Configuration

Command-line flags configure a single run. Profiles and targets make those settings reusable and committable so sift dump does the right thing without flags — suitable for CI and automated dumping.

Quick start: [sift] defaults

Put repository-wide defaults in .sift.toml (project-local, committed) so any clone can run sift dump with no flags:

[sift]
style = "xml"
budget = 40000
mode = "full"
output = "codebase.md"
smart = true
extensions = ["go", "md", "ts"]
ignore = ["vendor/**", "dist/**"]
prompt = "Review for correctness."
# prompt_file = "prompts/review.md"   # file path instead of inline prompt

Flags still win: sift dump --budget 80000 overrides budget = 40000.

Profiles

Profiles are named presets you select with --profile (or via a target):

[profiles.review]
style = "markdown"
budget = 60000
mode = "signatures"
prompt = "Review this codebase for correctness and security issues."
extensions = ["go", "md"]

[profiles.claude]
style = "xml"
budget = 60000

[profiles.rust-strict]
extends = ["claude"]           # inherit, then override
extensions = ["rs"]
mode = "signatures"
prompt_file = "prompts/rust-review.md"
output = "rust-context.md"

sift dump . --profile review
sift dump . --profile rust-strict

extends composes profiles: ancestors are overlaid in order, then the profile’s own fields win. Cycles are an error.

Targets: multiple artifacts from one file

Each [[targets]] is one dump artifact. Select one with --target:

[[targets]]
name = "full"
profile = "claude"             # optional; inherits that profile first
output = "codebase.md"

[[targets]]
name = "sig"
profile = "rust-strict"
output = "codebase.sig.xml"
style = "xml"

[[targets]]
name = "docs"
extensions = ["md"]
prompt = "Summarize the architecture for a new teammate."
output = "docs-context.md"
budget = 12000

sift dump --target full        # codebase.md
sift dump --target sig         # codebase.sig.xml
sift dump --target docs        # docs-context.md (md-only, short budget)

Target fields (style, budget, mode, output, prompt/prompt_file, extensions, ignore, scoring) override the resolved profile and [sift] defaults. Flags override everything.

Prompts

Inline prompt or file-backed prompt_file on [sift], any profile, or any target:

[profiles.review]
prompt_file = "prompts/review.md"

[[targets]]
name = "api"
prompt_file = "prompts/api.md"

A reusable prompt library is also available:

[prompts.review]
text = "Review for correctness and security."

[prompts.docs]
file = "prompts/docs.md"

Selected with the resolver’s WithPromptRef("review") (programmatic) or via future --prompt-ref CLI flag.

Automation hints

Declarative hints for watch / hooks (opt-in; nothing runs from TOML alone):

[automation]
watch_interval = "500ms"
git_hook = "post-commit"       # `sift hook install` reads this
hook_targets = ["full", "sig"]

Precedence

Built-in defaults (Config.New())
  → global ~/.config/sift/config.toml
  → local .sift.toml [sift] defaults
  → selected profile (+ extends expansion)
  → selected target overrides
  → explicit --flags (pflag.Changed guard)

Local .sift.toml overlays global; sift dump /path resolves .sift.toml relative to the scanned root, not just cwd.

Legacy

default_profile = "NAME" in either config file, and bare [profiles.NAME] without [sift]/[[targets]], continue to work:

[profiles.claude]
style = "xml"
budget = 60000

extensions and ignore as TOML arrays (e.g. extensions = ["go", "md"]) are canonical; comma strings are a CLI concern (--ext go,md), not TOML.

Tuning selection ([scoring])

Any [sift], profile, or target may carry a scoring table. Unset keys keep built-in defaults; retention overrides per-role retention under budget (config priority over language default):

[profiles.min]
scoring = { retention = { docs = 0.0 } }        # don't guarantee docs under budget

[profiles.lean]
[profiles.lean.scoring]
recency_weight = 0.40
churn_weight   = 0.20
centrality_weight = 0.20
role_weight    = 1.0
full_band      = 0.55
skip_band      = 0.20
base_min       = 0.01
base_max       = 1.0
relevance_weight = 0.20
preference_bonus = 0.05
signature_bonus  = 0.05
skip_multiplier  = 0.35
sig_quality_min  = 0.35
sig_quality_max  = 0.85
retention        = { entrypoint = 0.30, docs = 0.22, config = 0.18, api = 0.12 }

Top-level [scoring] is also honored.

Settings reference

Output

Flag	Default	Purpose
--output	codebase.md	Output file; - writes to stdout
--style	markdown	plain, markdown, json, or xml
--mode	full	full content or signatures outline
--clipboard	off	Copy the rendered output to the system clipboard
--target	—	Target from .sift.toml (--target NAME)
--profile	—	Profile from config (--profile NAME)

Context and tokens

Flag	Default	Purpose
--budget	0	Maximum output tokens; 0 is unlimited
--tokenize-model	cl100k_base	BPE encoding used for counting
--prompt, -p	—	Task directive prepended to the output

Filtering

Flag	Default	Purpose
--ext	—	Only include these extensions (comma-separated)
--ignore	—	Extra gitignore-style patterns (comma-separated)
--max-size	0	Max file size in MB; 0 is unlimited
--hidden	true	Ignore hidden files; set false to include them
--git	true	Ignore Git-ignored files; set false to include them
--binary	off	Include binary files (skipped by default)
--smart	off	Skip generated/lock/minified/oversized files
--smart-max-tokens	15000	Per-file token ceiling for the smart filter

Safety

Flag	Default	Purpose
--secrets	true	Scan and redact detected credentials
--force-secrets	off	Include secrets instead of redacting

Terminal display

Flag	Default	Purpose
--highlight	true	Terminal syntax highlighting
--theme	auto	Highlight palette: auto, none, dark, light
--no-color	off	Disable color output
--window-title	—	Set the picker’s terminal title
--no-window-title	off	Disable terminal title updates
--no-nerd-fonts	off	Plain ASCII glyphs in the picker

Scan behavior

Flag	Default	Purpose
--dir	.	The root directory to scan
--concurrent	true	Concurrent file processing
--workers	CPU count	Number of concurrent workers
--timeout	—	Maximum execution time (e.g. 30s, 5m)
--show-skipped	off	List excluded files and reasons
--progress	off	Show progress information

Diagnostics

Flag	Default	Purpose
--verbose	off	Debug-level logging
--quiet	off	Only warn/error logging
--log-level	—	Set the logging level explicitly

Interactive theme vs. syntax theme

Two related but separate settings:

• --theme controls terminal syntax highlighting of rendered output (auto, none, dark, light).
• The picker’s interactive color palette (chosen with t in the TUI) is stored separately under the user’s application configuration directory and reloaded on the next picker session. Inside the picker, the active palette also drives preview syntax highlighting; --theme still governs non-interactive output.

Where state lives

Per-project dump state (used by delta) and interactive preferences are kept outside the scanned repository — by default under the application configuration directory (e.g. ~/.config/sift/). Nothing is written into your project, so the context document and repo stay clean.