died with `Permission denied`); the smoke test's
+  `--serve` GGUF path corrected to
+  `~/Models/FlameF0X/Apodex-1.0-mini-Q4_K_M-GGUF/apodex-1.0-mini-q4_k_m.gguf`;
+  `bin/models-detect.sh` env var renamed from the typo `APEDEX_URL` to
+  `APODEX_URL`.
+- **Weight truth.** The Apodex Q4_K_M weights are **21.7 GB**, not the ~4–6 GB
+  the plan assumed — it ca
+
+### Files
+- *(carried from the frozen CHANGELOG.md)*

package/docs/fixes/sessrumnir/unversioned-2026-09-16-the-visualizer-finds-its-db-and-the-seat-hall-actually-b.md

@@ -0,0 +1,22 @@
+## sessrumnir · unversioned · 2026-09-16 — the visualizer finds its DB, and the seat-hall actually builds
+
+### Why
+Two more readers left behind — both discovered by *starting the apps*, not by any
+gate. Each failed silently in its own way.
+
+- **`scripts/start.sh` pointed the smithy at a repo path that no longer holds the
+  DB.** `0003-private-data-separation` moved it to `$YMIR_HOME/smidja/smidja.db`,
+  but the starter still passed `CMD_DB=<repo>/apps/smidja/smidja_data/smidja.db`
+  — so the visualizer API died on boot with `smidja.db not found` and `:8437`
+  answered nothing. It now resolves the same pair `bin/smidja-bootstrap.sh` does
+  (home first, in-repo fallback, `SMIDJA_DB` override). The asset already
+  *described* this resolution; the code now implements it.
+- **`bin/sessrumnir-ensure.sh` never built the app.** `[ "$built_present" ]` tests
+  a **literal, non-empty string** — always true — so `ensure --install` skipped
+  `build_app` and reported `built: no` with exit 0. Sessrúmnir only built when the
+  build was run by hand. Now `built_present` (the function) is called.
+- `galdr-reread`: `.agents/skills/galdr-ymirsystem/assets/smidja.md` — the DB
+  path, the `scripts/start.sh` DB resolution, and the observer's read list.
+
+### Files
+- *(carried from the frozen CHANGELOG.md)*

package/docs/fixes/sessrumnir/unversioned-2026-09-17-0-1-13-the-re-voiced-seat-arrives-and-the-packages-are-l.md

@@ -0,0 +1,8 @@
+## sessrumnir · unversioned · 2026-09-17 — 0.1.13: the re-voiced seat arrives, and the packages are listed
+
+### Why
+- **`@zerwiz/sessrumnir` 0.1.9** — the seat that says Sessrúmnir and "Brokk is working", with the fork`s name gone from its source, its bundles, its installer and its package identity. Pinned so a fresh install receives it.
+- **The README lists every package** on npm, and why the smithy is `@zerwiz/smidja-factory`.
+
+### Files
+- *(carried from the frozen CHANGELOG.md)*

package/docs/fixes/sessrumnir/unversioned-2026-09-17-sessr-mnir-stops-wearing-pi-s-clothes.md

@@ -0,0 +1,23 @@
+## sessrumnir · unversioned · 2026-09-17 — Sessrúmnir stops wearing Pi's clothes
+
+### Why
+- **The naming law, broken at the root.** Clicking the Sessrúmnir icon opened a
+  window titled **Pi Desktop**: `apps/sessrumnir` is the Pi Desktop shell adopted
+  whole, and its product name was still Pi's —
+  `export const PI_DESKTOP_PRODUCT_NAME = 'Pi Desktop'` — rendered as the heading on
+  the home screen, the sidebar and the chat panel. That is the title the Allfather
+  saw, twice, on every launch.
+- **Mended at the name:** `SESSRUMNIR_PRODUCT_NAME = 'Sessrúmnir'` (Odin's hall of
+  many seats), the three rendering components walked, and the old constant kept as an
+  alias so no unwalked import breaks. `tsc --noEmit` exit 0.
+- **Still owed:** 1,255 English strings, **86 of them naming Pi** ("Pi is working",
+  "Quit Pi Desktop", "Pi Desktop v{{latestVersion}} is available", "Start Pi/OMP
+  before planning with Council"), and Pi's logo on the home screen. The locale is the
+  one place they live — that pass is the order, not a patch.
+- Also in the launcher this session: an icon click now raises THE SYSTEM (converge
+  the service, reborn a windowless app, focus a live window), and `stop --view X`
+  stops only X — the loop that killed every view is gone. The Hall's own outage had
+  a cause: an Astro dev server on :4323 while the app waited on :4322.
+
+### Files
+- *(carried from the frozen CHANGELOG.md)*

package/docs/fixes/sessrumnir/unversioned-2026-09-17-the-hearth-re-seeds-on-its-container-sessr-mnir-s-chat-f.md

@@ -0,0 +1,23 @@
+## sessrumnir · unversioned · 2026-09-17 — the hearth re-seeds on its container: Sessrúmnir's chat fire returns
+
+### Why
+- **The ember glow was gone from the chat because the canvas sized itself once
+  at mount and only re-seeded on `window.resize`.** Sessrúmnir's split panes
+  (file tree · editor · image) resize the chat column without any window resize,
+  so the hearth kept a stale or zero-sized canvas — invisible. Both the shared
+  hearth (`midgard/design-system/ember.js`) and Sessrúmnir's React port
+  (`ember-background.tsx`) now watch their container with a ResizeObserver and
+  re-seed only when the measured size actually changes (never per-frame).
+  Hlidskjalf inherits the same mend through the shared module.
+- **Verified live**: Sessrúmnir restarted on the fresh build; a screen capture
+  of the running window shows the fire drawing (ember + warm-glow pixels across
+  the hearth area). The running app had been serving a stale bundle (missing
+  `react-i18next` — deps installed, rebuilt).
+
+galdr-reread: `.agents/skills/galdr-ymirsystem/assets/hlidskjalf-ui.md` — the
+EmberBackground hearth section (shared module + ResizeObserver mend).
+
+# CHANGELOG
+
+### Files
+- *(carried from the frozen CHANGELOG.md)*

package/docs/fixes/sessrumnir/unversioned-2026-09-17-the-packaged-seat-carries-the-watch.md

@@ -0,0 +1,11 @@
+## sessrumnir · unversioned · 2026-09-17 — the packaged seat carries the watch
+
+### Why
+- **The @zerwiz/ymir tarball now ships every surface the runtime reads**: the
+  harness surfaces (`.pi/`, `.opencode/`, `.claude/`, `.codex/`, `.cursor/`),
+  `assets/`, `data/`, `deploy/` — so an npm-installed seat has the watchers,
+  the session-start hooks and the extensions, not only the repo clone (0.1.16
+  → 0.1.17).
+
+### Files
+- *(carried from the frozen CHANGELOG.md)*

package/docs/fixes/sessrumnir/unversioned-2026-09-17-the-readme-lists-the-packages-and-the-pin-moves-to-the-r.md

@@ -0,0 +1,16 @@
+## sessrumnir · unversioned · 2026-09-17 — the README lists the packages, and the pin moves to the re-voiced seat
+
+### Why
+- **The README says what is on npm.** Five packages serve — `@zerwiz/ymir` (the
+  distro) and the four surfaces it depends on — and the section says so with a
+  table, plus the one line that matters to a reader: *why the smithy is not
+  `@zerwiz/smiddja`*. That name was accepted twice and served never (package
+  document 404, website 403, republish refused as already published); it was
+  renamed to `@zerwiz/smidja-factory`, which resolves.
+- **The distro's pin moves to `@zerwiz/sessrumnir@0.1.9`** — the seat that speaks
+  its own name rather than the fork's. A fresh install was still receiving the
+  old voice through the `0.1.8-alpha` pin.
+- **`@zerwiz/ymir@0.1.13`** is the packet that carries them.
+
+### Files
+- *(carried from the frozen CHANGELOG.md)*

package/docs/fixes/sessrumnir/unversioned-2026-09-17-the-seat-s-name-and-the-visualizer-s-home.md

@@ -0,0 +1,12 @@
+## sessrumnir · unversioned · 2026-09-17 — the seat's name and the visualizer's home
+
+### Why
+- **omarchy, not omarchy-1** — the machine is `omarchy`; the fleet config docs
+  and the tailscale-sync map say so.
+- **The Smíðja visualizer's home is the smidja repo clone** —
+  `apps/smidja-factory/apps/visualizer`; the skill tree carries no copy (the
+  stale build was moved aside). `smidja.md` (the governed asset), the smithy
+  SKILL, and the install asset all point at the clone.
+
+### Files
+- *(carried from the frozen CHANGELOG.md)*

package/docs/fixes/skills/unversioned-2026-09-11-galdr-tyr-assets-symlink-not-copies.md

@@ -0,0 +1,11 @@
+## skills · unversioned · 2026-09-11 — Galdr/Tyr assets: symlink, not copies
+
+### Why
+- **Fix:** `tyr-check/assets` is now a **symlink** to the canonical
+  `galdr/assets`, so the two can never drift. A background process had been
+  rewriting the registry, flipping the sync gate red; a symlink makes drift
+  structurally impossible. Compliance back to 8/8 stable.
+- **Verified:** compliance 8/8, smoke 8/8, lint 4/4.
+
+### Files
+- *(carried from the frozen CHANGELOG.md)*

package/docs/fixes/skills/unversioned-2026-09-11-load-the-owning-asset-before-editing-a-governed-path-5-l.md

@@ -0,0 +1,21 @@
+## skills · unversioned · 2026-09-11 — Load the owning asset before editing a governed path (5 layers)
+
+### Why
+The runtime drifted from its documentation because the asset that governs a
+subsystem was never loaded before the code changed. Five layers now prevent it:
+
+- **1. `AGENTS.md`:** a new `governed[6]` table maps each governed path to the
+  asset to load first, and `manual[]` gained `installation`, `ui`, `runtime-spec`.
+- **2. Session digest:** `saga-session-start.sh` prints an `ASSET ROUTING` block
+  every session, so the mapping is always in context.
+- **3. Router loadable:** `galdr`'s `disable-model-invocation` is removed (it hid
+  the one router that owns the assets); `tyr-check` keeps it (deliberate).
+- **4. Enforced seatbelt:** new `bin/syn-asset-pretool-check.sh` denies an
+  `edit`/`write` of a governed path until its asset was read (`state/asset-reads`);
+  the Pi extension relays it and records asset reads.
+- **5. Compliance gate:** `compliance-check.sh` gains an `assets` check that
+  FAILS when a governed path changed without its asset (it caught this very
+  change), and `runtime-compliance.md` documents G12 + the routing map.
+
+### Files
+- *(carried from the frozen CHANGELOG.md)*

package/docs/fixes/skills/unversioned-2026-09-11-omarchy-first-host-skills-machine-learning-native-fixes.md

@@ -0,0 +1,27 @@
+## skills · unversioned · 2026-09-11 — Omarchy-first: host skills, machine learning, native fixes
+
+### Why
+- **New skill `ymir-omarchy`:** Omarchy-native operation — the two laws (never
+  edit `/usr/share/omarchy/`; Hyprland owns placement), host detection, the
+  monitor/scale coordinate trap, the amdgpu GPU crash and mitigation, desktop-app
+  placement, and a safe-customisation table.
+- **New skill `ymir-thjazi`:** Þjazi (herdr-first) — protocol floors (14+ panes,
+  0.8.0 presentation spaces), verification, installation.
+- **Ymir learns the host:** `bin/omarchy-sense.sh` snapshots and *diffs* the
+  machine (Omarchy version, explicit packages, config files, monitors, scale) into
+  `state/omarchy-setup.json`; `bin/omarchy-hook-install.sh` adds an Omarchy
+  `post-update` hook so it re-learns after every `omarchy update`.
+- **Þjazi in the installer:** `bin/herdr-ensure.sh` + new `step_backend`
+  (herdr, else tmux — never a silent fallback).
+- **New `step_omarchy`** (SKIP on a non-Omarchy host).
+- **Desktop fixes:** window placement is now LOGICAL and work-area clamped, and
+  prefers a non-primary display (the old math used physical pixels on a scale-1.5
+  panel and filled the screen). Apps are identified by their command line, not the
+  `.bin/electron` shim pid — which fixes the **double-launch** bug. `stop` kills
+  every matching pid.
+- **GPU mitigation:** `YMIR_DESKTOP_DISABLE_GPU=1` adds
+  `--disable-gpu --disable-gpu-compositing` for small-VRAM iGPUs (the crash was
+  `amdgpu: Not e
+
+### Files
+- *(carried from the frozen CHANGELOG.md)*

package/docs/fixes/skills/unversioned-2026-09-13-the-imported-names-leave-the-halls-norse-naming-purge.md

@@ -0,0 +1,26 @@
+## skills · unversioned · 2026-09-13 — the imported names leave the halls (Norse naming purge)
+
+### Why
+- **"firstmate" retargeted to Ymir's own names across every Ymir-owned surface.**
+  The upstream distro's nautical vocabulary is gone from the docs, skills, `bin/`,
+  config, and the Hlidskjalf review copy: `firstmate` → **Brokk**, `secondmate` →
+  **Eindri-home**, `crew`/`crewmate` → **Eindri**, `captain` → **Allfather**
+  (`docs/lore.md`, `docs/Architecture.md`, `.agents/agents/brokk.md`,
+  `.agents/skills/herdr-panes/assets/{herdr,tmux}-backend.md`,
+  `.agents/skills/eindri-homes/assets/control-plane.md`,
+  `.agents/skills/saga-bearings/assets/board-template.html`,
+  `.agents/skills/ymir-host/assets/thjazi.md`, `bin/README.md`,
+  `apps/hlidskjalf/**`).
+- **Engine literals kept.** Every real upstream identifier stays verbatim — the
+  Herdr labels (`firstmate`, `2ndmate-<id>`, `firstmate-<id>`), the home marker
+  `.fm-secondmate-home`, the envelope `FIRSTMATE_OP: ` and `[fm-from-firstmate]`,
+  and all `FM_*`/`fm-*` names — because `.agents/backend/` (the vendored engine)
+  and `assets/reference/` (provenance) are untouched.
+- **Licensing.** `NOTICE` now records the upstream **firstmate** distro
+  (`github.com/kunchenguid/firstmate`, MIT, © kunchenguid) and a README "Built on"
+  section points to it; the MIT copyright/permission notice is retained.
+- **Governed asset:** `galdr-ymirsystem/assets/hlidskjalf-ui.md` records the Allfather
+  review copy in the same change. `complianc
+
+### Files
+- *(carried from the frozen CHANGELOG.md)*

package/docs/fixes/skills/unversioned-2026-09-16-gunnl-joins-the-hall-skillopt-integration.md

@@ -0,0 +1,17 @@
+## skills · unversioned · 2026-09-16 — Gunnlöð joins the hall (SkillOpt integration)
+
+### Why
+- **SkillOpt integration.** `pip install skillopt` into `.venv/`; training
+  loop (`skillopt-train`), eval (`skillopt-eval`), and nightly
+  self-evolution (`skillopt-sleep`) available for Smíðja prompts and
+  Ymir skills. Nightly Nornir job at 00:30 (bin/nornir-job-skillopt-sleep.sh),
+  staged artifacts only — Allfather approves before adopt. One-time setup:
+  `bin/skillopt-setup.sh`.
+- **Naming.** SkillOpt adopted the name **Gunnlöð** (keeper of the mead of
+  poetry — distills trajectories into refined skill artifacts). Added to
+  `.agents/assets/agents/naming.md` (platform map, 28 subsystems),
+  `.agents/skills/galdr-ymirsystem/assets/registry.md` (skills + external tools),
+  `README.md` (System Map + Skills + gratitude), `NOTICE` (MIT attribution).
+
+### Files
+- *(carried from the frozen CHANGELOG.md)*

package/docs/fixes/skills/unversioned-2026-09-16-one-skills-tree-every-harness-the-config-stomp-mended.md

@@ -0,0 +1,25 @@
+## skills · unversioned · 2026-09-16 — one skills tree, every harness; the config stomp mended
+
+### Why
+- **Every harness now reaches `.agents/skills`.** Pi was already right — its own
+  code walks up from the cwd to `.agents/skills` (and `~/.agents/skills`), so it
+  discovers Ymir's skills natively with no link and no config; a second root
+  under `.pi/` would only double-load. Claude Code, Codex and Cursor read project
+  skills from their own directory, so `bin/valknut-load.sh` now binds
+  `.claude/skills`, `.codex/skills` and `.cursor/skills` to `../.agents/skills`.
+  opencode reaches the tree through `skills.paths`. The install step `loaders`
+  runs the loader, so a fresh install gets all of it.
+- **The `harnesses` gate (G14) now asserts the skills surfaces too** — opencode's
+  config, Pi's native root, and the three links — so a harness silently loading
+  nothing can no longer pass.
+- **Defect mended: the loader was stomping `opencode.json`.** `bin/valknut-load.sh`
+  renders `opencode.json` from `opencode.json.example`, and that example carried
+  only `llama.cpp` — so a loader run silently **deleted the Apodex provider**,
+  which had only ever lived in the generated file. The example now carries Apodex
+  as well, with the served id `apodex-1.0-mini`, so a re-render cannot drop it.
+- `bin/valknut-load.sh --status` now reports every surface (agents *and* skills)
+  with a true row count instead of a hardcoded header.
+- Assets: `harness-integration/README.md` gains the skill-loc
+
+### Files
+- *(carried from the frozen CHANGELOG.md)*

package/docs/fixes/skills/unversioned-2026-09-16-the-skill-loading-audit-wrong-skills-unloaded-the-index-.md

@@ -0,0 +1,26 @@
+## skills · unversioned · 2026-09-16 — the skill-loading audit: wrong skills unloaded, the index made true
+
+### Why
+- **Three phantom skills stopped loading.** opencode was loading `galdr-compliance`
+  and `galdr-crafter` — both superseded, their work long since moved to
+  `tyr-check` — and `NSR`, the NSR scaffolding spec, which sat at
+  `assets/nsr/SKILL.md` carrying frontmatter. A recursive scanner walks
+  `skills.paths` to *any* depth, so all three were announced to every agent as
+  live skills. The two galdr crafters are removed; the NSR spec is renamed
+  `assets/nsr/scaffold-spec.md` and stays a document (its sibling templates
+  carry no frontmatter and were already inert).
+- **Two gates added, one per failure class** (`compliance-check.sh`):
+  - **`harnesses` (G14)** — every link in `.claude/agents`, `.codex/agents`,
+    `.cursor/agents`, `.pi/agents`, `.opencode/agent` must land, one hop, in the
+    canonical `.agents/agents/` tree (Rule 02), and no nested `SKILL.md` may
+    carry frontmatter. One hop deliberately: Galdr's agent file *is* the skill,
+    by the dual-surface rule.
+  - **`skillindex` (G15)** — every real skill dir must be named in
+    `.agents/skills/README.md`, and every `.agents/skills/<name>` path the assets
+    cite must exist. Lines marked planned/legacy/superseded/removed/abandoned/
+    retired are exempt by intent.
+- **The drift they exposed, mended:** the canonical index listed 23 skills while
+  26 existed (`groa-update`, `lifecycle`, `rules-check-drift` were mis
+
+### Files
+- *(carried from the frozen CHANGELOG.md)*

package/docs/fixes/skills/unversioned-2026-09-17-0-1-9-both-shapes-reach-the-registry.md

@@ -0,0 +1,17 @@
+## skills · unversioned · 2026-09-17 — 0.1.9: both shapes reach the registry
+
+### Why
+- **`ymir raise` works on a packaged install.** 0.1.8 shipped before the resolver,
+  so a user who installed from npm hit `cd …/apps/hlidskjalf: No such file or
+  directory` — every package had arrived, and eighteen scripts were looking in the
+  clone's `apps/`. `bin/app-lib.sh` resolves a surface in either shape now.
+- **A clone and a package are one tree to the scripts.** Proven against both: the
+  five surfaces resolve from `$HOME/Ymir` and from an npm-installed
+  package, and `ymir raise` against a packaged tree fetches the app's
+  dependencies, raises the gate API, starts Nornir cron, and brings up Bifrost and
+  the well.
+- **The patience words** — `style_patience` — open the installer's long chain and
+  the raise path's build, so a long hour says what it is doing.
+
+### Files
+- *(carried from the frozen CHANGELOG.md)*