AI agent skill · by @samugit83
Add Partial Recon
Adding partial-recon support for a tool: running a single pipeline phase on demand from the workflow graph, reading its inputs from the existing Neo4j graph and merging results back. The input-node map decides the modal UI, and some input types must never be manually entered. Trigger: adding a tool to partial recon; a new recon/partial_recon_modules/*.py; editing SECTION_INPUT_MAP in nodeMapping.ts; wiring the partial-recon modal or a single-phase re-run.
Archived 2026-10-23 · Source updated 2026-09-17 · License: MIT. Check the original repository for the latest version.
SKILL.md instructions
## When to Use - Making an existing recon tool runnable as a single on-demand phase from the graph. For adding the tool to the full pipeline, use `recon-tool-integration`. For the graph write itself, use `graph-db-writes`. This skill is the partial-recon path: graph-sourced inputs, the modal, and the single-phase re-run. --- ## Critical Rules - **NEVER let partial-recon results create duplicate nodes.** Partial runs merge into the **existing** graph and must dedup via `MERGE` on the tenant triple - see `graph-db-writes`. Partial and full share the same graph. - **NEVER expose a graph-only input type as a manual textarea.** `SECTION_INPUT_MAP` in [nodeMapping.ts](../../webapp/src/components/projects/ProjectForm/nodeMapping.ts) lists what the tool reads from the graph, but only the types that make sense for manual entry get a user field. Types like `Port` and `Endpoint` come from the graph only and must not be hand-entered. - **ALWAYS wire the shared entry so it works in BOTH the full pipeline AND partial recon.** Partial recon is spawned as a **separate** container from the full pipeline (`recon_orchestrator/container_manager.py`); hook the shared entry function once so both inherit (see `recon-ai-enrichment` for the same rule). - **ALWAYS mirror the reference impl matching your tool's input shape**, not an arbitrary one: Naabu (`Subdomain` + `IP`), Masscan (`IP` only), Nmap (`IP` + `Port`), Katana (`URL`). The input-node shape drives the whole wiring. --- ## Input-node shape -> reference impl | Tool reads | User enters | Mirror | | --- | --- | --- | | Subdomain + IP | Subdomain + IP (two textareas) | Naabu | | IP only | IP (one textarea) | Masscan | | IP + Port | IP (Port from graph) | Nmap | | BaseURL | URL (maps to a BaseURL node) | Katana | ## Commands ```bash # recon/*.py is spawned fresh per job (volume-mounted) - no rebuild ./redamon.sh test unit # recon + root-recon sections ``` ## Resources - [docs/readmes/coding_agent_prompts/PROMPT.ADD_PARTIAL_RECON.md](../../docs/readmes/coding_agent_prompts/PROMPT.ADD_PARTIAL_RECON.md) - full walkthrough, modal UI, input validation, reference impls - Related skills: `recon-tool-integration`, `graph-db-writes`, `recon-ai-enrichment`