cc-os/docs/memory-system/03-architecture-decisions.md

# Architecture Decision Records

_Last updated: 2026-07-03_

A running log of decisions and *why*. Format per entry: Context · Decision · Rationale ·
Alternatives rejected · Status. Newest decisions extend the log; supersede rather than delete.

---

## ADR-001 — Two memory types, kept as separate systems

- **Context**: Earlier attempts to make one tool serve both "what happened" and "how do we do
  X" felt forced (e.g. trying to make memsearch filter knowledge by tags).
- **Decision**: Model **episodic** memory and **semantic/knowledge** memory as two separate
  systems with different tools.
- **Rationale**: They have different lifecycles (episodic accretes and decays; knowledge is
  deliberately maintained), different write paths (auto-captured vs curated with guardrails),
  and different query patterns ("when did we…" vs "how do we…"). Separation dissolves the
  earlier integration tension entirely.
- **Alternatives rejected**: One unified store (memsearch-for-everything, or OpenBrain's single
  `thoughts` table) — conflates the two and forces awkward filtering.
- **Status**: Accepted.

## ADR-002 — memsearch for the episodic layer

- **Context**: Need timeline/"what happened" memory (Goal 3) that's NL-queryable and lazy.
- **Decision**: Adopt **memsearch** (Zilliz) off-the-shelf for episodic memory.
- **Rationale**: It already implements the OpenClaw daily-notes + "dreaming" pattern and the
  markdown-as-truth / disposable-shadow-index philosophy we'd otherwise build. Embedded
  **Milvus Lite** (single file), hybrid BM25+vector+RRF search, local ONNX embeddings (no API
  key/cost), a FileWatcher that handles deletions — **no Docker, no server**. Two-line install.
- **Alternatives rejected**: claude-mem (MCP-based — Claude must actively call search; opaque
  blobs vs readable markdown; overkill features). Hand-building daily notes + dreaming
  ourselves (reinventing a solved tool).
- **Status**: Accepted.

## ADR-003 — Flat vault with namespaced tags, not folders

- **Context**: Connelly/Huryn organize by folders (`tools/`, `domain/`). User wants a flat
  Obsidian vault with tags as virtual indexes, and cross-cutting filters (client × tool ×
  convention).
- **Decision**: One **flat markdown vault**; organize via **namespaced, nested tags**
  (`tool/`, `client/`, `domain/`, `convention/`, `scope/`). Slashes are valid Obsidian nested
  tags, so `#tool` matches all children.
- **Rationale**: A note can carry several namespaces at once (`tool/semrush` +
  `client/sesame3g` + `convention/react-ts`) — folders can't express that. Enables "filter by
  client+tool to narrow the index." Enumerable virtual indexes ("what clients/tools exist").
- **Alternatives rejected**: Folder hierarchy (single-axis; can't do cross-cutting filters).
  Pure-prefix path filtering via memsearch `source_prefix` (would force directories back in).
- **Trade-off accepted**: Tags give the *human/Obsidian* free filtering, but the *AI* gets
  nothing for free from tags — we must materialize them into a queryable index (see ADR-004).
- **Status**: **Refined by ADR-011** (type/ and project/ namespaces added; hierarchy-vs-facets
  clarified). Core decision — flat vault, namespaced tags — stands.

## ADR-004 — SQLite + Sequel (Ruby) tag index as the knowledge-layer cache

- **Context**: The AI can't use Obsidian tags directly; tag filtering needs a machine-queryable
  index. A previous `~/Documents/SecondBrain/` tag database was lost track of.
- **Decision**: A small **Ruby program using the Sequel ORM over SQLite**, exposed as a **CLI**.
  Schema: `files(path, mtime, summary, scope)`, `tags(name)`, `files_tags` join
  (`many_to_many`). The summary is a **column on `files`** (an attribute), not a join.
- **Rationale**: Normalized `tags` table makes enumerating the vocabulary a first-class cheap
  query (the "virtual index" goal). The `summary` column is what turns the index from a
  *finder* into a *router* — the AI sees enough to pick a file without opening it (progressive
  disclosure, low tokens). Ruby + Sequel + CLI keeps the contract clean and the DB swappable;
  the AI never touches SQLite directly.
- **Failure-mode guard (the lost-SecondBrain lesson)**: **markdown is always authoritative; the
  SQLite file is a disposable cache** that is never synced and can be rebuilt from frontmatter
  anytime (`index update --rebuild`).
- **Alternatives rejected**: Plain-markdown generated `INDEX.md` (must regenerate; grep-at-scale
  is token-heavy). Frontmatter grep on demand (scales badly). Milvus/Postgres for knowledge
  (overkill; QMD/memsearch prove SQLite is enough — see ADR-006/008).
- **Query output**: returns **path + summary + matched tags** (option C) — tags are cheap and
  show *why* a result matched, useful for cross-client queries.
- **Status**: **Superseded by ADR-010** (Graphify replaces the Ruby/SQLite tag index). The
  `summary` + namespaced-tag frontmatter this ADR introduced is **retained as note metadata**;
  only the bespoke Ruby/SQLite index and its CLI are dropped.

## ADR-005 — Structured-first; semantic search over the vault deferred

- **Context**: Tag filtering ("client/sesame3g + tool/semrush") may miss notes whose wording
  doesn't match the query ("how do we use semrush" vs a note titled "search analytics
  integration").
- **Decision**: Ship the knowledge layer **structured-only** (tags + summaries). **Defer**
  meaning-based search over the vault until it demonstrably bites.
- **Rationale**: Structured tagging is the lightweight/fast thing the user wants, and the
  summary+tag design is built to make it work. Follow the video's "only level up when it bites."
- **Status**: **Superseded by ADR-010.** The premise (ship structured-only, bolt on semantics
  later) no longer holds: Graphify makes the knowledge layer a graph from day one, giving
  structured *and* connection-based recall together. The "only level up when it bites" instinct
  carries forward to whether a *vector* layer is ever needed on top of the graph.

## ADR-006 — QMD as the (deferred) semantic-over-knowledge layer

- **Context**: When ADR-005's structured-only proves insufficient, we want a set-and-forget
  semantic layer over the vault, local and Docker-free.
- **Decision**: Earmark **QMD** (github.com/tobi/qmd) for that role; do **not** install yet.
- **Rationale**: Local markdown search using **SQLite + FTS5/BM25 + local vector embeddings
  (EmbeddingGemma-300M GGUF) + LLM rerank**; CLI + optional **MCP server**; no Docker, no API
  keys. Validates that SQLite + a local vector model suffices (no Milvus/Postgres for
  knowledge). Complements the tag index (QMD filters by path/collection context, not
  first-class frontmatter tags), so it adds semantic recall without replacing structured
  filtering.
- **Alternatives rejected**: Pointing memsearch at the vault (mixes episodic and knowledge
  corpora; its filtering is path-prefix not tags). A bespoke embedding index (reinvents QMD).
- **Status**: **Superseded by ADR-010.** Graphify's knowledge graph fills the
  semantic-recall role (traversal/`explain` over connections) without a separate vector system,
  so QMD is no longer earmarked. Revisit a vector layer only if graph traversal demonstrably
  misses cases where embedding similarity would win.

## ADR-007 — Lazy freshness: write-hook + session-start reconcile, no daemon/cron

- **Context**: The cache must reflect new/edited/deleted/renamed notes without becoming a
  resource hog or going stale on renames.
- **Decision**: **Option A (lazy).** A `PostToolUse` hook updates the index on **AI** writes
  (single-file, prunes on delete). **Manual** edits are caught by a **session-start reconcile**
  (`index update --since` + prune of vanished paths). **No daemon, no cron.**
- **Rationale**: The AI is the primary writer, so write-time hooks give event-driven freshness
  with no polling. The user rarely edits the vault by hand, so a session-start reconcile is
  enough; a continuous `inotify` daemon (the `listen` gem) would add an always-on process to
  manage/sync for negligible benefit. Matches the user's "lazy sync is fine" stance.
- **Alternatives rejected**: `inotify`/`listen` daemon (live freshness, but always-on process
  to manage — unnecessary). Cron reconcile ("seems silly" per user; session-start covers it).
- **Status**: Accepted.

## ADR-008 — Markdown-as-truth; sync the vault, not the indexes

- **Context**: Must be accessible on a VPS / multiple machines but run local-fast (Goal 4).
- **Decision**: Sync the **markdown vault** to the VPS via **git or Syncthing** (choice deferred
  to build time). **Graphs/indexes (Milvus Lite, Graphify `graphify-out/`) are rebuilt per
  machine and never synced.**
- **Rationale**: Markdown is plain text — git/Syncthing sync it trivially; lazy (hourly or
  continuous-async) is enough. Indexes are disposable caches; syncing binary DBs invites
  conflicts for no gain. Local reads stay fast; ownership and portability stay with the user.
- **Alternatives rejected**: **OpenBrain / Mem0** hosted DBs — always-remote, adds per-query
  latency and monthly cost, conflicts with local-fast; ownership weaker (Mem0 especially).
  Only worth it for real-time cross-tool memory, which the user called overkill.
- **Status**: Accepted.

## ADR-009 — Package as a global Claude Code plugin with skills

- **Context**: Every project, on every machine, should know how to use the vault — write
  conventions, query patterns, the hooks, and the CLI — without per-project setup.
- **Decision**: Ship hooks + scripts + CRUD know-how as a **global Claude Code plugin with
  skills**, installed at the user level.
- **Rationale**: Skills carry the "when to write / what conventions / how & when to query"
  guidance to the model; the plugin registers the session-start / session-end / PostToolUse
  hooks and wires up Graphify (extraction/update/query + MCP server). Global install =
  consistent behavior everywhere; single source of truth for the conventions themselves.
- **Status**: Accepted (to be built — see 04-build-plan.md).

## ADR-010 — Graphify knowledge graph as the knowledge layer (supersedes ADR-004/005/006)

- **Context**: ADR-004 specced a hand-built Ruby/Sequel/SQLite tag index (+ CLI) as the
  machine-queryable layer over the vault, with ADR-005/006 deferring meaning-based recall to a
  future QMD vector layer. Before building any of it, we evaluated **Graphify** (`graphify`,
  PyPI `graphifyy`) — a tool that turns a folder into a queryable knowledge graph (local
  tree-sitter AST for code, local-SLM entity/relationship extraction for docs). See
  `06-graphify-evaluation.md`.
- **Decision**: Use **Graphify as the knowledge-layer engine** over the vault, with a **local
  Ollama** backend for doc extraction and free AST for per-project code graphs. **Drop** the
  Ruby/SQLite tag-index CLI (ADR-004) and the earmarked QMD layer (ADR-006); **retain** the
  `summary` + namespaced-tag frontmatter from ADR-003/004 as note metadata.
- **Rationale**: One off-the-shelf tool delivers both what the tag index was for (structured
  retrieval) and what QMD was deferred for (connection/meaning-based recall via graph
  traversal + `explain`) — without writing or maintaining a bespoke index, and without a vector
  store. Code graphs come free. Keeps the markdown-as-truth, no-Docker, no-API-key, local-first
  properties (extraction runs against local Ollama). Net scope reduction: the entire Ruby build
  (old critical-path Step 2) and the QMD layer are removed.
- **What's retained / changed**: `summary` stays the human-written router hint Graphify does not
  generate; namespaced tags stay useful for Obsidian filtering and as node attributes. How
  tightly metadata should feed graph queries is a **build-time refinement**, not settled here.
- **Trade-off accepted**: Graphify's `--update` doesn't prune deleted nodes (stale-node drift) —
  mitigated by a periodic `--force` rebuild on the session-start staleness check (ADR-007's lazy
  model still applies). Graphify also moves fast (flags are version-dependent; anchored to
  v0.8.30) and its headline token-savings numbers are corpus-dependent — benchmark our own.
- **Alternatives rejected**: Building the Ruby/SQLite index as originally planned (more code to
  own; no semantic recall); adding QMD as a second system on top (two stores where one graph
  suffices).
- **Status**: Accepted (to be built — see 04-build-plan.md and 06-graphify-evaluation.md).

## ADR-011 — Faceted tag taxonomy: six independent namespaces (refines ADR-003)

_Date: 2026-06-04_

- **Context**: ADR-003 introduced five namespaces (`tool/`, `client/`, `domain/`,
  `convention/`, `scope/`). During vault-reuse assessment (ADR-012) it became clear that (1)
  the existing SecondBrain vault uses a de-facto first-tag convention for note kind
  (research/plan/log/adr/howto) that should be made explicit and machine-queryable, and (2)
  for a freelancer working many projects per client, project identity deserves a first-class
  namespace rather than being implied by `client/` or `domain/`.
- **Decision**: Knowledge-vault notes are classified by **six independent, flat tag facets**
  that sit side-by-side, never nested into one another:
  - `type/` — note kind: `research`, `howto`, `adr`, `hub`, `plan`, `log`, `clip`, etc.
  - `client/` — which client
  - `project/` — which project (first-class; a freelancer's projects are the primary unit of
    work)
  - `domain/` — knowledge domain / topic area
  - `tool/` — tool-specific knowledge
  - `convention/` — conventions
  - …plus `scope/global` or `scope/project` (retained from ADR-003)

  Hierarchy and relationships are expressed via **hub notes** (`type/hub`), **wikilinks**, and
  **Graphify knowledge-graph edges** — NOT via nested tag paths.

  By convention `type/` is listed **first** in frontmatter, preserving the SecondBrain vault's
  existing type-first ordering habit and making the note kind immediately visible.
- **Rationale**: The vault is flat — hierarchy is not expressed through folder paths or tag
  nesting. The user's reality is many-to-many (many projects per client, knowledge domains
  spanning clients), which a single-parent tree models badly and forces false hierarchy. A
  project hub note links out to both its `client/` and relevant `domain/` tags rather than
  being buried under either. Per-type `_templates` will be provided for **core types only**
  (research, howto, adr, hub); the long tail stays freeform until a pattern earns a template.
  Consistent per-type structure also improves Graphify's local-SLM extraction reliability.
- **Alternatives rejected**: Hierarchical nesting in the style of John Conneely's
  `domain/{product}/{project}.md` folder structure (from the youngleaders.tech article "How I
  finally sorted my Claude Code memory" — **secondary/interview-grade source, not verified
  against primary implementation**). Rejected because: (1) the vault is flat — hierarchy is
  not expressed through folder paths; (2) the user's many-to-many reality maps badly onto a
  single-parent folder tree and forces false hierarchy; (3) nesting one facet through another
  (e.g. `domain/client/project`) creates Law-of-Demeter-style traversal coupling. Conneely's
  structure was the inspiration but diverges here on hierarchy-vs-facets. Faceted parallel tags
  are the flat-vault analogue of what the Graphify graph already does with edges, so they
  compose naturally with the chosen knowledge layer.
- **Status**: Accepted (supersedes the namespace list in ADR-003; core flat-vault +
  namespaced-tags decision stands).

## ADR-012 — Reuse the existing SecondBrain vault as the knowledge vault

_Date: 2026-06-04_

- **Context**: The design called for a flat markdown vault as the semantic knowledge layer
  (ADR-003/008/010). The question was whether to stand up a new `~/brain` vault from scratch
  or adopt the existing `~/Documents/SecondBrain` vault.
- **Decision**: **Adopt `~/Documents/SecondBrain`** as the knowledge vault rather than
  creating a new vault.
- **Rationale**: Assessment found the SecondBrain vault is already flat (all notes at root,
  only a `_templates/` exception — exactly what the design permits), already articulates the
  correct "durable knowledge, not working memory" role in its `CLAUDE.md` and
  `vault-conventions.md`, and contains ~20 real notes. It also includes two patterns that
  **improve on the current design** and should be adopted:
  1. `vault-conventions.md`'s "act without being asked" section specifying *when* the AI
     should proactively query the vault — a behavioral spec the cc-os docs lacked.
  2. Project-config hub notes with a tag-inference table (auto-tag by path pattern) that
     operationalizes *how* to tag a note from a given project.
- **Adaptations required (migration cost)**:
  - Add `summary:` frontmatter to existing notes.
  - Migrate flat unnamespaced tags to the six-facet namespaced form (per ADR-011).
  - Add `scope/global` or `scope/project` to each note.
  - Initialize git in the vault (no `.git` exists yet — required by ADR-008's sync strategy).
  - Replace the vault's `~/.claude/scripts/vault_search.rb` reference (script does not exist)
    with `graphify query` (ADR-010).
  These are mechanical schema migrations, not structural rework.
- **Alternatives rejected**: Starting fresh with a new `~/brain` vault. Rejected because the
  hardest design decision — flat structure, durable-knowledge-only role, governance philosophy
  — is already made and practiced in SecondBrain. The improved behavioral patterns
  (proactive-query spec, tag-inference table) and the existing notes are worth preserving; the
  remaining work is mechanical migration.
- **Status**: Accepted.

## ADR-013 — Build-first / migrate-incrementally (build-order inversion)

_Date: 2026-06-04_

- **Context**: The build runbook (`05-implementation-process.md`) originally front-loaded bulk
  vault migration as Step 1 — migrating all ~20 existing SecondBrain notes and all projects to
  the ADR-011 six-facet taxonomy before the system existed to validate them. This committed to
  a schema and workflow (the tag taxonomy from ADR-011, the vault-reuse choice from ADR-012,
  and Graphify extraction behavior) before any end-to-end path had been exercised. The risk:
  locking in an approach that fails at scale, with no feedback loop until the entire vault has
  been touched.
- **Decision**: **Invert the build order.** The full system is built and validated against a
  small **5–10 note fixture set** first. Bulk vault migration is deferred to the final stage.
  The first real-data validation uses **one small project that contains both code AND
  documents**, exercising both the local-SLM doc-extraction path and the tree-sitter code path
  in the same run. After that single project validates end-to-end, remaining projects are
  onboarded **one at a time** with an observe-and-adjust step between each.
- **Rationale**: Validates the ADR-011 taxonomy and ADR-012 vault conventions against the real
  Graphify extraction pipeline before the entire vault is committed. The first mixed code+docs
  project surfaces both extraction paths (SLM for docs, tree-sitter for code) early, when
  corrections are cheap. Per-project rollout keeps the blast radius of any schema or workflow
  correction small; each project is an opportunity to observe and adjust rather than discover
  problems across 20 notes at once. This is consistent with the "markdown-as-truth, indexes are
  disposable" principle (ADR-008): the vault notes are durable, but the extraction schema should
  be validated before it shapes all of them.
- **Alternatives rejected**:
  - **Keep migration-first (status quo)**: Front-loads all ~20 notes and all projects before
    any end-to-end validation exists. Commits to ADR-011's taxonomy and ADR-012's migration
    steps against the full vault without a feedback loop — exactly the gap this decision closes.
  - **Big-bang migrate everything after build**: Build against fixtures, then migrate all notes
    and all projects in one batch at the end. Avoids the pre-build commitment problem but still
    risks a single large irreversible migration with no observe-and-adjust loop between units.
    Per-project rollout with intermediate checkpoints is strictly safer.
- **Cross-references**: ADR-011 (six-facet tag taxonomy — the schema being validated);
  ADR-012 (SecondBrain vault reuse — the migration steps this order defers).
- **Status**: Accepted (updates `05-implementation-process.md` build order).

## ADR-014 — Graph connectivity comes from authored structure; migration scaffolding is a first-class prerequisite

_Date: 2026-06-05_

- **Context**: ADR-011 specified hub notes + wikilinks + Graphify graph edges as the mechanism
  for expressing hierarchy and cross-note relationships, with ADR-013 deferring bulk vault
  migration to the final stage. Before build began, a discriminating empirical test compared a
  cached-replay graph (per-fixture isolated extractions) against a clean single-pass deep
  extraction (`graphify extract ~/Documents/SecondBrain --backend ollama --model
  qwen25-coder-7b-16k --max-concurrency 1 --token-budget 8000 --mode deep --exclude
  .obsidian`) against the real `~/Documents/SecondBrain` vault under Graphify 0.8.31 +
  qwen2.5-coder:7b. See `07-graph-connectivity-findings.md` for the full data and methodology.
  [primary/measured — 2026-06-05 session]
- **Decision**: **The connective spine of the knowledge graph must be author-provided.** Hub
  notes and wikilinks are not optional scaffolding to add "someday" — they are the mechanism
  by which Graphify connects thematically related notes, and they must be authored **as part of
  the migration step**, not deferred to bulk import. Migration scaffolding (hub notes +
  wikilinks for key concepts) is treated as a **first-class build deliverable** in the
  migrate-incrementally phase of ADR-013.
- **Rationale**: The empirical test found that Graphify is a **structure extractor, not a topic
  clusterer**. Even at `--mode deep --token-budget 8000`, no emergent shared-topic hub nodes
  appeared (no "Pest Control" node, no "Niche Prospecting" node). All cross-note edges observed
  came from explicit references, wikilinks, or document-level semantic similarity — not from
  shared thematic identity. A practical test query ("how do we do niche prospecting outreach for
  pest control?") returned 3 starting notes and traversal could not reach the email templates /
  ACV data / business-model notes (separate communities, no connecting edges). This confirms
  that useful retrieval is gated on migration scaffolding, not on Graphify's extraction power.
  The clean single-pass run also showed the cached graph was partially a build artifact (cross-
  note edges rose from 41% to 78% in a single-pass run), but the structural finding — no
  emergent hub nodes — held in both runs.
- **Relationship to ADR-011**: Validates the hub-notes + wikilinks half of ADR-011 empirically.
  The facet-tag half is **not yet validated**: no edge was observed to arise from shared
  frontmatter facet tags alone. Whether `client/X` or `tool/Y` tags create graph connectivity
  is an **open question** — see "Deferred" below. Do not assume facet tags contribute to graph
  traversal retrieval until tested.
- **Relationship to ADR-013**: Refines the migrate-incrementally stage. "Migration" must be
  defined to include hub note authoring and wikilink addition for key concepts, not just
  frontmatter schema migration (adding `summary:` and namespaced facet tags). The build plan
  (`04-build-plan.md`) should be updated to name this deliverable explicitly.
- **Alternatives rejected**: Relying on the SLM to auto-cluster topics and synthesize hub
  entities — **empirically does not happen** at 7B model size with `--mode deep`. The design
  already intended human-authored hub notes for this; the test confirms that intent was correct
  and the fallback assumption ("maybe the LLM will do it") is false.
- **Deferred**:
  1. **Facet-tag-to-graph-edge question**: Do shared frontmatter facet tags (`client/`,
     `tool/`, `domain/`, etc.) cause Graphify to create edges between notes, or does graph
     connectivity come only from explicit wikilinks/references and semantic similarity? This was
     NOT tested. Resolve before designing graph-traversal retrieval skills.
  2. **Larger extraction model**: Whether a substantially larger SLM (14B, 30B) would
     synthesize emergent topic-hub nodes is untested. Secondary — the design does not depend on
     it — but worth one test run before the build ships.
  3. **`reasoning_effort:"none"` patch**: The clean run required a local patch to `graphify/
     llm.py`. Track the upstream Graphify issue tracker for an official fix; treat the
     installed version as pinned until resolved.
- **Status**: Accepted. Refines ADR-013 (migrate-incrementally phase scope) and empirically
  validates the hub-notes/wikilinks mechanism of ADR-011 while flagging its facet-tag half as
  an open question.

## ADR-015 — memsearch episodic memory version-controlled in a dedicated private repo, auto-synced via cc-os SessionEnd hook

_Date: 2026-06-09_

- **Context**: memsearch's memory store at `~/.memsearch` accumulates daily session-summary
  files (`memory/YYYY-MM-DD.md`) that are the only irreplaceable data in the episodic layer —
  the Milvus Lite index (`milvus.db`) and the bge-m3 ONNX embeddings model are derived/
  disposable and can be rebuilt at any time via `memsearch index`. With Step 4 (episodic layer)
  now live, preserving episodic memory across machines and protecting against local disk loss
  required a sync strategy. The question was: what venue, what scope, and who owns the sync?
  memsearch's own stated design philosophy is "markdown files are the canonical data store; the
  vector database is a derived index" and notes that markdown is "git-friendly" and the index
  rebuildable from markdown — making git a natural fit for the markdown layer.
- **Decision**: `~/.memsearch` is a **dedicated git repo** on branch `main` with remote `origin`
  pointing to a **private self-hosted Forgejo repo**
  (`ssh://git@forgejo.swansoncloud.com:2222/jared/memsearch.git`;
  web: `https://forgejo.swansoncloud.com/jared/memsearch`). A whitelist `.gitignore` tracks
  **only** `memory/*.md` (the daily session files) and `.gitignore` itself. Excluded as
  derived/disposable: `milvus.db` (Milvus Lite index — rebuildable any time via
  `memsearch index`), `config.toml`, and the bge-m3 ONNX embeddings model (lives in
  `~/.cache/huggingface`, not in the store). Auto-commit and push are wired into the **cc-os
  memory plugin's own `session-end.sh`** hook, not the marketplace plugin. The appended block
  guards on `~/.memsearch/.git` existing, runs `git add -A` (whitelist makes it safe), commits
  only when something is staged (message: `memsearch: session memory <date>`), and pushes with
  `timeout 30 ... || true`. The entire block is wrapped in a subshell with a trailing `|| true`
  so it can **never fail session shutdown**.
- **Rationale**: 
  - **Dedicated repo (not folded into the vault or a project repo):** ADR-001 established
    episodic and semantic/knowledge as separate systems by design, and `~/.memsearch` is a
    global, cross-project store with no natural home in any single project repo. Committing
    episodic session logs into the Obsidian vault repo would conflate the two systems and
    violate the separation of concerns that ADR-001 is built on.
  - **Whitelist — commit markdown, exclude rebuildable index/model/config:** consistent with
    memsearch's own design philosophy ("markdown is canonical; index is derived") and with the
    cc-os principle from ADR-008 ("sync the vault, not the indexes"). The 544 MB bge-m3 ONNX
    model is not even in the store; the Milvus Lite DB rebuilds in one command. Committing them
    would bloat the repo for zero durability gain.
  - **Sync lives in cc-os hook, not the marketplace plugin:** the marketplace plugin's hook
    scripts (`stop.sh`, `session-start.sh`, `user-prompt-submit.sh`, `session-end.sh`) perform
    no git operations — only a read-only `git rev-parse --show-toplevel` for scoping. Adding
    sync to a marketplace plugin that can be clobbered by an upstream update is fragile; owning
    it in the cc-os `session-end.sh` keeps the sync logic under our control and version-tracked
    in this repo.
  - **Fail-safe contract:** the `|| true` wrap and `timeout 30` on push ensure that a network
    outage, an unreachable Forgejo instance, or any git error cannot prevent a session from
    closing cleanly. The SessionEnd hook's harness timeout is ~10 s, so push is effectively
    capped; commits land locally first, and any unpushed commit is carried forward by the next
    session's push (daily files are append-only, so no conflict risk).
- **Commingling — resolved 2026-06-09**: `~/.memsearch` aggregates session summaries across
  **all clients** into one global store. This is an **explicit design choice**, not a tolerated
  risk: the user deliberately accepts a single commingled global store across all clients.
  Private self-hosted Forgejo (user-owned infrastructure) is the chosen sync venue; client-
  agreement compliance is the user's knowing responsibility. The concern was previously recorded
  as user-owned and unresolved; that is now closed: single global store is the intended design.
  This also aligns with a forward direction of minimizing dedicated per-client working
  directories — since memsearch captures memory globally regardless of cwd, a single
  `clients/` area for non-project client work becomes viable (direction stated; not yet designed).
- **Alternatives rejected**:
  - **Fold into the Obsidian vault repo:** violates ADR-001's episodic/semantic separation.
    Episodic logs have different lifecycle (accrete and decay), different write patterns
    (auto-captured every session), and would clutter a vault intended for curated durable
    knowledge.
  - **Auto-commit in the marketplace plugin:** the marketplace plugin can be overwritten on
    update, losing the sync logic silently. Out-of-band ownership in cc-os is safer.
  - **Commit the Milvus Lite index or the embeddings model:** both are large binaries, derived
    from the markdown source, and rebuildable. Committing them wastes space and provides no
    additional durability. The markdown files are the canonical source; the index follows from
    them.
  - **Syncthing or rsync instead of git:** git provides both version history and conflict-free
    daily-append semantics; Syncthing is continuous-async (suitable for the vault where changes
    are sparse); git's push-on-session-end cadence matches how memsearch produces data (one
    daily file per day, append-only). Git was already chosen for ADR-008's vault sync rationale;
    applying the same mechanism to the episodic store is consistent.
- **Status**: Accepted (commingling resolved 2026-06-09). Initial commit `106cebc` in the
  Forgejo repo; git repo initialized 2026-06-09 via the `git-context:repo-init` skill. The
  memsearch auto-commit+push was relocated to a dedicated SessionEnd hook entry
  (`memsearch_sync.py`) by ADR-016, 2026-06-12; behavior preserved.

## ADR-016 — Memory plugin sourced from cc-os git repo; bash→Python deep-module port; memsearch-sync split; symlink cutover

_Date: 2026-06-12_

- **Context**: The global Claude Code os-vault plugin was developed iteratively in-place at
  `~/.claude/plugins/os-vault/` — never tracked in version control. Four bash hook scripts
  (`session-start.sh`, `session-context.sh`, `post_tool_use_write.sh`, `session_end.sh`) each
  independently parsed YAML config, decoded stdin JSON, and referenced shared filesystem
  contracts (e.g. a temp file keyed by `$SESSION_ID`) via ad-hoc inline code with no
  abstraction boundary. Additionally, the memsearch auto-commit+push block (ADR-015 behavior)
  was embedded in `session_end.sh` alongside the vault journal logic. The plugin was untracked
  and fragile; the bash scripts had no shared modules, no tests, and a jq-vs-python3
  inconsistency across hooks. An OpenSpec change (`memory-plugin-source-and-port`) was
  initiated to move the plugin into git and port the hooks to Python.
- **Decision (1) — Source location**: The canonical source for the global os-vault plugin is now
  `cc-os/plugins/os-vault/` (this repo, git-tracked). `~/.claude/plugins/os-vault/` is a symlink
  pointing to `cc-os/plugins/os-vault/`; it is no longer an independent directory.
- **Decision (2) — Module shape: deep-module Python architecture**: The four bash hooks were
  ported to Python as a **deep-module** design — three shared modules providing clean, stable
  abstractions, plus thin entry-point scripts:
  - `config.py`: `load_config() -> Config` frozen dataclass — handles YAML parse, path
    expansion, defaults, missing-file case. Replaces 3 separate inline YAML parsers.
  - `hook_io.py`: `read_input() -> HookInput` dataclass (session_id, cwd, file_path, reason)
    — replaces 4 separate inline stdin parsers and eliminates the jq-vs-python3 inconsistency.
  - `session_state.py`: `record_touch(session_id, path)` / `read_touches(session_id) ->
    list[str]` — encapsulates the `/tmp/claude-vault-touched-$SESSION_ID` temp-file contract.
    The temp file still exists by necessity (separate process invocations); the win is an
    explicit, named, testable contract.
  - Thin entry-point scripts: `hooks/session_start.py`, `hooks/session_context.py`,
    `hooks/post_tool_use_write.py`, `hooks/session_end.py`, `hooks/memsearch_sync.py`.
  - Every `main()` is wrapped in `try/except`: uncaught exceptions log to stderr and `exit 0`
    (fail-open invariant — a hook crash cannot block a session).
- **Decision (3) — memsearch-sync split**: The memsearch auto-commit+push block from ADR-015
  was extracted out of `session_end.py` (vault journal) into its own dedicated SessionEnd hook
  entry `memsearch_sync.py`, registered as a separate hook in `~/.claude/settings.json` with a
  30-second timeout. This is a **relocation of ADR-015 behavior, not a reversal** — the
  auto-commit+push logic, whitelist `.gitignore`, fail-safe `|| true` contract, and Forgejo
  remote are byte-for-byte preserved. The benefit is separation of concerns: vault journal
  failures and memsearch sync failures are now independently observable and independently
  fail-open.
- **Decision (4) — Deployment-mechanism deviation (symlink)**: The OpenSpec plan described
  repointing the `local-plugins` marketplace source path to `cc-os/plugins/os-vault/`. In
  practice, there is no `marketplace.json` registry manifest to repoint — the marketplace's
  `known_marketplaces.json` manages marketplace sources, not per-plugin paths, and the hook
  scripts are entirely decoupled from the marketplace (they are registered in `settings.json`
  by hardcoded absolute path). The functional cutover was therefore accomplished via two
  actions: (1) `~/.claude/settings.json` hook entries were rewritten to invoke `/usr/bin/python3`
  with absolute paths into `cc-os/plugins/os-vault/hooks/`, which is the real behavioral cutover;
  (2) a symlink `~/.claude/plugins/os-vault → cc-os/plugins/os-vault/` was created so that skill
  resolution (which follows the plugins directory) continues to resolve correctly. Pre-cutover
  backups were taken: `~/.claude/settings.json.bak-precutover-2026-06-12`,
  `~/.claude/known_marketplaces.json.bak-precutover-2026-06-12`, and
  `~/.claude/plugins/os-vault.bak-bash-precutover/`.
- **Rationale**:
  - **Source in git**: an untracked plugin is invisible to review, diff, and rollback; living in
    cc-os gives change history, golden fixtures, and CI-comparable testing.
  - **Deep-module over OO hook-class hierarchy**: four hooks that run as separate processes
    with no shared runtime have nothing to gain from a class hierarchy — inheritance is just
    complexity. Shared modules with named functions and typed dataclasses give all the benefits
    (single parse path, named contracts) with no class machinery. The three modules provide a
    deep public surface (one function call = full parse + validation) behind a thin API, which
    is the Ousterhout deep-module criterion.
  - **memsearch-sync split**: the original monolithic `session_end.sh` mixed two distinct
    concerns (vault journal vs external git push). Split entry points mean each has its own
    timeout budget, its own failure domain, and its own log line.
  - **Symlink + settings.json rewrite**: the symlink is the minimum viable mechanism consistent
    with how Claude Code actually resolves plugins. Attempting to retool `known_marketplaces.json`
    for per-plugin path overrides would have required reverse-engineering an undocumented format;
    direct hook path rewriting in settings.json is explicit, transparent, and reversible.
- **Alternatives rejected**:
  - **OO hook-class hierarchy**: a base `Hook` class with subclasses per hook type. Rejected
    because hooks run in separate processes — there is no shared runtime state to inherit.
    Shared logic belongs in functions, not class trees.
  - **Keep bash, just move into git**: bash lacks typed dataclasses, structured exceptions, and
    unit-testable modules. The jq-vs-python3 inconsistency (pre-existing bug) would be carried
    forward unchanged, and the temp-file contract would remain implicit.
  - **Repoint `known_marketplaces.json`**: no per-plugin path override semantics confirmed in
    the actual marketplace format; would require undocumented hacking with no rollback path.
- **Consequences / ongoing contracts**:
  - `~/.claude/plugins/os-vault` is a symlink; do not replace it with a directory (e.g. on plugin
    reinstall) without checking cc-os source first.
  - `~/.claude/settings.json` hook entries now invoke Python 3 via absolute path into cc-os;
    keep these entries current when hook filenames change.
  - The fail-open invariant (`exit 0` on any unhandled exception) is a hard contract: never
    wrap a hook's `main()` in code that propagates exceptions to the harness.
  - A fresh-session cutover test was run 2026-06-12 and passed: all five hooks fired from the
    cc-os Python sources.
- **Cross-references**: ADR-009 (global plugin design); ADR-015 (memsearch sync — relocated,
  not reversed).
- **Status**: Accepted. Cutover verified 2026-06-12.


## ADR-017 — Project graph onboarding assesses the repo and writes `.graphifyignore` before extracting

_Date: 2026-06-17_

- **Context**: The `/os-vault:onboard-project` skill's onboarding flow ran `graphify extract . --backend ollama --model qwen2.5-coder:7b` directly. graphify does NOT honor `.gitignore`; it uses a separate `.graphifyignore` file (same syntax — see `docs/graphify/02-installation-setup.md`). With no `.graphifyignore`, onboarding `viking-warrior-training-log` (135 tracked files) walked `node_modules/` (5,858 files / 161MB) and routed every non-code file there through the Ollama doc pass. The run was killed after ~1 hour with no `graph.json` produced. Cost driver: only non-code files hit the Ollama LLM; code uses the free tree-sitter AST pass.
- **Decision**: Onboarding is now assessment-first. Before running `graphify extract`, the skill surveys the repo, classifies directories and file types into include/exclude (weighing non-code file count since only non-code files cost LLM time), confirms the proposed ignore list with the user, writes `.graphifyignore` at the repo root, ensures `graphify-out/` is in the project's `.gitignore`, then runs extract. The ignore list is generated per-project — what counts as noise varies by stack — not from a static template.
- **Rationale**: graphify's `.gitignore` blindness is by design; the correct lever is `.graphifyignore`, not a workaround. Assessment before extraction surfaces borderline dirs (migrations, seeds, fixtures, sample data) that need a human call. Per-project generation avoids the false safety of a shared template that silently mismatches a new stack.
- **Alternatives rejected**: **Static `.graphifyignore` template** — rejected; dependency/build/cache dirs vary by project stack; a template would both over-exclude (suppressing wanted source) and under-exclude (missing stack-specific dirs) for any given project. **Auto-write without user confirmation** — rejected; borderline directories (e.g. migrations, seeds, fixtures) require human judgment; mechanical exclusion would silently drop legitimate content. **AST-only mode (skip doc pass entirely)** — deferred; the `--backend ollama` flag stays; the ignore list is the correct cost lever, not backend switching.
- **Status**: Accepted 2026-06-17.

**Addendum — 2026-06-17:**
- **Default-exclude taxonomy is type-based, not name-based.** The skill's assessment step organizes exclude candidates into 11 categories by KIND (fetched deps, build/generated output, caches, VCS internals, editor/AI-tooling dirs, lockfiles, coverage/logs, bulk data/databases, binaries, secrets, graphify-out/) with a meta-principle: index what a human authored; exclude anything fetched, generated, cached, tooling config, bulk data/binary, or secret. Example names within each category are illustrative, not exhaustive — the goal is to recognize the kind even in an unfamiliar stack. See SKILL.md for the full table.
- **Onboarding uses the 16k-context model.** The `graphify extract` command now passes `--model qwen25-coder-7b-16k` (the `ollama_model` from config.yaml), a 16k-context build of qwen2.5-coder:7b. Inference is GPU-bound (~65 tok/s); the main speed lever is chunk-count × generation — fewer chunks per doc via a larger context window. `GRAPHIFY_OLLAMA_NUM_CTX` does not propagate through graphify's OpenAI-compatible endpoint; the larger context must come from the model's Modelfile.

## ADR-018 — Plugin renames must update the marketplace manifest, not just the directory and settings.json

_Date: 2026-07-03_

- **Context**: The `memory` plugin directory was renamed to `os-vault` (git rename,
  `plugins/memory/` → `plugins/os-vault/`), the `~/.claude/plugins/os-vault` symlink was
  updated, and `settings.json`'s `enabledPlugins` was flipped to `os-vault@local-plugins`.
  Despite this, no `os-vault` slash commands (skills) were available in-session. Root cause:
  the `local-plugins` marketplace manifest (`~/.claude/plugins/.claude-plugin/marketplace.json`)
  — which explicitly lists each plugin's `name`/`source` rather than auto-discovering the
  directory — still declared only `memory`/`./memory`, and `installed_plugins.json` still held
  a stale install record for `memory@local-plugins`. Hooks kept firing throughout (they're
  wired by absolute path directly in `settings.json`, bypassing plugin resolution entirely),
  which masked the break — everything looked healthy except the skills/slash-commands.
- **Decision**: A local-plugin rename/move is a three-file operation: (1) the plugin directory
  or symlink, (2) the marketplace manifest entry (`name` + `source`), (3) `settings.json`
  `enabledPlugins`. After all three, run `claude plugin marketplace update <marketplace>` to
  re-validate, `claude plugin install <new>@<marketplace>` to create the new install record,
  and `claude plugin uninstall <old>@<marketplace>` to drop the stale one — then verify with
  `claude plugin list` and `claude plugin details <new>@<marketplace>` (skills/agents/hooks
  inventory). Documented as a standing procedure in `CLAUDE.md` under "Renaming or moving a
  local plugin."
- **Rationale**: The marketplace manifest is the actual source of truth for what plugins exist
  under a directory-source marketplace — the directory listing itself is not consulted for
  plugin identity. Because hooks are independently wired by path, they give a false signal that
  the plugin is fully operational; skill/slash-command registration must be checked separately
  via `claude plugin details`.
- **Alternatives rejected**: **Rely on directory listing / auto-discovery** — not how
  `local-plugins` (a directory-source marketplace) resolves plugins; the manifest is
  authoritative. **Skip the install/uninstall refresh, just fix the manifest** — leaves a stale
  cached install record in `installed_plugins.json` under the old name, which is confusing and
  can mask a future rename of the same kind.
- **Consequences / ongoing contracts**: any future rename of `os-vault` (or a new local plugin)
  must touch the marketplace manifest, not just `settings.json`. `claude plugin details
  <plugin>@<marketplace>` is the verification step to run after any plugin registration change
  — it surfaces the skill/agent/hook inventory that a passing hook test would not.
- **Status**: Accepted 2026-07-03. Fixed and verified same day (`os-vault@local-plugins` shows
  enabled with 5 skills resolved; stale `memory@local-plugins` record removed).

## ADR-019 — Global os-orchestration plugin supersedes per-project orchestration text

_Date: 2026-07-03_

- **Context**: Session orchestration guidance (how Claude Code should delegate vs execute directly) had accumulated in two incompatible places. (1) A permissive global plugin (`os-orchestration`, moved into cc-os from a standalone `~/dev/cc-plugins/orchestration/` repo) injects `ORCHESTRATION.md` as SessionStart additionalContext carrying the rule: "do single-file/≤2-tool-call ops directly; delegate only when work is parallelizable across independent files, spans many files, or needs isolated/large context." (2) cc-os's own `CLAUDE.md` carried a stricter per-project override: "Delegate all file I/O and shell commands to subagents via the Agent tool. No exceptions by default." The two rules contradicted each other and lived in different places, creating confusion about which one applies when.
- **Decision**: **Adopt the plugin's permissive rule as the canonical global default.** cc-os relinquishes its stricter local override entirely. Every project now follows the plugin's behavior: single-file/≤2-tool-call ops are executed directly; larger or cross-file work is delegated. The `os-orchestration` plugin is migrated into cc-os (`plugins/os-orchestration/`) following the same pattern as `os-vault` (git-tracked, symlinked into `~/.claude/plugins/`, registered via `local-plugins` marketplace).
- **Rationale**: One authoritative orchestration rule, maintained in one place (the plugin, under version control), is simpler than per-project copy-paste text that drifts. The permissive rule reflects actual practice (small direct edits are efficient; large multi-file refactors warrant delegation). Consolidation into a global plugin means all projects get the same behavior without duplication, and the rule can be evolved once and picked up everywhere.
- **Alternatives rejected**:
  - **Keep cc-os's stricter override as a local addition on top of the global plugin**: would recreate the incompatibility problem; cc-os would be a special case instead of a normal project.
  - **Make the strict version the global default instead**: reverses the decision in favor of the permissive rule. The permissive approach scales better (most tasks fit the single-file/≤2-tool threshold); strict delegation adds ceremony for no gain on small work. The plugin's default was chosen by the user for good reason.
- **Cross-references**: ADR-016 (os-vault plugin sourced from cc-os git repo, establishing the git-tracked plugin pattern), ADR-018 (plugin marketplace mechanics for local plugins).
- **Status**: Accepted 2026-07-03. cc-os's CLAUDE.md updated to remove the stricter orchestration section and note that behavior is now global; `os-orchestration` component added to Implemented Components.

## ADR-020 — os-adr plugin built; implementation-time resolutions of the design's open questions

_Date: 2026-07-03_

- **Context**: The `os-adr` plugin (OpenSpec change `add-os-adr-plugin`, built from `docs/adr-system/04-plugin-requirements.md` + `05-plugin-prd.md`) was implemented. Its design deliberately left two questions open — the migration pilot's low-confidence flag-rate threshold, and whether ADR frontmatter needs a Graphify tag/edge convention for retrieval layer 3 — and the pilot forced several heuristic refinements within the design's declared mechanical/LLM boundary.
- **Decision**: (1) **Pilot gate threshold: 25%** low-confidence flags, recorded in `Adr::MigrationReport::FLAG_RATE_THRESHOLD` and printed in every migration report. Final pilot rates (sandboxed copies of the three surveyed projects): viking-warrior-training-log 0%, delta-refinery 0%, llf-schema 8.3% — gate passed. (2) **No Graphify tag/edge convention for ADRs**: retrieval layer 3 expands the *query* file paths one hop through `graphify-out/graph.json` node `source_file` attributes and matches the expanded set against `affected-paths` frontmatter; ADRs never need to exist as graph nodes. Verified against the real llf-schema graph. (3) **Heuristic-boundary refinements from the pilot**: heading-synonym fallback mappings (Why/Rationale/Problem/Background→Context, Impact/Implications/Trade-offs→Consequences) are mechanical copies but cap the file at `migration_confidence: medium`; status synonyms SETTLED/Confirmed→Accepted, OPEN→Proposed map mechanically; bold-label paragraphs (`**Context.** …`) and bulleted labels (`- **Status:** …`) parse as sections; multi-decision files *without* mechanically splittable unit headings classify `unrecognized` (report-only) rather than converting as one blob.
- **Rationale**: The threshold was set empirically at pilot time per the PRD ("do not ship on an unmeasured guess"); 25% sits well above the observed worst case (8.3%) while still failing a genuinely broken heuristic run (pre-tightening iterations hit 100% on delta-refinery and llf-schema). Query-path graph expansion keeps ADRs plain markdown with zero indexing obligations — retrieval reuses the existing project graph instead of imposing a new convention on it (and sidesteps the open ADR-014 facet-edge question entirely). Capping synonym-mapped files at medium keeps the mechanical-vs-judgment boundary honest: the copy is mechanical, the mapping is a judgment call, so it stays flagged.
- **Alternatives rejected**:
  - **Indexing ADRs as Graphify nodes with a tag/edge convention**: adds an extraction pass and a freshness obligation for no retrieval gain at current corpus sizes; revisit only if one-hop query expansion proves too shallow.
  - **Treating heading synonyms as high-confidence mechanical fills**: hides an interpretive mapping from the flag mechanism.
  - **Auto-splitting multi-Status files that lack unit headings**: the split points would themselves be guesses; `unrecognized` + manual handling is the non-destructive answer.
- **Cross-references**: `docs/adr-system/04-plugin-requirements.md` (locked requirements), `05-plugin-prd.md` (phases), `06-eval-scenarios.md` (held-out eval sketches — the eval itself stays deferred), ADR-013 (build-first/migrate-incrementally precedent), ADR-018 (three-place plugin registration, followed during install), ADR-019 (plugin conventions).
- **Status**: Accepted 2026-07-03. Plugin live at `cc-os/plugins/os-adr/`, installed as `os-adr@local-plugins`. Rollout order locked: pilot projects' real migration (interactive, per project) → cc-os retrofit → wider.

## ADR-021 — Model-tier skill-execution eval for os-adr (Eval A), run in-session via subagents

_Date: 2026-07-03_

- **Context**: The os-adr build validated every model-touching surface (skill-following, migrate fills, find judgment) only at the session model used to build it (Fable). Confidence that the skills execute correctly on the *weakest* tier (Haiku) would imply confidence on all stronger tiers. The locked eval methodology (`04-plugin-requirements.md`) covers only the deferred held-out unprompted-behavior eval (requirements 4–5, "Eval B") and says nothing about model tiers. Separately, headless `claude -p` runs consume more of the user's subscription credit than in-session work.
- **Decision**: (1) A second, non-held-out eval — **Eval A: prompted skill-execution across model tiers** — lives at `plugins/os-adr/eval/`: two fixture projects (generated with the plugin's own CLIs), six scenarios (S1 create, S2 create+supersede, S3 find/conflict, S4 find/distractor, S5 init, S6 migrate+fills), a deterministic Ruby checker (`bin/check`, structural invariants only — never prose quality), sandbox + headless-runner scripts, and a runner-prompt template. (2) **Primary run mode is in-session**: a driver session spawns Agent-tool subagents with `model:` pinned to the tier under test; each subagent reads the SKILL.md file directly (uniform across tiers; dispatch is deterministic plumbing). Headless `claude -p` is the fidelity fallback only. (3) The **`autoresearch` skill (Classic mode)** wraps the grid as its metric to iterate SKILL.md *wording only* — checker, fixtures, scenarios, and runner prompt are frozen during a loop, as the guard against metric-gaming. (4) **Eval B gains a model axis** when it eventually runs: scenario × model tier, pass rate per tier as the autoresearch metric.
- **Rationale**: The deterministic core (CLIs, hook, index) is model-independent and already tested; what varies by tier is instruction-following, so the cheap, high-signal eval is exactly that surface with machine-checkable pass criteria (the plugin's invariants double as scoring rules). In-session subagents preserve the user's credits and parallelize; the fidelity gap (no SessionStart context, no slash dispatch) is irrelevant for *explicitly invoked* skills. Eval A prompts are deliberately not held-out — only Eval B's are — so they can be iterated on freely without contaminating the deferred methodology.
- **Alternatives rejected**:
  - **Headless-only harness**: higher subscription cost per run, no parallel driver, no benefit for prompted-execution scoring; kept as fallback for full-fidelity checks.
  - **Folding tier-testing into Eval B**: conflates two questions (can the model *execute* a skill it was told to run vs. does it *recognize* when to run one) and would burn held-out scenarios on mechanical failures.
  - **LLM-judged scoring**: reintroduces the model-capability variable into the scorer; structural invariants are sufficient and reproducible.
- **Cross-references**: `plugins/os-adr/eval/README.md` (procedure + autoresearch invocation), `docs/adr-system/06-eval-scenarios.md` (Eval B sketches, now noting the model axis), ADR-020 (pilot gate the S6 fixture mirrors).
- **Status**: Accepted 2026-07-03. Harness built and self-tested (all six scenarios: perfect-run PASS, untouched-sandbox FAIL). The grid itself has not yet been run against haiku/sonnet.

## ADR-022 — os-status plugin: aggregated in-process SessionStart status checks

_Date: 2026-07-06_

- **Context**: Two real incidents showed that cc-os plugins have per-project and per-environment preconditions nothing verifies: (1) `CLAUDE_CODE_SUBAGENT_MODEL=haiku` in `~/.claude/settings.json` silently downgraded all 23 verified subagent spawns found by the WS1 orchestration audit — and the E1 canary proved the downgrade is unobservable to the model at spawn time, so protection must be deterministic; (2) projects routinely lack artifacts plugins assume (ADR system, vault hub note), with each plugin inventing its own SessionStart reminder in isolation (os-adr, os-doc-hygiene) or having none. The WS3 plan (`docs/plans/ws3-status-convention-plugin.md`) proposed a per-plugin status-check convention plus a glue plugin; two perspective reviews (simplifier, implementer, 2026-07-06) pressure-tested where check code should live.
- **Decision**: New global plugin **`os-status`** (`plugins/os-status/`), one SessionStart hook running all checks **in-process**: plain Python functions in `hooks/checks.py` with the uniform signature `check(ctx) -> CheckResult(status, message)` and a single registry list. Three result states: `ok` (silent), `note` (near-zero-token line, never snoozed/suppressed — exists so os-adr's Eval-B-tuned usage note survives conversion verbatim), `warn` (aggregated into at most one banner; once-per-day snooze + permanent suppress in a gitignored per-project `.cc-os/` dir — state only, never code). Initial checks: `subagent-model-env-override` (env + settings.json `env` block; runs outside git projects too), `adr-system-present` (verbatim port of os-adr's hook, honoring legacy `.os-adr/suppress`), `vault-hub-note-present` (config-first slug in `.cc-os/config`, else `type/hub` + `project/<name>` facet-tag scan of the vault). os-adr's own SessionStart hook was removed atomically in the same change, with the cutover verified against the refreshed plugin cache.
- **Alternatives rejected**:
  - **Option A — per-project code copies** (`.cc-os/status.d/*.py`): copies drift; the 2026-07-04 stale-cache incident is exactly this failure class; adds a per-project install step.
  - **Option B as sketched — per-plugin scripts + subprocess/JSON enumeration**: overbuilt for three trivially cheap probes with one author in one repo (spawn/parse/timeout/error-normalization protocol for three registrants); and enumeration from plugin caches would silently run stale code — os-vault is wired by absolute path to git source and its cache copy is already stale (implementer finding). The uniform function signature is kept as the seam so extraction to per-plugin checks later is mechanical.
  - **Blanket silent-on-ok**: would regress os-adr's present-state usage note; hence the third `note` state.
- **Deferred**: `context` injection field (no consumer yet), consolidation of `.os-adr/`/`.dochygiene/` state dirs, conversion of os-doc-hygiene's reminder, Phase 2 cache-freshness "collection manager" duties, subprocess enumeration.
- **Cross-references**: OpenSpec change `add-os-status-plugin` (proposal/design/specs/tasks), `plugins/os-status/invariants.md`, WS1 findings (`docs/orchestration-audit/2026-07-06-findings.md`), ADR-018 (marketplace-manifest procedure followed for install).
- **Status**: Accepted 2026-07-06. Built, tested (36 model-free tests), installed, smoke-tested (including the regression canary: exported `CLAUDE_CODE_SUBAGENT_MODEL=haiku` produces the override warn).

## Rejected tools (summary)

| Tool | Why rejected for our use |
|------|--------------------------|
| MemPalace (L4) | Storage not readable markdown; isolated drawers (knowledge not interconnected); fights self-managing + cross-linking goals |
| Recall / LightRAG (L5) | Content knowledge bases / deep research, not operational memory; Recall = hosted, you don't own data; LightRAG = enterprise overkill |
| OpenBrain / Mem0 (L6) | Always-remote DB → latency + cost; conflicts with local-fast lazy-sync; only pays off for real-time cross-tool memory (user: overkill) |
| Postgres / Milvus server | Unnecessary — Graphify's local graph (knowledge) + Milvus Lite (memsearch episodic) cover everything locally with no Docker |
| claude-mem | MCP-based (Claude must call search); opaque blobs vs readable markdown; feature overkill |
| Ruby/SQLite tag index CLI; QMD vector layer | Superseded by Graphify before build — one knowledge graph replaces both the structured index and the deferred semantic layer (ADR-010) |