Context

/archcore:context

Pull-mode project context. Surfaces the rules, decisions, specs, and patterns that apply to a code area before you change it — or recaps current focus when picking up work.

Not related to the AI context window or session state — this is about the .archcore/ knowledge base.

When to use

• "What rules apply to src/payments/?"
• "Before I refactor the auth module, show me what I should know"
• "Show me the decisions for src/api/"
• "What rules apply to my current changes / what I just edited?" → --git-changes
• "Pick up where we left off"
• "What was I working on in payments?"
• "Where is the checkout work right now?"
• "Load project context"

Not context:

• Creating documentation → /archcore:capture, /archcore:decide
• Planning a feature → /archcore:plan
• Detecting stale docs → /archcore:audit --drift
• Health audit, counts, or status breakdown → /archcore:audit (--deep for the full audit)

Routing table

Classify $ARGUMENTS into one mode:

The bare words changes or git (no leading --) stay topic — only the exact flag --git-changes or the natural-language signals above switch to git-changes mode, which takes its scope from the working tree (Step 2) and falls back to an empty state when git is unavailable.

Proactive use. --git-changes is the one mode the agent MAY invoke without the user asking: when there are already uncommitted changes and you keep working over them, run it once per task to load the rules for what you've touched, then reuse that result. Do NOT re-run it per edit, and do NOT run it on a clean tree (it returns nothing). Every other mode is user-driven.

Execution

Step 1: Classify

Determine mode from $ARGUMENTS per routing table.

Step 2: Query

Git-changes mode:

Resolve the scope from the working tree, running the resolver once via Bash:

"${CLAUDE_SKILL_DIR}/../../bin/git-scope" --git-changes

${CLAUDE_SKILL_DIR} is this skill's own directory (skills/context/), set by Claude Code for Bash calls during skill execution; the resolver lives two levels up at the plugin root. On a host that does not set it, resolve bin/git-scope relative to this skill file (two directories up). Parse stdout:

• A lone sentinel (__CLEAN__, __NOT_REPO__, __NO_GIT__, __USAGE__) → render the matching empty state (Step 6) and stop.
• Otherwise each plain line is a directory; __TOTAL__ <M> is the raw directory count for cap reporting.

For each directory, call in parallel mcp__archcore__search_documents(path_ref="<dir>", limit=10, sort="relevance") — a smaller limit than path mode, because results aggregate across directories and each matched document carries its full relation graph; a large limit times N directories floods the context. Merge the result sets, dedupe by document path, and tag each result with the directory that surfaced it (used as via in Step 5). Then proceed through Steps 3–5 unchanged.

Ambiguity: if the mode came from a natural-language signal (not an explicit flag) and the resolver returns __CLEAN__, ask one AskUserQuestion — "Working tree is clean; did you mean a specific path or topic?" — and reclassify on the answer. An explicit --git-changes skips the question and shows the clean empty state.

Path mode:

Normalize the argument: trim whitespace, convert \ to /, strip trailing /.

Call mcp__archcore__search_documents(path_ref="<normalized>", limit=50, sort="relevance").

Topic mode:

Call mcp__archcore__search_documents(content="<argument>", limit=50, sort="relevance").

Topic search is strict substring — singular/plural and near-synonyms do not match. If the first call returns empty, retry once with a shorter or alternate phrasing of the same term before falling through to the empty state.

Pickup mode:

Call in parallel:

• mcp__archcore__search_documents(types=["plan", "idea"], status="draft", limit=10, sort="mtime")
• mcp__archcore__search_documents(types=["adr", "rule"], status="accepted", mtime_after="30d", limit=10, sort="mtime")

If the recent-accepted call returns empty, retry once with mtime_after="90d".

Step 3: Group

Path and topic modes — group results by type:

Drop remaining types — accepted plan/idea, task-type, and vision/requirements (prd, mrd, brd, urd, brs, strs, syrs, srs). Results are already sorted by search_documents (specificity → type priority → mtime); keep the top 5 per section. Inside Reference the same sort applies, so rfc (typeRank 3) outranks guide (6) and doc (17) when specificity ties.

Pickup mode — three fixed sections:

• In Progress — results from the drafts call
• Recent Decisions — adr results from the recent-accepted call
• Recent Rules — rule results from the recent-accepted call

Step 4: Guide routing

For each item in the Rules, Decisions, or Specs sections, inspect its incoming_relations. If a relation of type implements or related points from a guide, inline that guide as an indented bullet under the parent. Skip a guide that is already inlined under a sibling in the same section. Track the set of guide paths inlined this way — any guide present in the search results but not in this set is an "orphan guide" and is routed to the Reference section in Step 3 instead of being dropped.

Step 5: Render

## Rules (N)
- **<title>** [rule · <status> · <match.kind>]
  `<path>`
  > <excerpt>
  - 📖 **<guide title>** [guide · <status>]

## Decisions (N)
- **<title>** [adr · <status> · <match.kind>]
  `<path>`
  > <excerpt>

## Specs (N)
...

## Patterns (N)
...

## Reference (N)
- **<title>** [<type> · <status> · <match.kind>]
  `<path>`
  > <excerpt>

## In Progress (N)
⚠️  **<title>** [plan · draft]
  `<path>`
  > <excerpt>

_Classified as: <mode>._

Do NOT emit a section header if its group is empty. The classification footer is always emitted.

Global sources (only when results carry them). If any result has global: true / read_only: true / source_kind: "global", load skills/_shared/globals.md and apply its rendering and precedence rules (badge the document, keep it in its normal type section, mark local-overrides-global for same-slug pairs). If no result is global (the default — and the case when the CLI is too old to emit the fields on search), render exactly as above: no badge, no change.

Git-changes mode — two render additions:

• Under each item add a _via: <dir>_ line naming the changed directory that surfaced it (on multiple matches, the most specific dir).
• Replace the footer with _Classified as: git-changes — <D> dirs._ (append (capped at 20 of <M>) when __TOTAL__ exceeds 20).

If all sections are empty, fall through to Step 6.

Step 6: Empty state

• Call mcp__archcore__list_documents(limit=1). If the knowledge base is empty → "No documents indexed yet. Run /archcore:capture or /archcore:decide to seed .archcore/."
• Otherwise → "No documents reference <scope>. Consider using @<scope> in a rule or ADR so future work in this area surfaces the context automatically."

Resolver sentinels:

• __CLEAN__ → "Working tree is clean — no staged, unstaged, or untracked changes to scope."
• __NOT_REPO__ / __NO_GIT__ → "Git scope unavailable here — pass a path or topic instead."
• Resolver returned directories but no documents matched → use the "No documents reference …" message above, scoped by the directory list.

Result

A grouped markdown surface of the rules, ADRs, specs, patterns, reference docs (doc, rfc, orphan guide), and in-progress work that applies to the requested scope — or a pickup summary of draft work + recent accepted decisions and rules when called with no argument. With --git-changes, the scope is derived from the working tree (staged, unstaged, and untracked changes) and the footer reports the directory count. A classification footer identifies which mode was chosen.