Vestige

Mine recent local coding-agent transcripts for memories worth keeping, using the `vestige_scan_sessions` MCP tool. Fire at session start (to catch up on what happened in prior sessions), when the user says "scan my sessions", "what did we work on recently?", "catch up from my transcripts", "ingest recent sessions", "mine my history", or when picking up a project after a pause. The tool hands you redacted, cursor-advanced turns from this project's Claude Code and Codex transcripts; you extract decisions/notes/preferences/questions inline and file each via `vestige_propose_candidate`. Opt-in — requires `[mcp] allow_scan_sessions = true` in `.vestige/config.toml`. Candidates land in the assimilation inbox for human review, never as durable memories directly.

Use this skill when you need the raw source receipts for a memory or candidate — the typed evidence rows, not the narrative walk. Fire when the user says "what sources back mem_X?", "show me the source receipts", "list the sources for mem_X", or "filter sources by kind agent_session". Use `vestige-why` when you want the full provenance narrative including events; use this skill when you want the tabular source index only.

Use this skill when you want to audit what queries the agent or CLI has run against project memory. Fire when the user says "what queries did the agent run?", "show recent trace events", "did anyone search for X recently?", or "list traces by caller mcp". Every vestige search, recall, expand, and context call writes a query_events row automatically — this skill surfaces them. Use vestige-trace-show to expand a single trace.

Use this skill to re-run a stored trace against the current memory store and diff the results against the original — answering "would the agent get the same answer today?". Fire when the user says "re-run trace_X", "replay that search", "diff trace_X against current state", or "did the answers drift?". Read-only: writes a NEW query_events row tagged with replay_of; never mutates the original trace or any memory. provider_match=false flags embedding drift.

Use this skill when you want to expand a specific trace to its full detail — the query text, mode requested vs resolved, provider, result IDs, and scores. Fire when the user says "expand trace_X", "show me what trace_X returned", or "what did that search find?". Use vestige-trace-list first to find the trace_id, then this skill to inspect one in full.

Use this skill when you want to know where a memory came from, need to explain its provenance, or must audit what was captured and when. Fire when the user says "where did this come from?", "explain that memory", "why do we have mem_X?", "show provenance for X", "what's the source of mem_X?", or "trace this memory back". Works for both memories and pre-approval candidates. Soft-deleted memories are included — provenance is always inspectable.

Review Vestige code against this repo's CODESTYLE.md, CLAUDE.md, and PRD — naming, progressive disclosure, error handling, agentic patterns, and the seven hard architectural rules (soft-delete only, project-scope boundary, immutable migrations, one-way crate deps, newtype IDs, MCP intent-not-mechanics, no daemon). Use when the user says "vestige-review", "review this", "review my branch", "check my code", "look over this PR/diff", "roast this", "is this ready to merge", or after finishing work in any Vestige crate. Prefer over the global code-review skill inside this repo — it layers Vestige-specific rules on top of the style guide and produces findings grouped by severity (blocker / major / minor / nit) with file:line citations and concrete fixes.

Fires automatically — without the user asking — when the conversation produces something a future session would want. Watch for decisions ("we'll go with…", "settled — X"), preferences ("I prefer…", "always…", "never…", "make sure to…"), open questions ("TBD", "unclear whether…", "follow-up:"), TILs / aha moments, gotchas ("careful — X breaks Y"), smells, workarounds, root causes, and broken assumptions. Probe `vestige recall` first to dedup, then dispatch to `vestige <subcommand> add`. Surface: `Recorded <kind> mem_<ULID>.` Do not wait for "remember this".

Render the project's Vestige context pack — summary, recent decisions, open questions, recent activity — as a budget-bounded brief. Use at session start, before unfamiliar work, after a long pause, or when the user says "what's the state of this project?", "where are we?", "catch me up", "what have we decided?", "what's outstanding?", "give me the rundown". Returns a JSON envelope with `summary`, `decisions[]`, `open_questions[]`, `recent[]`. Token budget defaults to 1200 — pass `--budget-tokens` to adjust.

Soft-delete a Vestige memory by its handle (`mem_<ULID>`) when the memory is wrong, superseded, or stale. Fire when a decision has been reversed, a note is no longer accurate, an open question has been answered, or the user says "forget memory <id>", "that's outdated", "remove <id>", "drop that memory", "we don't do that anymore". Flips status to `deleted` and drops the memory from search results — the row stays in the journal. Reversible via `vestige-restore`.

Search Vestige memory before a decision, refactor, or library choice — any time you suspect "we already discussed this". Fire when the user says "have we decided this?", "what did we say about X?", "look up our notes on Y", "is this already settled?", "search memory for Z". Lexical (BM25) by default; `--hybrid` for merged lexical + semantic recall (requires `vestige embed --all`); `--semantic` for cosine-only. Returns ranked compact cards with `mem_<ULID>` handles. Use `vestige-show <id>` to expand.

Capture a project decision to Vestige memory when committing to an architectural choice, picking approach A over B, settling a tradeoff, or choosing a library / framework / pattern. Fire on "we'll go with…", "let's use…", "decision:…", "I'm choosing X over Y because…", "settled — we use…", "going with X", or when the user says "capture that decision", "remember this", "record it". Decisions ground future reasoning and feed every `vestige context` pack. Captures body, rationale (the why), and importance (default 0.7). Returns the new handle (`mem_<ULID>`).

Capture a general project note to Vestige memory for non-trivial facts, workarounds, gotchas, aha moments, TILs, code smells, or durable TODOs — anything worth keeping that isn't a decision, preference, or open question. Fire on "good to know:", "TIL", "note that…", "heads up:", "FYI:", "turns out…", "the reason X happens is…", "TODO:", "smell:", "careful — X", or when the user says "jot this down", "note that down", "remember this fact". Default importance 0.5. Returns the new handle (`mem_<ULID>`).

Capture a user preference to Vestige memory when the human expresses a convention, opinion, or "how I like things done". Fire on "I prefer X", "always use Y", "don't do Z", "we always…", "never…", "I like…", "I don't like…", "make sure to…", "do not…", "convention:…", "rule:…", "house style:", "in this project we…", or when correcting your own output to match what the user just said. Preferences constrain every future agent run via the `vestige context` pack. Default importance 0.6. Returns the new handle (`mem_<ULID>`).

Capture an open question to Vestige memory when an ambiguity can't be resolved now and needs a human, investigation, or later session. Fire on "TBD:", "open question:", "we should figure out…", "unclear whether…", "follow-up:", "to investigate:", "leaving X open", or when the user says "capture that as a question", "we'll come back to that", "park that for now", "good question — note it down". Questions surface in every `vestige context` pack until `vestige-forget`'d (typically when answered via `vestige-record-decision`). Default importance 0.5. Returns the new handle (`mem_<ULID>`).

Restore a previously soft-deleted Vestige memory by its handle (`mem_<ULID>`). Use when a memory was forgotten in error, when the situation has reverted (a decision that was reversed gets re-affirmed), or the user says "bring back <id>", "restore <id>", "undo that forget", "that memory should still be there". Re-flips status from `deleted` to `active` and re-indexes the memory's representations into FTS. Note — embeddings are left stale by restore (they'll re-embed on the next `vestige embed` run).

Expand a Vestige memory at higher fidelity once you have its handle (a `mem_<ULID>` string). Use this skill after `vestige-recall` returns a card whose `one_liner` isn't enough, when the user says "show me memory <id>", "expand <id>", "what does <id> say in full?", "read me memory <id>", or whenever you need the *body*, *rationale*, or *source content* of a memory rather than the compact handle. Depth ladder — one_liner → summary → compressed → full. Default is summary; pass `--depth full` to read the entire body.