Context Brief

Context Brief is Lerim's fast startup context for workflow-scoped agent work.

It is a generated Markdown view of durable context records. It is not a second memory store, and agents should not edit it by hand. The source of truth remains the durable context store.

In the current CLI, workflow scope is resolved through project registration. For customer pilots, the same idea can map to clients, engagements, research topics, teams, incidents, accounts, or other business scopes.

lerim context-brief show prepends live DB freshness before printing the current markdown snapshot. The preface is current at read time; the markdown is the last generated artifact.

The current view lives at:

~/.lerim/workspace/current/<project_id>/CONTEXT_BRIEF.md

Agents do not need to know the internal scope id ahead of time. They should run lerim context-brief show, status, or path from inside the current workspace, or pass --project <name-or-path>.

Flow

flowchart TD
    A["Agent starts project work"] --> B["lerim context-brief show"]
    B --> C{"Current artifact exists?"}
    C -- "yes" --> D["Read live freshness preface and CONTEXT_BRIEF.md"]
    C -- "no" --> E["Use lerim context-brief status"]
    E --> F["Suggested action: refresh"]
    D --> G{"Need deeper or newer context?"}
    G -- "yes" --> H["lerim context-brief status"]
    H --> I{"DB records changed since generation?"}
    I -- "yes" --> J["lerim context-brief refresh"]
    I -- "no" --> K["Use lerim query or lerim answer for deeper lookup"]
    G -- "no" --> L["Proceed with project work"]
    J --> M["Runtime generates dated artifacts"]
    M --> N["Copy latest markdown and manifest to workspace/current"]
    N --> D

Generation Architecture

flowchart TD
    A["Trigger: manual refresh, daily daemon, or curate changed records"] --> B["Resolve project from cwd, name, or path"]
    B --> C["Resolve scope id from workspace path"]
    C --> D["Read current manifest from workspace/current"]
    D --> E["Count record_versions changed after previous generated_at"]
    E --> F{"Current file exists and changed count is 0 and not --force?"}
    F -- "yes" --> G["Return skipped result and record service run"]
    F -- "no" --> H["Create dated run folder under workspace/YYYY/MM/DD/context-brief"]
    H --> I["Load active candidate records from context store"]
    I --> J{"Any candidates?"}
    J -- "no" --> K["Build empty Context Brief draft without model call"]
    J -- "yes" --> L["Context Brief synthesis agent"]
    L --> M["Compile cited startup context"]
    M --> N["Structured output: summary sections and cited lines"]
    N --> O["Validate every line cites known record IDs and fixed sections"]
    K --> P["Render Markdown"]
    O --> P
    P --> Q["Write run artifacts: CONTEXT_BRIEF.md, manifest, events, agent log, trace"]
    Q --> R["Copy CONTEXT_BRIEF.md and manifest into workspace/current/<project_id>"]
    R --> S["Return generated result and record service run"]

Agent Boundary

The Context Brief feature has two layers:

• lerim.context_brief owns deterministic use-case logic: project resolution, changed-record detection, candidate loading, validation, rendering, manifests, status, artifact paths, deterministic Start Here, fixed section order, and live freshness prefaces.
• lerim.agents.context_brief owns model synthesis only. It receives bounded candidate records and returns structured cited lines for the model-filled sections.

LerimRuntime.context_brief() ties those layers together. The daemon calls the runtime for all registered projects during the daily pass, and after curate only when curate changed records.

Refresh Rules

Context Brief refresh is intentionally not part of the ingest hot path.

• lerim context-brief show, status, and path are fast local reads.
• show prepends live DB freshness before printing the static markdown snapshot.
• lerim context-brief refresh generates only when records changed, unless --force is passed.
• The daemon runs a daily pass across registered projects and skips unchanged projects.
• curate triggers Context Brief only when it creates, updates, archives, or otherwise changed records.
• Empty projects get an empty-state Markdown file without a model call.

Fixed Markdown Shape

Generated Context Brief has a stable section order so agents can scan it predictably:

1. Summary
2. Start Here
3. Current Handoff
4. Decisions
5. Constraints & Preferences
6. Project Facts
7. Open Risks / Review Queue
8. Follow-up Queries
9. Sources

Summary is the compact startup cache. Start Here is deterministic and rendered by Lerim from project metadata and artifact status. Current Handoff must come only from recent episode evidence; without that evidence, it should explicitly say no persisted implementation handoff is available and point agents back to the current chat, workspace state, and relevant checks for live work.

Decisions, Constraints & Preferences, and Project Facts hold durable records. Open Risks / Review Queue and Follow-up Queries are populated only from records that explicitly describe unresolved work, review concerns, or questions worth asking next. Sources lists the cited record IDs used in the body. Any test/build result in these sections is historical persisted evidence; agents must rerun relevant checks after making edits.

Artifact Layout

Each generation writes a dated run folder:

~/.lerim/workspace/YYYY/MM/DD/context-brief/context-brief-<timestamp>-<id>/
  CONTEXT_BRIEF.md
  manifest.json
  events.jsonl
  agent.log
  agent-trace.json

The latest successful run is copied to the stable current path:

~/.lerim/workspace/current/<project_id>/
  CONTEXT_BRIEF.md
  manifest.json

The manifest records the project, project_id, run folder, generated time, previous generation time, trigger, candidate count, included record IDs, and changed-record count before generation.

What Agents Should Do

At startup, an agent working in a workspace should:

1. Run lerim context-brief show from the workspace.
2. If the file is missing or the task depends on newest context, run lerim context-brief status.
3. If status reports changed DB records, suggest or run lerim context-brief refresh.
4. Use lerim query for exact inspection and lerim answer for synthesized answers across more context.