Configuration¶

sminer runs with no configuration at all. When you do want to change something, every setting can come from a flag, an environment variable or the config file.

Precedence¶

The first layer that sets a value wins:

Command-line flags, such as --since 30d
Environment variables: SMINER_ plus the key in upper case, with dots as underscores. classifier.model is SMINER_CLASSIFIER_MODEL.
The config file: the file named by --config, or else the first of these that exists: $XDG_CONFIG_HOME/sminer/config.yaml (by default ~/.config/sminer/config.yaml), ~/.sminer.yaml, ./.sminer.yaml
Built-in defaults

sminer config show prints every effective value and which layer set it:

KEY               VALUE              SOURCE
window            30d                env SMINER_WINDOW
offline           true               flag --offline
classifier.model  claude-haiku-5-5   file /home/me/.config/sminer/config.yaml
min_count         3                  default

The config file¶

Every key is optional. config.example.yaml in the repository is a commented starting point:

roots: [~/.claude/projects]
window: 90d
min_count: 3
offline: false
exclude_repos: ["~/personal/*"]

classifier:
  model: claude-haiku-5-5
  max_calls_per_run: 500
  concurrency: 4

embeddings:
  provider: auto
  model: nomic-embed-text
  url: http://localhost:11434

cluster:
  threshold: 0

redact:
  patterns: ["(?i)aws_secret", "ghp_[A-Za-z0-9]+"]

remedies:
  hooks_dir: ~/src/claude-hooks
  global_claude_md: ~/.claude/CLAUDE.md
  skills_dir: ~/.claude/skills
  settings: ~/.claude/settings.json

Keys¶

Key	Default	Meaning
`roots`	`[~/.claude/projects]`	Directories searched for `*.jsonl` transcripts. Flag: `--root`
`window`	`90d`	How far back to ingest and analyze: `90d`, `2w`, `12h`. Flag: `--since`
`min_count`	`3`	Occurrences a finding needs to be reported. Flag: `--min-count`
`offline`	`false`	Never call the network. Flag: `--offline`
`exclude_repos`	`[]`	Globs of repos never ingested; a pattern also matches subdirectories
`classifier.model`	`claude-haiku-5-5`	Model for classification and proposal drafting
`classifier.max_calls_per_run`	`500`	Cap on API calls per run, across both uses
`classifier.concurrency`	`4`	Parallel classifier calls
`embeddings.provider`	`auto`	`auto` (ollama if it answers, else built-in), `ollama` or `hash`
`embeddings.model`	`nomic-embed-text`	ollama embedding model
`embeddings.url`	`http://localhost:11434`	ollama server
`cluster.threshold`	`0`	Cosine similarity to join a cluster; `0` uses the embedder's default (0.45 built-in, 0.78 ollama)
`redact.patterns`	`[]`	Extra regexes to redact, on top of the built-ins
`remedies.hooks_dir`	`~/src/claude-hooks`	Where hook proposals are written, and existing hooks are read
`remedies.global_claude_md`	`~/.claude/CLAUDE.md`	The CLAUDE.md for rules that span repos
`remedies.skills_dir`	`~/.claude/skills`	Where skill proposals go, and existing skills are read
`remedies.settings`	`~/.claude/settings.json`	Claude Code settings, read for existing hooks
`db`	`$XDG_DATA_HOME/sminer/sminer.db`	The database. Flag: `--db`
`patch_dir`	`$XDG_DATA_HOME/sminer/patches`	Where accepted patches are written
`json`	`false`	Print machine-readable JSON on stdout. Flag: `--json`
`quiet`	`false`	Print nothing on stderr except errors. Flag: `--quiet`
`verbose`	`false`	Print detail on stderr. Flag: `--verbose`
`no_color`	`false`	Disable color. Flag: `--no-color`

Paths may start with ~. Invalid values (a bad window, a min_count below 1, an unknown embeddings.provider, or quiet and verbose both set without either flag) are rejected up front with exit code 2.

Credentials¶

The classifier uses the Anthropic SDK's standard credential lookup: ANTHROPIC_API_KEY most commonly. sminer never takes a key as a command-line argument.

Locations¶

sminer follows the XDG base directory spec, on macOS too:

What	Where
Config file	`$XDG_CONFIG_HOME/sminer/config.yaml`, default `~/.config/sminer/config.yaml`; then `~/.sminer.yaml` and `./.sminer.yaml`
Database and patches	`$XDG_DATA_HOME/sminer/`, default `~/.local/share/sminer/`

Better clustering with ollama¶

The built-in embedder hashes words, so it groups corrections that share wording but can miss paraphrases. A local embedding model groups by meaning, and nothing leaves your machine either way:

ollama pull nomic-embed-text
ollama serve   # sminer picks it up automatically with provider: auto

Changing the embedder recomputes all embeddings on the next run. Finding IDs are matched by similarity, so most survive the switch.