pi-workbench/offload-rules.example.json at main · Stinocon/pi-workbench · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
{
  "rules": [
    {
      "id": "adversarial-review",
      "type": "block",
      "harnessRoots": [
        "/absolute/path/to/your-mirror-repo"
      ],
      "criticalPaths": [
        "extensions/offload-rules.ts",
        "extensions/verify-gates.ts",
        "extensions/declarative-hooks.ts",
        "extensions/web.ts",
        "extensions/delegate.ts",
        "extensions/anon-guard.ts",
        "extensions/shell-guard.ts",
        "extensions/plan-mode/**",
        "extensions/subagent/**",
        "hooks.json",
        "offload-rules.json",
        "shell-guard.json",
        "decisions/*",
        "git-hooks/*",
        "scripts/check-*.cjs",
        "scripts/test-*.mjs",
        "scripts/test-*.sh",
        "scripts/install-*.sh",
        "anon.py",
        "deanon.py",
        "skills/anon/**",
        "skills/memory/**",
        "skills/security-review/**",
        "scripts/**",
        "extensions/**"
      ],
      "criticalPathsAnywhere": [
        "AGENTS.md",
        "mcp.json",
        ".pi/verify.json",
        ".pi/decisions/*"
      ],
      "exemptPaths": [
        "*.example.json",
        "README.md",
        "docs/**"
      ],
      "reviewAgents": [
        "miner-medium",
        "miner-high",
        "miner-xhigh",
        "miner-max"
      ],
      "maxReminders": 1,
      "message": "Adversarial review required: a guarded surface was modified and no review by a different model has covered it. Run ONE round as your LAST act, after your final write to a guarded surface. The round discharges ONLY the paths its task names: name ALL guarded paths of this change-set, or none. Brief: the DIFF (not the whole file), the invariants it must not break, a cap on findings with a severity each. Discovery on miner-high; verification of the fixes on miner-medium; miner-xhigh or miner-max only when the surface is a security one (guards, gates, credentials, HA/MCP writes); decision records and other doc-level diffs on miner-medium. The task parameter is mandatory. Fixes are then verified by the deterministic gates and the tests they add — not by another review round.",
      "discoveryAgent": "miner-high"
    }
  ],
  "directives": [
    "Offload calibration: do not offload work to the cheap tier merely because the task involves tool calls. The cheap model is appropriate only for low-risk mechanical slices: simple repo search, file reading, information extraction, mechanical reformatting, simple verification, unambiguous local transformations, tool-free workers. Architectural changes, non-trivial debugging, requirement interpretation, tasks governed by local AGENTS.md files, implementation decisions, complex refactors, multi-step tasks, and work that must understand conventions distributed across the repository stay on the primary.",
    "Prose-to-structure offload (mandatory): converting a given document, notes, transcript, or list into a structured table / summary / extraction is a mechanical slice and always eligible for the cheap tier; when the request is such a conversion, offload it to a bounded worker — delegate (default target: your configured draft worker) or a miner-medium subagent — and review the output cell-by-cell — do not perform the conversion inline. Deterministic extraction/enumeration/sort is never an LLM task: grep/awk/python only."
  ]
}