hforge-recursive-investigate

recursive investigation orchestration using harness forge. use when the task is hard, ambiguous, cross-module, or likely to benefit from Typed RLM, bounded subcalls, durable iterations, and replayable recursive artifacts instead of chat-only reasoning.

HForge Recursive Investigate

Trigger Signals

  • the root cause is unclear and needs staged evidence gathering
  • the task crosses multiple modules, services, or ownership boundaries
  • the prompt is getting expensive and would benefit from compact root frames and durable artifacts
  • the work needs Typed RLM, bounded code cells, proposal artifacts, or replayable investigation state
  • a normal one-pass repo scan is likely to lose context or miss intermediate findings

Do not use this for straightforward single-file edits where ordinary repo inspection is enough.

Inspect First

  • AGENTS.md
  • .hforge/agent-manifest.json
  • .hforge/generated/agent-command-catalog.json
  • .hforge/runtime/index.json
  • .hforge/runtime/recursive/language-capabilities.json
  • .hforge/runtime/recursive/runtime-inventory.json
  • .hforge/runtime/recursive/escalation-heuristics.json
  • .hforge/runtime/recursive/sessions/
  • commands/hforge-recursive-investigate.md

Workflow

  1. confirm the workspace is initialized with hforge status --root . --json
  2. if runtime artifacts are missing, initialize first with hforge bootstrap --root . --yes
  3. inspect recursive support with hforge recursive capabilities --root . --json
  4. inspect host-runtime posture with hforge recursive runtimes --root . --json
  5. provision Python or PowerShell explicitly with hforge recursive provision-runtime <python|powershell> --root . --json only when a workspace-managed alias is helpful
  6. create a durable session with hforge recursive plan "<objective>" --task-id <taskId> --root . --json
  7. prefer hforge recursive execute with a typed bundle before using freeform structured-analysis snippets
  8. use hforge recursive run only for one bounded fallback step when Typed RLM is not the right fit
  9. inspect iterations, subcalls, cells, promotions, meta-ops, score, and replay to summarize the work from durable artifacts
  10. stay honest about support posture, especially for Cursor and OpenCode where recursive support is translated rather than native

Output Contract

A good recursive-investigate result should include:

  • the reason recursive mode was chosen
  • the session id and the posture discovered from recursive capabilities and runtime inventory
  • whether the agent used Typed RLM, bounded structured analysis, or both
  • the most relevant iteration, scorecard, replay, or proposal artifacts
  • the next recommended action based on the durable evidence