docs(prompts,skills): bound recon output for shared-disk hygiene

[devin-ai-integration]

Summary

Agents on a deep black-box scan can write very large, uncapped recon artifacts (e.g. katana JSONL crawl output) into the single shared /workspace disk, with no self-limiting. Nothing in the prompt or the katana skill currently nudges them to bound crawl output or clean it up, so a broad -jsl/-kf all crawl on a large site can grow until the sandbox disk saturates. This adds lightweight, best-effort guidance (no code/runtime changes) so agents keep recon output bounded and tidy.

Layered so the rule is seen regardless of which skills are loaded:

• agents/prompts/system_prompt.jinja — always-on DISK & SCRATCH HYGIENE note in the shared-workspace (AGENT ISOLATION & SANDBOXING) block: /workspace is shared/finite, prefer bounded recon over "collect everything", send large output to /workspace/scratch/ and remove raw output once distilled, and — when disk is tight — clean up only files from your own task (don't delete another agent's in-use artifacts on the shared disk). The PHASE 1 "CRAWL thoroughly" bullet also gains a bound-and-tidy reminder.
• skills/tooling/katana.md — document the flags that actually bound crawl output size (from projectdiscovery's katana docs) and a reduce-then-delete workflow:
  • -mdp (max-domain-pages — katana's default is unlimited), -fsu (collapse near-identical URLs), -fs fqdn (default rdn crawls all subdomains), -f url (plain URL list vs verbose JSONL), -or/-ob (omit raw/body), -mrs (cap per-response bytes).
  • Baseline command now ships -ct 10m -mdp 2000 -fsu; deep-JS example is time-bounded.
  • New "Keeping output small" section: bound scope/volume, shrink each record, reserve -jsl/-kf all for narrowed targets, then distil (-f url / short path list) and delete the raw crawl; sanity-check with du -sh.

Intentionally soft in tone (guidance, not hard mandates) and free of specific size thresholds; a model told to "go very deep" can still ignore it, so this is best paired with a hard inner-sandbox disk guard as the real backstop. This PR is the prompt/skill layer only.

Link to Devin session: https://app.devin.ai/sessions/dad023e379e942f287bcf6822463b7a4
Requested by: @0xallam

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring

[greptile-apps]

Greptile Summary

This PR adds guidance for keeping reconnaissance output bounded on the shared workspace. The main changes are:

• Adds shared-disk and scratch-file hygiene to the system prompt.
• Bounds crawls by target, depth, duration, and output volume.
• Documents Katana flags for smaller, scoped output.
• Recommends reducing crawl results before deleting raw artifacts.

Confidence Score: 5/5

This looks safe to merge.

• Cleanup guidance now protects files belonging to other active tasks.
• URL deduplication guidance is limited to an extracted URL list.
• No blocking issues remain in the updated documentation.

Important Files Changed

Filename	Overview
strix/agents/prompts/system_prompt.jinja	Adds bounded reconnaissance and ownership-aware cleanup guidance for the shared workspace.
strix/skills/tooling/katana.md	Adds scoped Katana examples and guidance for limiting, reducing, and cleaning crawl output.

_{Reviews (3): Last reviewed commit: "docs(skills): add research-backed katana..." | Re-trigger Greptile}

[devin-ai-integration]

@greptile

[devin-ai-integration]

@greptile