Claude Code Plugin

Claude Code plugin

Install it with /plugin marketplace add ratomirjovanovic/pandaward and then /plugin install pandaward@pandaward (macOS, Linux) or /plugin install pandaward-windows@pandaward (Windows), and restart Claude Code. Details, updating and removal are on Installation.

Hands-free use

You do not need the commands below. Ask Claude to check the project's security (or run /pandaward:start) and it does the rest:

1. Scans the project, has read-only specialist agents review what automation could not decide, triages the results, builds the fix plan and writes the reports, with a live checklist.
2. Asks one question at the end: fix the plan (all steps or the first N), record the triage results, install missing optional scanners, or nothing.
3. Carries out what you chose without further prompts: one fix at a time with a retest after each, disposition changes through the confirmation prompt described under Hooks, and fresh reports at the end.
4. Keeps watching for the rest of the session: after every turn that edits code, the Stop hook retests the changed files and Claude tells you about anything new or reopened.

What stays with you: revoking leaked secrets at the provider, confirming disposition changes in the prompt, product decisions a fix depends on, and network tests, which need a scope you approve in your own terminal (Network Testing).

Asking in your own words works for single steps too, for example "what is still untested?" or "explain finding PW-0007".

Skills

Skill	Purpose
/pandaward:start [quick|deep]	The whole assessment with a visible checklist, then one question about what to do next, carried out for you (see Hands-free use)
/pandaward:init	Initialize and summarize the inventory
/pandaward:plan	Explain the plan and suspicious N/A decisions
/pandaward:scan	Passive scan with an honest summary
/pandaward:status	Resume from the ledger
/pandaward:findings [ID]	List or explain findings with evidence
/pandaward:review [domain|triage]	Manual-assisted review by read-only agents
/pandaward:fix-plan	Remediation plan; asks once which steps to apply, then applies them one by one with a retest after each
/pandaward:retest	Targeted re-scan of what changed, then the delta; Claude also runs it on its own after editing code
/pandaward:graph	Cross-domain graph, optionally as mermaid
/pandaward:report	Render reports
/pandaward:doctor	Tool check
/pandaward:test	Network test flow; only you can start it
/pandaward:feedback [PW-ID|missed]	Prepare a report about a wrong or missed finding for the Pandaward project; you review and submit it

Agents

All agents are read-only and record only hypotheses.

Agent	Domains
pandaward:architect	A: inventory gaps, trust boundaries, threat model
pandaward:identity	B, C: authentication, sessions, authorization, tenant isolation
pandaward:api	D, F, G, H: injection paths, webhooks, business logic, frontend sinks
pandaward:cloud	I, J, N, O: supply chain, IaC, CI/CD, logging
pandaward:mobile	K, L: Android and iOS static review
pandaward:ai	P: prompt injection paths, tool authorization, MCP
pandaward:triage	Verifies suspected findings and recommends a disposition

Hooks

Event	Behavior
SessionStart	Adds a short ledger summary (latest assessment, coverage, open findings) and the suggested next step, which Claude offers and runs instead of asking you to type it
PreToolUse	Asks before approvals, network tests, disposition changes, direct ledger access and manifest edits; denies edits to .pandaward/private/
PostToolUse	Marks edited files as stale so coverage is not overstated
Stop	After a turn that edited files, runs a changed-only passive retest (at most 40 seconds) and hands new or reopened findings back to Claude
PreCompact	Records a checkpoint event; the ledger stays authoritative

All hooks are silent when the project is not initialized, and none sends network traffic.

The Stop hook is silent when nothing was edited or the retest found nothing new or reopened. If the last assessment took longer than the 40-second budget, it skips the retest and says the changed files are STALE; if a retest runs out of time or an adapter fails, it reports the assessment as INCOMPLETE. It never reports a finding as fixed unless the ledger marks it REMEDIATED. Turn it off with hooks.auto_retest: false (Configuration).

MCP tools

Tool	Arguments	Effect
pandaward_get_status		read
pandaward_list_coverage_gaps	domain	read
pandaward_get_control	control_id	read
pandaward_get_findings	status, confidence, control_id, limit	read, snippets redacted
pandaward_get_evidence	evidence_id	read, integrity checked
pandaward_get_graph	edge_id	read
pandaward_list_entrypoints	kind, service, limit	read: routes, GraphQL, gRPC, workers, cron with file and line
pandaward_scan_passive		local scan, writes the ledger
pandaward_retest	changed_only (default true)	targeted re-scan of what changed, writes the ledger
pandaward_build_report	formats	writes reports
pandaward_record_hypothesis	control_id, file, start_line, end_line, title, severity, rationale, fix, proposer	records a HYPOTHESIS finding
pandaward_prepare_feedback	kind, finding_id, rule, file, line, note, include_path, include_snippet	prepares a report and a prefilled issue link; sends nothing
pandaward_request_active_test		explains authorization; never executes

Arguments are validated strictly; unknown fields are rejected. Every response carries a notice that repository-derived fields are untrusted data. The tool names, inputs and annotations are frozen; see Schemas and compatibility.