README.md

Getting started: one decision-aware task

Lore helps developers and coding agents understand the decisions that matter before changing code. Start with a task, not an interactive lesson, browser view or generated wiki. The product scope explains the boundary; learning-engine migration explains removed flags and JSON responses.

Install

Build from the Lore repository using a stable Rust toolchain:

cargo install --path . --locked
lore --version                       # lore 0.10.0

Ensure Cargo’s binary directory (normally $HOME/.cargo/bin) is on PATH. With the locked dependencies already cached, add --offline for a network-free install. See the actual clean-source install and first-use receipt; it does not establish model availability or a published release.

First contact (no configuration, model or write)

Run inside a project with Markdown documentation:

cd /path/to/your-project
lore context "Add retry diagnostics without changing accepted policy"
lore --json context "Add retry diagnostics without changing accepted policy"
lore onboard                           # optional short project orientation

Ordinary unpinned context and onboard before initialization use the same bounded lore.bootstrap_context schema-1 documentary source selection. The output retains quoted source excerpts, paths, line spans, file SHA-256 and explicit omissions, with zero model calls and automatic writes. It does not inspect the checkout, prove implementation, certify policy authority or execute tests. First-contact discovery honors local exclusions, file/read limits and configured Markdown roots.

Read the original cited passage. Check whether it is accepted, proposed, rejected or a report. A missing passage can mean it was excluded or omitted, not that a constraint is absent.

Optional practice project

For a disposable, offline example, copy the HarborPay fixture, not the original Lore checkout:

cp -R examples/tutorial-project /tmp/lore-practice
cd /tmp/lore-practice
lore context "Can production charge retries increase to five?"

PowerShell from the Lore checkout:

Copy-Item -Recurse examples/tutorial-project "$env:TEMP\lore-practice"
Set-Location "$env:TEMP\lore-practice"
lore context "Can production charge retries increase to five?"

Use a new destination if the copy already exists. In the fixture, an accepted three-attempt rule and a proposed five-attempt RFC have different decision status. An incident report is not a runtime test. Nothing here requires an answer key, lesson assessor, learning profile or model download.

Compile durable knowledge (deliberate opt-in)

lore init --configure-only --source ./docs
# Inspect lore.yml, the selected model and privacy limits.
lore doctor --inference    # optional: makes synthetic provider calls
lore init                  # may perform many inference calls
lore status
lore context "Add retry diagnostics without violating approval"
lore decisions "retry policy"
lore cases "duplicate charge"

init builds the retained evidence registry; new projects now default to output.mode: registry_only and do not create wiki pages. If you need optional Markdown export, create a new config with lore init --configure-only --publish-markdown --source ./docs or choose output.mode: markdown. Existing wiki projects keep their output until explicitly migrated (guide); 0.9.0-rc.1 always published pages. Compiled context uses schema-5 intelligence by default or a labeled deterministic fallback; context "TASK" --fast is explicitly model-free schema-2 retrieval only after initialization. Exact retained quotations are available with lore evidence ID. OpenWiki snapshots are optional; source Markdown is enough to get started.

A call to lore update may invoke configured models; lore status is read-only. Keep .lore/ private and excluded from version control, because it may contain source evidence and model responses.

Give a coding agent the same context

lore integrate codex                 # preview instructions, no writes
lore integrate codex --apply         # explicit install into AGENTS.md
lore integrate claude-code           # preview CLAUDE.md instructions
# Then: lore integrate claude-code --apply

Integration gives no additional filesystem, network, model-cost or code-edit authority. The agent still owns changes and actual tests. Agent integration details.

Inspection, evidence and limits

lore context "TASK" --inspect grants read-only bounded checkout inspection for the invocation on an initialized project; --no-inspect prohibits it even when a host grant exists. Static inspection is not executed verification. Hosted models and checkout egress require separate explicit grants; project text never grants them. Use --json for exact schema and provenance, read omissions, and check material discrepancies yourself or with your coding agent.

If a command fails: uninitialized_project means a pinned/fast mode requires lore init; an absent model may produce a labeled deterministic fallback on initialized context; a budget error needs a larger --max-tokens or narrower task; and pending publication requires deliberate recovery. Operations and privacy, command map.

Not available anymore: onboard --mode, hints, exercises, answer assessment, lesson revisions, solution reveal and transfer activities. These arguments now fail explicitly. The previous lore.human_experience and lore.bootstrap_onboard JSON contracts were replaced by the existing context contracts; consult the migration guide. Archived project-onboarding designs and empirical protocols are historical records, not feature documentation.

For discovery use lore search "QUERY"; for recorded reasons and history use lore decisions "QUERY" or lore cases "QUERY"; follow returned citations with lore evidence ID. The former read/explore workflows are retired (migration).