Menu

Search

Documentation Index

Theme: Symblix Light

Configuration

Last updated:

# Configuration

Audience: users and framework developers who configure a workspace. Purpose: the complete relurpify_cfg/ reference: every file, its keys, and the validation rules. Prerequisites: Getting started ; environment variables live there, not here.

# Where configuration lives

All workspace configuration lives in relurpify_cfg/ at the workspace root. The tree is generated from an embedded template bundle; doctor materializes it once on an empty workspace and never overwrites your edits (--fix is the explicit overwrite). Treat the tree like source code: it is what the runtime loads and relurplint validates.

For the Relurpify repository itself, the checked-in relurpify_cfg/ is a materialization of the same bundle, and a CI gate (make check-config-tree-drift) fails if the two diverge. Framework config changes go into the templates first, then regenerate.

# Validation rules

Every config file is validated strictly at load:

Rule Detail
Schema declaration The first non-comment line must be schema: <kind>/<version>, for example schema: relurpify/workspace/v1
Unknown fields Rejected; a typo is an error, not a silent default
YAML anchors and aliases Rejected outright
Secret-like fields Field names such as apikey, token, password, secret, privatekey, credential are banned in any config file
Registered schemas The declared schema kind/version must be registered in the loader

Secrets never live in config files. RELURPIFY_LLM_API_KEY is the only secret, and it is read from the environment (see the env var table ).

# The tree 

relurpify_cfg/
├── workspace.yaml                  # provider, model, sandbox, logging, postures, runner
├── model/
│   ├── provider/                   # one manifest per inference provider (4 ship)
│   └── profiles/                   # per-model behavior profiles (3 ship)
├── security/                       # sandbox, shell blacklist, tool policies, ingestion (4 files)
├── tools/                          # 99 tool manifests
├── prompts/                        # the prompt corpus; see PROMPT_DSL
└── euclo/                          # thoughtrecipes + families.yaml; see RECIPE_DSL

# workspace.yaml

Shipped defaults:

Key Default Description
schema relurpify/workspace/v1 Schema declaration (required first line)
paths.state_dir .relurpify_state Runtime state directory, relative to the workspace
model.provider ollama Active provider name; must match a provider manifest
model.name gemma4:e4b Active model; must be served by the provider
sandbox.backend gvisor gvisor or docker
selection.posture deterministic deterministic (ambiguous intents ask you) or model_assisted (tier-2 model may break ties)
logging.level info Log level
logging.format json Log format
audit.retention_days 7 Audit chain retention
audit.enforcement strict Audit enforcement posture
telemetry.enabled false Telemetry writing
runner.enabled true Background service runner
runner.workers 2 Runner worker count
runner.queues [knowledge] Queues the runner drains
runner.poll_interval 500ms Spool poll interval
runner.heartbeat 5s status.json heartbeat interval
runner.refresh_interval 0s Cadence for scheduled knowledge.refresh submission; 0s disables it

Strict-load interaction: with RELURPIFY_STRICT set, the five keys with built-in defaults (paths.state_dir, logging.level, logging.format, audit.retention_days, telemetry.enabled) must be present or loading fails; see Getting started .

# Provider manifests

One file per inference provider under model/provider/, schema relurpify/model/provider/v1.

Shipped file Kind Endpoint Models Native tool calling
ollama.provider.yaml ollama http://localhost:11434 gemma4:e4b, gemma4:12b, codellama:13b yes
lmstudio.provider.yaml lmstudio http://localhost:1234 discovered no
openai_compatible.provider.yaml openai_compatible http://localhost:8080/v1 yours yes
offline.provider.yaml offline none gemma4:e4b, offline-synthetic scripted

Implemented kinds: ollama, lmstudio, openai_compatible, offline, and tape (record/replay for tests; no manifest ships). Manifest fields include the request timeout, the available model list, the native-tool-calling flag, and a concurrency cap (max_concurrent).

# Model profiles

Glob-matched behavior per model under model/profiles/, schema relurpify/model/profile/v1. A default profile is mandatory; profiles tune tool-calling intent, context size, generation parameters, and quirks.

Shipped file Pattern Notable settings
default.llm.yaml * intent auto, 4 concurrent tools, 8192-token context, temperature 0.2
gemma4.llm.yaml gemma4* native tool calling, 2 concurrent tools, repair.strategy: llm (bounded repair of malformed decisions for thinking models)
qwen2.5-coder.llm.yaml qwen2.5-coder* native tool calling, 1 concurrent tool, double_encode_args: true (tool-args quirk), temperature 0.1

# Security policies

Four files under security/; missing pieces fail closed, they never open up.

File Schema What it controls
sandbox.policy.yaml relurpify/policy/sandbox/v1 Container constraints: protected_paths (shipped: relurpify_cfg), no_new_privileges, orphan reaping, and image_digest (pin it for a reproducible sandbox)
shell.policy.yaml relurpify/policy/shell/v1 Shell command blacklist; ships deny-git-reset-hard blocking git reset --hard
localtool.policy.yaml relurpify/policy/localtool/v1 Per-tool execute posture; ships bash: ask and cli_git: ask
workspaceingestion.policy.yaml relurpify/policy/ingestion/v1 Which files the knowledge layer may ingest; ships an allow rule for the workspace

Design rules the policies sit on top of:

  • The default tool posture is ask; permissive defaults are rejected by a dedicated lint gate.
  • Egress to private or loopback hosts is unconditionally denied.
  • .git and configuration roots are protected paths.
  • Effect/risk/trust class strings in capability blocks are normalized at the load boundary; make lint-class-normalization fails any code path that skips the normalization.

# Tool manifests

The 98 files under tools/ declare every model-callable tool: parameters, execution backend, flag expansion, sandbox limits, result chunking, exit-code error maps, rate limits, and a mandatory capability block declaring trust, risk, and effect classes.

Category Count Examples
file/ 6 file_read, file_write, file_list
git/ 5 cli_git (blame, branch, commit, diff, history)
lsp/ 7 lsp_get_diagnostics, lsp_get_definition
search/ 3 search_grep, search_semantic, search_find_similar
shell/ 73 build, text, network, system, archive, fileops tools plus bash
test/ 5 python_pytest, python_unittest, go_test, node_npm_test, rust_cargo_test

Tool names in recipes and policies resolve through the registry, never by literal string match; see RECIPE_DSL for the resolution order.

# Safe and dangerous changes

Safe:

  • Scalar values in workspace.yaml (provider, model, posture, logging).
  • New provider manifests and model profiles.
  • Security policy values within the fail-closed rules above.
  • New tool manifests, thoughtrecipes, and prompt files.
  • Anything under euclo/: your edits win over the templates.

Dangerous:

  • Hand-editing the generated tree inside the Relurpify repository itself (the drift gate fails; edit the templates and regenerate).
  • Renaming config keys (unknown fields are errors downstream).
  • Trying to set a permissive default tool posture (rejected by lint).
  • Declaring network egress to private or loopback hosts (denied regardless of policy).
  • Wrong-case class strings in capability blocks (normalized at load, but the lint gate flags unnormalized sites in code).

# Load semantics

  • Precedence: CLI flag > environment > workspace config > built-in default.
  • A missing workspace.yaml is fine: defaults apply.
  • A present but partial workspace.yaml under strict mode is a load error that names the defaulted keys.
  • Doctor reports provider health, profile match, and API-key presence as blocking diagnostics; the UI stays locked until they resolve.
  • After every load the runtime logs a fingerprint of the resolved, secret-free configuration.

# Where to go next

On this page