Recon Tool Integration

When to Use

• Adding a brand-new scanning/enrichment tool to the scan pipeline.

For adding AI decision-making to an existing tool, use recon-ai-enrichment. For the graph write, use graph-db-writes. For the settings, use project-settings-cascade. This skill is the module + pipeline wiring.

Critical Rules

• NEVER ship the enrichment module without its _isolated wrapper. run_<tool>_enrichment_isolated() is the actual call path for the GROUP 3b parallel fan-out AND for every unit test; the plain run_<tool>_enrichment() alone is never fan-out-safe. Reference: recon/main_recon_modules/censys_enrich.py:369.
• NEVER vary the top-level result key. The module writes combined_result["<tool>"] and the wrapper returns snapshot.get("<tool>", {}) with the same identifier used everywhere else in the pipeline (censys_enrich.py:365).
• NEVER collect a field and not write it to the graph. Every field in the output dict must land on a node/relationship or it is silent data loss - see graph-db-writes.
• NEVER add a tool setting without updating BOTH preset layers. Add it to the Zod recon-preset-schema.ts and to RECON_PARAMETER_CATALOG in webapp/src/app/api/presets/generate/route.ts. Miss the Zod schema and AI-generated presets silently strip the setting; miss the catalog and the preset LLM never knows the tool exists.
• ALWAYS prefix every print() log [symbol][ToolName] ([*] progress, [+] success, [-] skipped, [!] error). Recon stdout is tailed into the SSE recon drawer; a bare print() is invisibly formatted.
• ALWAYS place a fan-out tool behind the deep-copy _isolated wrapper and in the correct execution group in recon/main.py; never parallelize across a dependency boundary (a tool needing live URLs cannot run before GROUP 4).
• ALWAYS route a new external call through recon/helpers/circuit_breaker.py. An API endpoint goes through a Breaker (guarded_call, or a provider:endpoint breaker with record(classify_http(...))); its keys go through a KeyPool, not a raw main-key read. A loop over scan hosts (HTTP or a Docker tool) gates each host with host_health.allow(url) / scope.skip_if_down(url) and records the outcome with host_alive / host_failed — any HTTP response is life, only a connection failure counts. Never log a key or a response body: breaker messages carry the provider, endpoint and a refused key's 1-based rotation position only.
• ALWAYS declare what the tool cut to the coverage accumulator. Open a scope = circuit_breaker.scope((), label="Tool", unit="host(s)") at the tool's start and call scope.finish("<phase>", sources=[...], host_source="<graph source>", payload=result) at its end — so the end-of-run prune KEEPS a paused source's or a skipped host's prior findings instead of deleting them as "gone", and the run shows partial — N sources skipped. host_source must equal the source value the tool writes to the graph. A tool whose findings carry no host field passes host_field=False (the skip is reported at the source level).
• A new tool setting FAILS THE BUILD until it is in the registry. Every parameter is described once in recon_settings/registry.yaml, and a test walking Prisma.ProjectScalarFieldEnum fails until every column has an entry. Draft it with python3 tooling/scripts/extract_recon_registry.py, EDIT IT, then python3 recon_settings/build.py.
• Tuning is OPEN and controlled at the point of use, not by being refused. A rate is capped to the engagement ceiling at scan start; a wordlist path outside the project's own directories is dropped. A container image is a closed values: list and an out-of-set value is REFUSED at the write, because accepting one and replacing it at scan start made get_recon_settings echo an image the scan would never run.
• Two things stay closed, and only two. The engagement SCOPE (mcp: create_only, set once by create_project) and the engagement RECORD (mcp: never, deny_reason: engagement-record - the client, the contacts, the dates, the document). The engagement's LIMITS are ordinary mcp: settable fields in the engagement_limits group: they are reachable from the form and from MCP alike, and what makes them safe is that each is enforced at scan start whatever the setting says. See README.MCP.SERVER.md.
• A settable field with no form input fails parity.test.ts, and so does a form input with no classification. Neither door may reach something the other cannot. form_section is joined from the tool's own entry, so a field lands beside its tool automatically.
• A new rps field with traffic: active and no roe_capped: true fails the build. That gap is how three rate limits shipped reachable over MCP and outside the ceiling.