← The kit's docs / specs/

Commands and Agent Skills for context-v

Catalog of every command, slash command, and agent skill the kit ships — surfaced natively via `.claude/commands/` and `.claude/skills/`.

Path
context-v/specs/commands-and-agent-skills-for-context-v.md
Authors
Michael Staton
Augmented with
Claude Code (Opus 4.7)
Tags
Context-Vigilance · Agent-Skills · Slash-Commands · Claude-Code · Spec-Kit · Reference-Catalog

Audience

Adopters of context-vigilance-kit — anyone who runs cv-init on a pseudomonorepo. Also: lead engineers onboarding a team to the kit, and agents (Claude Code or similar) that need to discover the kit's vocabulary at session start.

Purpose

A consolidated catalog of every command, slash command, and agent skill the kit ships, with status, trigger, signature, output, and cross-references. The companion spec Systematizing Chroma as Loading Mechanism for Context V defines the system behavior (three loops, applicability, philosophy). This spec defines the surface area users and agents discover.

Two reasons this catalog matters:

  1. Adoption clarity. New users need a single place to see "what's in here, what does it do, when does it fire."

  2. Agent discoverability. Claude Code auto-loads .claude/skills/ and .claude/commands/ from project scope at session start. Spec-Kit (GitHub's spec-driven-development library) uses this exact mechanism to make commands available natively. The kit follows the same convention so adopting agents see the surface without extra wiring.

Loading model — how agents and users discover kit assets

The kit ships three classes of agent-facing assets, each with a different discovery mechanism:

Asset class Lives in Discovery mechanism
Agent Skills <senior-root>/.claude/skills/<skill-name>/SKILL.md (project scope) OR ~/.claude/skills/<skill-name>/SKILL.md (user scope, per feedback skill authoring in lossless skills) Claude Code auto-loads; activates when the user's intent matches the skill's description field
Slash Commands <senior-root>/.claude/commands/<command-name>.md Claude Code auto-loads; appear in the slash menu and as /<name> invocations
CLI Commands <kit_root>/scripts/<command> (Lossless: ai-labs/context-vigilance-kit/scripts/ ) Invoked in the terminal by the user; agents invoke via Bash

cv-init wires all three so adopting trees have the catalog active from the first session after init.

Agent Skills

Three skills total. One existing; two introduced by the Systematizing Chroma as Loading Mechanism for Context V spec.

search-lossless-corpus

Field Value
Status Existing (implemented, in active use)
Trigger shape Q&A phrasing — "did we already…" , "what did we decide…" , "how did we fix…" , "where did we put…"
Auto-fire? No — activates by intent match against user message
Purpose Ground answers about prior decisions, shipped work, history, failures, and locations in the corpus instead of training-data folklore
Output Cited answer ( source_path + timestamp + repo slug per claim); never a chunk dump
Discipline Decompose → execute → evaluate → synthesize; cap at 5 Chroma queries per question
Skill file context-v/skills/search-lossless-corpus/SKILL.md
Scoped to All four collections (selected per sub-query, not fanned out)

Not changed by this kit version. Stays scoped to Q&A grounding; the new skills handle session-start loading and recency catch-up. Anti-pattern : don't chain into kickoff/catch-up on the same trigger — distinct surfaces.

kickoff-context-v

Field Value
Status Planned (v0 implementation in Systematizing Chroma as Loading Mechanism for Context V Loop 1)
Trigger shape Auto-fires at session start when applicability passes AND opening user message lacks context-v file references
Auto-fire? Yes — agent enforces the discipline humans skip
Purpose Replace hand-loading and dir-scan with semantic file selection while keeping the user as the picker
Output Files loaded into context (via Read ) after the user picks from a three-bucket triage
Flow Agent asks "any known relevant context-v files?" → user lists files OR says "help me find" OR "no context needed" → on help-me-find, agent queries context-vigilance-corpus + changelog-corpus (roles), groups hits by source_path , presents three buckets (Hard include / Maybe — have on hand / Ignore), user picks, agent Read s the picks
Distance threshold None — conversational evaluation does the work; no hard numeric trip-wire
No-match fallback Loud — agent says so and falls back to dir-scan
Skill file context-v/skills/kickoff-context-v/SKILL.md (to be authored in lossless-skills)
Scoped to context-vigilance-corpus + changelog-corpus (roles, resolved via config.collections.* )

catch-up-on-context-v

Field Value
Status Planned (v0 implementation in Systematizing Chroma as Loading Mechanism for Context V Loop 3)
Trigger shape Auto-fires as nudge at session start (after kickoff resolves) when delta-meaningfulness threshold clears AND nudge.state == "on" AND now > muted_until ; also via explicit slash command
Auto-fire? Yes (nudge policy — auto opt-in, mutable)
Purpose Keep users ambiently aware of recent context-v / changelog activity; meeting-prep digest for distributed teams
Output Three-level digest (repo → tag → files, recency desc) OR overflow-mode truncation with auto-narrow prompt for wide windows
Default nudge threshold Last session >3 days ago OR ≥5 context-v files modified in last 2 weeks
Default window (nudge) 2 weeks
Mute mechanism mute-for-today / mute-for-week / cv-config nudge off
Overflow handling Default cap 50 files; auto-narrow prompt above
Skill file context-v/skills/catch-up-on-context-v/SKILL.md (to be authored in lossless-skills)
Scoped to context-vigilance-corpus + changelog-corpus (roles)

Slash Commands

Explicit, parameter-driven invocations of the retrieval surfaces. Live in <senior-root>/.claude/commands/*.md ; appear in the Claude Code slash menu.

Catch-up family

Command Signature Purpose Status
/catch-up-2w no args Fast trigger: 2-week recency digest Planned
/catch-up-1m no args Fast trigger: 1-month recency digest Planned
/catch-up --window <N{d|w|m|y}> Explicit window (3m/6m/1y trigger overflow handling) Planned
/catch-up --scope <source_repo_slug> Limit digest to one repo's subtree Planned
/catch-up --by-author <name> Filter by author (requires fm_authors allowlist patch from Loop 2) Planned (v0 data, v1 UI)
/catch-up --max-files <N> Override overflow cap (default 50) Planned
/catch-up --since <ISO date> Explicit cutoff instead of window Planned
/catch-up --alphabetical Tag buckets in alphabetical order instead of most-recently-active Planned (v1 toggle)

Kickoff (rarely needed explicitly — the skill auto-fires)

Command Signature Purpose Status
/kickoff no args Manual re-fire of kickoff (after no context needed was said earlier in session, or after Claude Code's context compaction) Planned

CLI Commands

User-facing scripts under <kit_root>/scripts/ . Invoked from terminal; agents invoke via Bash . All write to <senior-root>/.context-vigilance/ unless flagged otherwise.

Setup and configuration

Command Signature Purpose Status
cv-init [--senior-root <path>] One-time ceremony: asks for / accepts senior root, writes .context-vigilance/config.json , the SessionStart hook in <senior-root>/.claude/settings.json , the wrapper script, and per-collection manifest skeletons Planned
cv-dismiss [path] Write .context-vigilance-skip marker at path (or cwd) so any session under it sees kit as not-applicable Planned
cv-undismiss [path] Remove the marker Planned
cv-config get <key> / set <key> <value> / list Read or write .context-vigilance/config.json entries — nudge state, collection name overrides, thresholds Planned
cv-help no args List all kit commands + skills + slash commands with one-line summary; pointer to this spec Planned (v0 stretch)

Ingestion

Command Signature Purpose Status
ingest-to-chroma.py [--changed-since] / [--force-rebuild] / [--reset] / [--limit <N>] / [--query "..."] / [--collection <name>] Ingest context-v markdown into context-vigilance-corpus (role). Existing script; --changed-since + sidecar manifest support added in v0. Existing + Patch
ingest-changelogs-to-chroma.py same flag shape Ingest <repo>/changelog/ entries into changelog-corpus (role) Existing + Patch
ingest-claude-sessions-to-chroma.py [--reset] / [--limit <N>] Opt-in, privacy-sensitive. Ingest Claude Code session transcripts into claude-code-sessions . Never auto-fired by the SessionStart hook. Existing
ingest-all.sh [--with-sessions] / [--with-traces] / [--with-claude] / [--reset] / [--dry-run] / [--only-context-v] / [--only-changelog] / [--only-claude] Master orchestrator. Default runs the two safe ingesters. Opt-in flags add privacy-sensitive collections. Existing

Diagnostics

Command Signature Purpose Status
smoke-test-chroma.py no args Verify Chroma connectivity, list collections, sample-query Existing

Hooks (not user-invoked, but worth cataloging)

Hook Trigger Purpose Status
<senior-root>/.context-vigilance/scripts/session-start.sh Claude Code SessionStart event Runs applicability check, then incremental --changed-since ingest of context-vigilance-corpus + changelog-corpus (roles). Best-effort; fails silently if Chroma unreachable. Never fires session/trace ingesters. Planned

Status summary

For each asset class, where v0 lands:

  • Already shipped: search-lossless-corpus skill; all ingest-*.py scripts; ingest-all.sh ; smoke-test-chroma.py .

  • v0 implementation (per Systematizing Chroma as Loading Mechanism for Context V implementation slice): kickoff-context-v skill, catch-up-on-context-v skill, cv-init , cv-dismiss , cv-undismiss , cv-config , the SessionStart hook + wrapper script, the --changed-since patches to the two ingesters, the fm_authors allowlist patch, the /catch-up-* and /kickoff slash commands.

  • v0 stretch: cv-help , hook verbose-logging mode.

  • v1+: /catch-up --by-author UI surface, theme summarization for wide windows, per-project last-session tracking, resume-after-compaction trigger, owner: schema activation, kit governance for multi-steward adopters, distribute kit as PyPI / Homebrew package.

Discoverability story

Three entry points for users encountering the kit:

  1. This spec as the canonical catalog (versioned, lives in context-v/specs/ ).

  2. cv-help as the at-the-terminal quick reference (v0 stretch).

  3. Claude Code's native slash menu for the /catch-up* and /kickoff commands (auto-discovered from <senior-root>/.claude/commands/ ).

For agents adopting Spec-Kit-style command loading: the kit's .claude/commands/*.md files are the contract. Agents that scan project-scope .claude/ directories at session start (Claude Code's default behavior) discover them automatically once cv-init has run. No additional registration step.

Naming conventions

  • Skill names: kebab-case verb-phrases ( kickoff-context-v , catch-up-on-context-v ). The verb signals the trigger shape (kickoff at session start; catch-up at intervals). Sibling names share construction ( -context-v suffix).

  • Slash commands: kebab-case, prefer fast-trigger short forms ( /catch-up-2w ) over flag-driven full forms when the parameter is fixed and frequently used.

  • CLI commands: cv- prefix for kit-level commands ( cv-init , cv-dismiss , cv-config ). Existing ingester scripts keep their descriptive Python-style names ( ingest-to-chroma.py ) since they predate the cv- namespace.

  • Systematizing Chroma as Loading Mechanism for Context V — companion spec defining the system behavior (three loops, applicability model, philosophy). This spec is the surface catalog; that spec is the system contract.

  • ChromaDB as Context Improvement Across Everything Everyone — parent exploration both specs close.

  • search lossless corpus — the existing skill cataloged above.

  • context vigilance — the directory-roles practice this kit supports.

  • pseudomonorepos — the tree-shape discipline defining "senior root."

  • chroma local — generic Chroma usage reference.

  • Profile Beads — prior art for the cv-init / cv-dismiss / cv-config family (Profile Beads's bd init / bd setup / bd remember family). The cv-help pattern echoes bd prime 's discoverability mechanism.

  • Spec-Kit (github.com/github/spec-kit) — GitHub's spec-driven-development library; the model for .claude/commands/ -based command discovery this catalog follows.

Outcome

(Open. Update as commands and skills are implemented per Systematizing Chroma as Loading Mechanism for Context V implementation slice. Mark status: Final when every Planned item above is implemented and at least one external adopter has discovered the surface via cv-help or the Claude Code slash menu on their own tree.)