A local ZCode plugin of research subagents for repo, markets, and collection-pipeline work. Thirty agents. Twenty-one are read-only: they gather and report, and the main thread decides and edits. Nine carry chartered write lanes. claw (v0.8.0) executes the Calibre acquisition pipeline's file-side halves through the pipeline's own structural gates. Five audit-and-fix seats (v0.9.0: charter-architect, ci-posture-auditor, doc-drift-auditor, debt-census, test-gap-analyst) may apply the fixes their own audits find under the charter-fence model: report-only by default, a clean-tree precondition, writes confined to a charter-named scope, re-verification after applying, an itemized report, and never a commit. Two purpose-built write seats (v1.0.0): latro writes deep web research up as one new cited markdown report per dispatch (no Edit tool, overwrite refused, keyless sources only), and assay grades a whole repo on five metrics and appends a dated goals block to its roadmap.md (append-only, re-verified zero-deletion) or grades an idea read-only on four criteria. A third purpose-built seat (v1.1.0): imprint tends a repository's GitHub identity, the description and topics, through gh verbs alone; propose mode is the default, apply dispatches name their repos, every mutation is read back and itemized, and its writes are the roster's first that publish to github.com instantly with no local diff to review. The main thread reviews every diff and owns the release. No agent spawns subagents of its own. Every agent closes each dispatch with a ## Debug Report block of the friction it actually hit (tool, environment, model, charter, dispatch, idea categories, or - none on a clean run); the dispatcher files actionable findings as GitHub issues in this repo, so the roster keeps a dated record of what to improve.
| Agent | Job |
|---|---|
repo-cartographer |
Structured repo map (layout, stack, conventions, tests, risk areas) before planning sessions. |
doc-drift-auditor |
Standard-layout docs vs code reality: VERSION sync, roadmap boxes, spec claims, patchnotes hygiene; apply mode fixes the mechanical drift it found. |
cascade-checker |
Downstream call-site inventory for a changed shared-library API, with per-repo breakage risk. |
spec-compliance-reviewer |
Diff vs spec.md contract; violations and unaddressed claims with path:line evidence. |
ci-triage |
Failed CI run via gh to root cause, decisive log lines, and fix direction. |
slop-reader |
Fresh-eyes AI-prose pass on commissioned docs; flags quotes, never rewrites. |
release-auditor |
Pre-tag readiness and post-tag sweep: VERSION carriers, patchnotes entry style, tag message health. |
workspace-sentinel |
One sweep, every owned repo: dirty trees, unpushed commits and tags, untagged releases, stale catalog rows. |
ci-posture-auditor |
Workflows vs the house-hardened CI shape (SHA pins, permissions, concurrency, timeouts); apply mode writes the hardening, tightening only. |
debt-census |
TODO/FIXME/HACK and dead-reference census, dated by blame, as a cleanup worklist; apply mode removes named census items. |
code-reviewer |
Bug-hunting review of a diff: correctness, edge cases, scope creep, userspace-break risk. |
git-archaeologist |
When did this break and why: regression windows and provenance from log, blame, pickaxe; bisect plans, never executed. |
claim-verifier |
Independent re-verification of a findings list: confirmed / refuted / unverifiable per claim. |
dependency-auditor |
Declared vs imported, floors vs used APIs, stdlib purity, Flatpak vendor staleness, toolchain drift. |
test-gap-analyst |
Coverage map: untested behaviors, suite-discovery failures, duplicated test files; write mode adds the named missing tests and runs exactly those. |
duplication-scout |
Cross-repo similar-module detection feeding the library-graduation rule; evidence, never the recommendation. |
stock-broker |
Markets research briefs: companies, ETFs, peers, and the surrounding markets (rates, FX, commodities) from keyless public sources; portfolio-relative guidance, never execution. |
ledger-analyst |
Personal-finance briefs from hledger journals: cashflow, budget vs actual, net worth, cost basis; encrypted org files fenced off by charter. |
librarian |
Books-collection research over the Calibre library: catalog health, wings and tag tree, series gaps, ownership duplicates, EPUB audits, taxonomy drift; book IDs and dated counts, never writes. |
data-analyst |
Read-only duckdb profiling of named local datasets: schema, nulls, top-N, joins; compact findings without flooding the main thread. |
security-auditor |
Defensive review of owned repos: gitleaks (redacted), risky patterns, advisory cross-checks via gh; ranked findings only. |
charter-architect |
Designs new agents: scorecard verdict (declining is valid), live-probed charter draft, and the full release kit; apply mode lands the approved kit in the repo and validates it. |
syshealth-auditor |
Backup and system-health census: borg freshness from state reads, sys_maintain staleness and failures, failed units, disk space; root-gated checks reported as directions. |
schedule-analyst |
Weekly school-and-life planning digest: Todoist via td (overdue triage, per-course and per-project load, streaks) cross-referenced with gcalcli agendas; recommendations only, never task writes; reports calendar-unavailable honestly while the OAuth session is down. |
music-archivist |
Music-collection research over lattice's read-only audits: health digest, genre hygiene against taxonomy.md, duplicates, codec coverage, discography gaps, listening candidates; album paths, never a tag edit. |
claw |
Executes Calibre import phases 1 and 2: vets a downloads batch (battery, dup screen, gated repairs, stamps, staging, manifest) and imports the signed manifest; decision points return as prompts, the seal stays with the main thread. |
shadow |
Read-only compliance pass over one import phase: manifest versus disk, verdicts versus decisions, imported ID sets, stamp read-backs, bak diffs; blocking findings before sign or commit. |
latro |
Deep web research written up as one new cited markdown report per dispatch: project technologies, book acquisition, papers, languages, web design; WebSearch/WebFetch plus a keyless endpoint ladder (arXiv, Crossref, OpenLibrary); overwrite refused, never authenticates. |
assay |
Whole-repo report card on five metrics (code quality, code cleanliness, repo cleanliness, documentation, cohesion) with an append-only, zero-deletion goals block for roadmap.md; read-only idea grading on four criteria with web prior art and a worth-it verdict. |
imprint |
Tends a repo's GitHub identity: audits the description and topics against what the repo actually is (grounded in README, spec, history) and on apply dispatches updates exactly those two fields via gh, read back and itemized before/after; never files, tags, releases, or issues. |
In ZCode: Settings → Plugin Management → Discover tab → + → add this directory as a local marketplace → install dragon-agents. Plugins from local (cache) marketplaces like this one install disabled until first enabled; only bundled official marketplace plugins are enabled by default. After enabling, trust the enabledPlugins key in ~/.zcode/cli/config.json and a canary dispatch over the UI toggle, and remember the agent roster only lands in conversations opened after an app restart.
Any text model works; the agents were written for text-only subagents (no vision). Flash-tier models are the sensible default for read-only sweeps; on analysis-heavy dispatches (cascade-checker, spec-compliance-reviewer, stock-broker, ledger-analyst, security-auditor, charter-architect, claw, shadow, latro, assay, imprint) a stronger model is worth it, and the nine write-capable agents (claw, charter-architect, ci-posture-auditor, doc-drift-auditor, debt-census, test-gap-analyst, latro, assay, imprint) belong on the strong tier: writes and verdicts that gate work are judgment. Per-agent model picks live in Settings → Subagents.
python3 scripts/validate.py checks both manifests and every agent charter: frontmatter fields, per-class tools ([Read, Bash] for read-only agents, [Read, Bash, WebSearch, WebFetch] for assay, [Read, Write, Bash, WebSearch, WebFetch] for latro, [Read, Edit, Write, Bash] for the five write seats, [Read, Bash] for imprint, whose writes are fenced to the two gh metadata verbs), the read-only line for readers, the Clean tree and never-commit fences for writers, the new-file and never-authenticate fences for latro, the append-only fence for assay, the metadata-lane, two-mutations, never-commit, and forbidden-mutations fences for imprint, the no-subagent line everywhere, the ## Debug reporting duty on every charter, and the thirty-agent roster. CI runs the same script on every push. AGENTS.md documents the plugin's rules and the marketplace/runtime gotchas; read it before changing how the plugin is packaged.
MIT. See LICENSE.