brief - SKILL.md Agent Skill

name: brief description: > Session-start briefing from Wenlan. Reads the project status file (the /handoff-maintained ledger of Active/Backlog work), then loads identity, preferences, and topic-relevant memories so the agent walks in with context. Surfaces any memories the daemon has flagged for human revision before the session uses them. Invoked as `/brief [topic]`. Call FIRST at session start, before any other Wenlan verb. argument-hint: "[topic]" allowed-tools: ["Bash", "mcp__plugin_wenlan_wenlancontext", "mcpplugin_wenlan_wenlanrecall", "mcpplugin_wenlan_wenlan__list_pending_revisions", "mcp__plugin_wenlan_wenlan__accept_revision", "mcp__plugin_wenlan_wenlan__dismiss_revision"]

/brief

Pull a curated session brief from Wenlan. Three sources, in order:

Project status file — what /handoff last wrote. Authoritative for "what's left to do" right now.
context MCP — identity, preferences, topic-relevant memories. Background, not ledger.
list_pending_revisions — daemon-flagged memories awaiting review.

Status file wins on "what's next" because memories rank by topic similarity and surface stale items alongside fresh ones; the status file is the live ledger maintained per session.

1. Read project status file first

Detect project root:

Bash: cd_repo=$(git -C "$PWD" rev-parse --show-toplevel 2>/dev/null); echo "${cd_repo:-no-git}"

If output is a path → <project> = basename (e.g. wenlan).
If no-git → <project> = cwd basename.

Read ~/.wenlan/sessions/_status/<project>.md:

Bash: cat ~/.wenlan/sessions/_status/<project>.md

If the file exists, render its ## Last session, ## Active, and ## Backlog sections verbatim at the top of the brief output, under a heading like Status (last session <date>). This is the authoritative "what's left" frame.

If the file is missing, say nothing about it. First-time projects haven't been handed off yet.

2. Resolve the active space

Before any MCP call, resolve the active space by invoking the bundled resolver script via Bash:

"$CLAUDE_PLUGIN_ROOT/bin/resolve-space.sh" --cwd "$PWD" \
    ${SPACE_ARG:+--arg "$SPACE_ARG"} \
    ${TOPIC_ARG:+--topic "$TOPIC_ARG"}

The script prints <space>\t<source-layer> on stdout. Capture both: the <space> value is what you pass as space=... to MCP tools; the <source-layer> value is one of arg, env, cwd-config, cwd-repo, topic, default.

Print one line to the user before the MCP call:

Resolved space: <space> (from <source-layer>)

so the user can confirm the resolution before the brief proceeds.

3. Call context

Call the wenlan MCP server's context tool. If the user passed a topic argument, pass it through. Otherwise infer scope from the working directory and the conversation so far — don't ask the user.

context(topic="<args or inferred>", space=<resolved>)

Scope inference rules:

topic: if user omitted args, pass the most recent topic from the conversation (file or feature being discussed), or omit for a fresh general brief at session start.
space: use the value from the resolver script above. Always pass it; the resolver guarantees a value (falling back to "default" if nothing else matches).

When to use

Session start — call BEFORE any other Wenlan tool.
Major topic shift mid-session.
User says "catch me up", "what's the background on X", "remind me about Y".
Mid-session check-in to confirm assumptions.

When NOT to use

Specific factual lookup → use /recall (more targeted).
Storing a new memory → use /capture.
End of session → use /handoff.

How to use the result

Treat the status file as the live ledger (what's next), and the context memories as background (how the user thinks). When the status file says an item is done but a memory still says it's pending, trust the status file — memories don't auto-supersede. If you see drift, flag it inline rather than parrot the stale memory.

Model how the user thinks. Their preferences, corrections, and past decisions tell you how they want to be helped, not just what they already know. Don't just look things up: adjust your behavior.

4. Pending revisions check

After loading context, call:

list_pending_revisions(limit=10)

If the result is empty, say nothing. Do not print "0 pending revisions" or any "all clear" line. The brief is already noisy enough.

If the MCP call errors, log a one-line warning and continue. Do not error out the brief.

If the result is non-empty, render a block at the end of the brief (after the context summary, before handing back to the user). The daemon returns rows already sorted by last_modified DESC; just slice the first three.

Each PendingRevisionItem carries target_source_id, revision_source_id, revision_content, source_agent, and last_modified. The original memory's content is not on this response. The block displays the proposed revision text and the target memory id; if the user wants to see the original before deciding, they can run /recall <target_source_id> first.

Pending revisions (<N> total, top 3 shown):

1. target: mem_abc123  (proposed by <source_agent or "daemon">)
   revision: "Wenlan uses Turso libSQL fork (libSQL-server) for vectors..."
   Action: accept (replace original) | dismiss (drop revision) | skip

2. ...

Inline accept/dismiss verbs map to:

accept: accept_revision(target_source_id="<id>")
dismiss: dismiss_revision(target_source_id="<id>")
skip: no call; the revision stays pending for the next /brief

If <N> > 3, end the block with one line:

Run `/review revisions` to walk the rest.

Do not auto-action anything. The user picks per item.

Note: this block surfaces all pending revisions, not just this session's, because revisions are about memories you may still be using right now, regardless of when the contradiction was flagged.