Skip to content

Plugins

The marketplace ships fifteen plugins. The spine is retrieval — a routing agent that picks and composes modalities — surrounded by plugins for orchestration, steering, verification and impact analysis, multi-lens review, exhaustive corpus review, controlled runtime evidence, token economics, cross-session handoff, and authoring.

Start from the task

Use the cookbook for multi-plugin journeys, the security guide before enabling execution or retention, and troubleshooting for first-run and lifecycle work.

  • retrieval-core

    The spine: a retrieval-strategist agent + retrieval-strategy skill that choose and compose modalities. Other plugins depend on it.

    retrieval · shipped

  • code-search

    Lexical, structural, code-intelligence, structured-data, history, rewrite, metrics, and non-code doc search. Two skills split by corpus.

    retrieval · shipped

  • indexkit

    Local-first semantic RAG: a bin/indexkit CLI with a configurable Ollama endpoint, turbovec vectors, opt-in FTS5/BM25 reciprocal-rank fusion, and hybrid --allowlist retrieval.

    retrieval · shipped

  • obsidian

    Skill-only RAG bridge: turn a vault's graph and tags into a candidate set fed to indexkit.

    retrieval · shipped

  • adr-bridge

    Bridge an adrkit decision corpus to context-kit's durable memory and semantic retrieval without duplicating adrkit's own agent plugin.

    retrieval · shipped

  • plan-execute

    Plan-big / execute-small orchestration: a strong model plans and delegates token-heavy work to a cheaper executor.

    orchestration · shipped

  • context-steering

    Place guidance at the cheapest layer that still fires — memory, rules, skills, subagents, MCP servers, or hooks.

    steering · shipped

  • verify

    Read-only per-claim verification plus prospective change-impact and blast-radius analysis, with an enforced non-mutating inspection runner.

    verification · shipped

  • runtime-evidence

    Controlled runtime observation after static verification cannot settle a claim, using exact allowlisted command IDs with bounded artifacts — or an approved optional tool when no command fits.

    verification · shipped

  • corpus-review

    Exhaustive review of a corpus too large to read at once — hashed unit inventory, bounded resumable shards, and a provable coverage ledger.

    verification · shipped

  • deep-review

    Multi-lens evaluative critique: independent charters return typed findings, agreement merges into confidence, disagreement survives as an explicit tradeoff.

    verification · shipped

  • context-handoff

    Manual-first, bounded task-state handoffs with a read-only compiler and deterministic provenance/freshness validation.

    continuity · shipped

  • memory

    Reviewed durable memories with provenance, cue anchors, freshness, supersession, and an optional project-isolated MemPalace provider.

    continuity · shipped

  • token-economics

    Measure token spend from local host records, and prove a tool's savings with a controlled A/B that must preserve the answer.

    measurement · shipped

  • plugin-forge

    Author portable plugins with scaffolding, manifest/frontmatter checks, and a deterministic aggregate catalog discovery-quality gate.

    authoring · shipped

How they fit together

graph TD
    RC[retrieval-core<br/>routing agent + decision flow]
    CS[code-search]
    LR[indexkit]
    OB[obsidian]
    VF[verify]
    RE[runtime-evidence]
    CR[corpus-review]
    DR[deep-review]
    CH[context-handoff]
    MM[memory]
    PE[plan-execute]
    CST[context-steering]
    PF[plugin-forge]

    CS -->|depends on| RC
    VF -->|depends on| RC
    RE -->|depends on| VF
    CH -->|depends on| VF
    MM -->|depends on| CH
    CR -->|depends on| VF
    CR -->|depends on| PE
    DR -->|depends on| VF
    DR -->|depends on| PE
    OB -->|feeds --allowlist| LR
    RC -.composes.-> CS
    RC -.composes.-> LR
    RC -.routes recall to.-> MM
    RC -.routes coverage to.-> CR

    classDef spine fill:#4f46e5,stroke:#4338ca,color:#fff;
    class RC spine;
  • code-search and verify depend on retrieval-core.
  • runtime-evidence and context-handoff depend on verify, so installing either transitively pulls the spine.
  • memory depends on context-handoff, so it also pulls verify and the retrieval spine. Handoffs remain authoritative current task state; archived copies are historical evidence.
  • corpus-review depends on verify and plan-execute: it raises verify's unable-to-check discipline from one claim to a whole corpus, and reuses plan-execute's cheap-worker fan-out to read shards. Retrieval surfaces the relevant; corpus review proves the exhaustive.
  • deep-review depends on verify and plan-execute: it deliberately refuses to settle its own DEFECT findings and routes them to verify as claims, and reuses plan-execute's fan-out to run lenses independently. Verification asks whether a claim is true; deep review asks whether the work is good.
  • obsidian and indexkit pair: the bridge produces candidate note paths that feed indexkit's hybrid --allowlist search.
  • plan-execute, context-steering, and plugin-forge are independent — orchestration, steering, and authoring around the retrieval core. verify can optionally use plan-execute for broad read-only impact coverage, but does not depend on it.

Dependencies at a glance

Plugin Category Ships Depends on
retrieval-core retrieval agent + skill —
code-search retrieval 2 skills + tool checker retrieval-core
indexkit retrieval bin/indexkit CLI + skill ollama + turbovec + SQLite FTS5 for --hybrid
obsidian retrieval skill only indexkit (runtime)
plan-execute orchestration skill + command + workflow + subagent —
context-steering steering skill + examples —
verify verification subagent + 2 skills + command + stdlib inspection runner retrieval-core
runtime-evidence verification skill + command + subagent + stdlib runner verify → retrieval-core
corpus-review verification skill + command + subagent + 3 stdlib scripts plan-execute, verify → retrieval-core
deep-review verification skill + command + subagent + stdlib adjudicator plan-execute, verify → retrieval-core
context-handoff continuity skill + 2 commands + subagent + stdlib validator verify → retrieval-core
memory continuity skill + 4 commands + stdlib adapter + opt-in Claude hooks context-handoff → verify → retrieval-core; MemPalace optional
token-economics measurement 2 skills + 2 commands + 2 stdlib scripts verify → retrieval-core
plugin-forge authoring skill + command + validators/tests —