Repository navigation
Expand file tree
/
Copy pathbcs.conf.sample
More file actions
112 lines (94 loc) · 4.35 KB
/
Copy pathbcs.conf.sample
File metadata and controls
112 lines (94 loc) · 4.35 KB
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
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
# SPDX-License-Identifier: GPL-3.0-or-later
# bcs.conf — BCS toolkit configuration
# Install to: ~/.config/bcs/bcs.conf
#
# This file is sourced as bash. CLI flags override these settings.
# See: bcs check --help
# See: tests/accuracy/LLM-ACCURACY.md for detailed backend accuracy data.
# Default model. May be a built-in alias (sonnet, opus, haiku, flash, pro,
# flash-lite, gpt5, gpt5-mini, qwen, qwen-small) or a canonical model ID
# (claude-sonnet-4-6, gemini-2.5-flash, gpt-5, qwen3.5:14b, ...). The backend is
# resolved from the alias-expanded name:
# claude-* -> Anthropic API
# gemini-* -> Google Gemini API
# gpt-* | o[0-9]* -> OpenAI API
# claude-code[:...] -> Claude Code CLI
# (anything else) -> local Ollama (e.g. minimax-m2:cloud)
#
# Accuracy-tested recommendations (by use case):
# Quick check: gpt5 (10-15s, 0 false positives)
# Daily dev: sonnet (28-87s, best accuracy/speed)
# Pre-release: claude-code:opus (2-24 min, highest accuracy)
BCS_MODEL=sonnet
# Default analysis depth: low (or min), medium, high, xhigh, max
# Each level sets max output tokens AND (on capable models) the thinking
# budget or reasoning_effort. See bcs check --help for the table.
# 'medium' is the recommended default for all backends.
# 'high' benefits Claude/OpenAI reasoning models; marginal elsewhere.
# 'max' should be avoided for Ollama cloud models (hallucination risk).
BCS_EFFORT=medium
# Treat warnings as violations: 0 or 1
BCS_STRICT=0
# Prepend `shellcheck --format=json -x` output to the LLM prompt as
# static-analysis context. Auto-skipped if shellcheck is not in PATH.
# Override per-call with --shellcheck / --no-shellcheck.
BCS_SHELLCHECK=1
# Result cache (0 or 1; default 1). Successful check results are cached
# under ${XDG_CACHE_HOME:-~/.cache}/bcs/ keyed on script content, standard
# content, model, effort, and filters -- re-checking unchanged inputs is
# instant and free. Override per-call with --cache / --no-cache.
#BCS_CACHE=1
# Default tier filters (override per-call with --tier / --min-tier).
# Values: core | recommended | style. Leave unset for no filtering.
#BCS_TIER=
#BCS_MIN_TIER=
# Default --debug (0 or 1); announces raw-response dump path.
#BCS_DEBUG=0
# Default --json (0 or 1); emit structured JSON on stdout.
#BCS_JSON=0
# Override or extend the model alias map (built-in aliases listed above).
# Add your own (or override defaults) with: MODEL_ALIASES[name]=canonical-id
#MODEL_ALIASES[sonnet]=claude-sonnet-5
#MODEL_ALIASES[mycloud]=minimax-m2:cloud
#MODEL_ALIASES[reviewer]=qwen3-coder:480b-cloud
# Override effort -> API parameters. Defaults shown.
# Anthropic constraint: EFFORT_THINKING[X] MUST be < EFFORT_TOKENS[X] - 1024.
#EFFORT_TOKENS[low]=4000
#EFFORT_TOKENS[medium]=8000
#EFFORT_TOKENS[high]=24000
#EFFORT_TOKENS[xhigh]=40000
#EFFORT_TOKENS[max]=64000
#EFFORT_THINKING[low]=0
#EFFORT_THINKING[medium]=2000
#EFFORT_THINKING[high]=6000
#EFFORT_THINKING[xhigh]=12000
#EFFORT_THINKING[max]=16000
#EFFORT_REASONING[low]=minimal
#EFFORT_REASONING[medium]=low
#EFFORT_REASONING[high]=medium
#EFFORT_REASONING[xhigh]=high
#EFFORT_REASONING[max]=high
# Ollama Cloud models (run `ollama signin` once, then `ollama pull <model>`):
# ▲ Accuracy testing shows Ollama cloud models (minimax, glm, qwen) have
# very low recall (0-0.5/8 on complex scripts) and non-deterministic
# suppression handling. Use for offline/private checking only.
# Prefer gpt5 (OpenAI) or sonnet (Anthropic) for accuracy.
# Available cloud models: minimax-m2:cloud, minimax-m2.7:cloud,
# qwen3-coder:480b-cloud, glm-5.1:cloud, deepseek-v3.1:671b-cloud
# Ollama server address
#OLLAMA_HOST=localhost:11434
# Ollama context window in tokens (sent as options.num_ctx). Without it the
# server truncates the prompt to its 4096-token default and the model never
# sees the standard. It must hold the standard (~33k tokens) plus the script
# plus the reply; raise it for long scripts or -e high and above (costs VRAM).
#BCS_OLLAMA_NUM_CTX=40960
# How long Ollama keeps the model loaded after a check: a duration (30m, 1h)
# or seconds as an integer (-1 = forever, 0 = unload at once).
#BCS_OLLAMA_KEEP_ALIVE=30m
# API keys (set the one for the backend your -m/MODEL_ALIASES selects; the
# backend is resolved from the model name, not by probing keys in order)
#ANTHROPIC_API_KEY=
#GOOGLE_API_KEY=
#GEMINI_API_KEY=
#OPENAI_API_KEY=
#fin