Agent Skills: /smith-recon — guided, bounded multi-source investigation

/smith-recon — guided, bounded multi-source investigation

Hydrate context for a topic from chosen sources, then synthesize a verified, evidence-linked brief. Designed to be GUIDED and CHEAP — it asks scope before sweeping, rather than blindly firing a large subagent fleet (costly and low-quality). Argument = the topic/question.

A common case is resuming a work thread: /smith-recon "resume my work thread on «topic/ticket»: pull the tracker issue, any open stacked PRs, related docs, and chat references, then brief where I left off". The sources below already cover this — no separate "resume" command is needed.

Procedure

1. Scope first (ask ONE thing) — the argument is the precise question; output language defaults to zh-Hant (user preference). The single scope decision to ask is which sources to sweep — a multiSelect over jsonl convo history, Serena + Basic-Memory, Notion, Slack, Jira, Google Drive, GitHub (one coherent choice, not bundled with other decisions — see @smith-ctx-claude one-scope-decision-per-turn). Do NOT presume the source set. Confirm depth/breadth only if the chosen sources leave it ambiguous.
2. Plan the sweep — list the concrete reads per chosen source. Cap fan-out: prefer a few targeted reads (and, when iterating, a /loop/ralph loop that re-shares the relevant smith skills + topic + accumulated context — @smith-ralph, @smith-automation) over many one-shot parallel subagents. Load @smith-research/@smith-validation (verify before asserting) and @smith-subagents (returns are claims; restate in-band).
3. Read bounded — for noisy history use a scripted full scan (counts + denominator), not sampling (@smith-guidance Scope Verification: close gaps, don't just disclose them). Quote source lines / URLs as you go.
4. Cross-verify — reconcile sources; flag contradictions; mark anything unverified explicitly. Verify suspicious claims before relying on them. Reconciling agreement between sources is not the same as trying to break a finding — do both (@smith-validation Adversarial Verification).
5. Brief — return a synthesis with evidence URLs/file:line for every claim, in the requested language. Note any source deliberately not covered and why (no silent omissions).

If the topic is broad, widen only after asking — never silently.