CHANGELOG.md

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[0.3.31] - 2026-06-21

Added

• Added bundled research and compliance signal profiles, registered as first-class source profiles alongside coding, support, ops, and generic.
• Added realistic multi-turn example traces for support, incident, research, and compliance workflows, replacing the earlier two-line toy traces.
• Added before/after demo guides for support ops, incident ops, research, and compliance, each showing real extracted records captured through extraction.
• Added a custom and non-coding agents front-door guide covering the profile, clean-trace, import, query, and improvement-loop path.
• Added a defense-in-depth redaction helper (scripts/redact_trace.py) that strips common secrets and PII from a trace before import.
• Added an agent improvement loop diagram to the README and custom-agents guide.

Changed

• Updated the README with a Custom and Non-Coding Agents section, the bundled profile table, and links to the vertical demos.
• Updated the docs navigation, examples index, and custom-trace submission guide to cover the new profiles, demos, and redaction helper.
• Updated the custom-trace-folder and context-retrieval diagrams to use neutral support and research agent examples instead of removed third-party logos.

[0.3.30] - 2026-06-07

Fixed

• Fixed dashboard operation detail logs so they honor the selected project instead of falling back to shared logs.
• Fixed project-scoped operations activity by paging through the session catalog until matching project runs are found.
• Hardened source-session catalog migrations and FTS rebuilds so older or malformed catalogs recover into a healthy indexed state for graph and source views.

[0.3.29] - 2026-06-06

Fixed

• Fixed dashboard graph project switching by enforcing selected-project scope on graph responses and clearing stale graph data during project reloads.
• Made dashboard record, graph, artifact, operations, and source views fail closed instead of preserving stale rows after scoped load failures.
• Fixed the Records tab "All statuses" filter so it returns active and archived project records instead of falling back to the active-only default.
• Hardened the dashboard launcher against mismatched backend versions so an old local API cannot silently serve stale project data to a newer dashboard.
• Made invalid project selections return clear errors for status, memory, and Clinic endpoints, and kept refine reports degraded instead of crashing when the session catalog is unavailable.

[0.3.28] - 2026-06-06

Fixed

• Updated the offline release integration expectation for project-list count fields so the release workflow validates the new dashboard project counts.

[0.3.27] - 2026-06-06

Changed

• Made dashboard project selection visibly affect graph, records, brief, Clinic, source, operations, overview, and insights views, while keeping shared runtime surfaces clearly labeled.
• Separated current active, archived, and total persisted record counts across project lists, status payloads, and Run Clinic.

Fixed

• Fixed plain missing project names being interpreted as relative paths, so invalid dashboard project selections now return a clear error instead of falling back to the current workspace.
• Scoped graph queries, record filters, run detail reads, and artifact history through the selected project, including child workspace paths.
• Kept lerim serve and dashboard session-backed endpoints available in a degraded state when the source-session catalog is unhealthy.

[0.3.26] - 2026-06-06

Added

• Added project selection to the dashboard surfaces that read project-scoped data, including overview, analytics, context, graph, traces, memory artifacts, operations, and Run Clinic.
• Added the Run Clinic dashboard artifact surface and project-scoped generated artifact history views.

Changed

• Kept selected projects in dashboard navigation URLs while leaving shared Skills and Settings views unscoped.
• Labeled shared runtime and multi-project operation rows clearly when they appear inside a project-scoped dashboard view.

Fixed

• Scoped dashboard status, queue, source-session, graph, artifact-history, and record-open reads to the selected project, including child workspace paths.
• Prevented cross-project artifact history and mixed-project operation totals from leaking into project-scoped dashboard views.

[0.3.25] - 2026-06-02

Added

• Added README and docs guidance for registering skill targets, reviewing skill update proposals, and using opt-in auto-apply.

Changed

• Refined the Skills dashboard proposal review into a desktop queue/detail layout with explicit risk labels and cleaner applied-proposal state.
• Reduced normal dashboard tab gutters so pages use the available workspace with compact responsive padding.

[0.3.24] - 2026-05-31

Added

• Added skill stewardship registration, refresh, proposal review, diff preview, and apply/reject flows across the CLI, API, and dashboard.
• Added scoped record loading, artifact-surface validation, and opt-in auto-apply policy controls for registered instruction targets.

Fixed

• Hardened skill proposal application so stale concurrent apply attempts cannot revert files after a proposal has already been marked applied.

[0.3.23] - 2026-05-30

Fixed

• Removed stale release and docs workflow references to the deleted support-boundary SVG generator.

[0.3.22] - 2026-05-30

Changed

• Refactored context-answering, context-brief, trace-ingestion, and working-memory agents onto DSPy pipelines.
• Simplified Working Memory handoff output so startup context stays focused on useful continuation context instead of task-list noise.

Fixed

• Restored the release metadata path after the failed post-0.3.11 tags by publishing the current release candidate under a fresh version.

[0.3.18] - 2026-05-28

Changed

• Added a persistent node type legend to the dashboard Context Graph canvas so record-kind colors are readable without opening the inspector panel.

[0.3.17] - 2026-05-28

Changed

• Replaced the dashboard Context Graph renderer with AntV G6, keeping the graph focused on topic clustering and clustering-off modes.
• Improved the clustering-off graph with record-kind node colors, a compact topology layout, and a record type legend.

Fixed

• Returned durable record kinds in graph API node payloads so dashboard node colors reflect the real data catalog instead of the derived context-node type.

[0.3.16] - 2026-05-28

Added

• Added the Working Memory dashboard surface, API/runtime support, CLI and concept documentation, and unit coverage for working-memory behavior.

Changed

• Improved the Context Graph topic visualization with real node-bounded cluster clouds, cleaner semantic cluster labels, draggable cluster handles, overlap guarding, and less noisy focused-edge rendering.
• Updated dashboard navigation, context brief surfaces, and supporting docs for the working-memory flow.

[0.3.15] - 2026-05-28

Added

• Added an Insights memory timeline with created, revised, archived, and active context signals over the selected time window.
• Added clickable changed-memory drilldown rows with a detail modal for memory body, typed fields, provenance, and version metadata.

Changed

• Grouped Insights charts into separate Memory Timeline and Source Activity sections while keeping one shared page-level time window selector.
• Hid empty model and tool diagnostic charts until indexed trace metadata exists, while keeping the readiness note visible.

[0.3.14] - 2026-05-27

Added

• Made lerim dashboard a local dashboard launcher that ensures the backend is reachable, installs dashboard npm dependencies when needed, starts the Next.js UI, and prints the dashboard URL.
• Added lerim dashboard --port for choosing the local dashboard UI port.

[0.3.13] - 2026-05-27

Added

• Added the open-source dashboard to the main Lerim repository.

Fixed

• Served the context graph dashboard from learned graph data, including real context_nodes and context_edges, instead of the old record-only fallback.

[0.3.12] - 2026-05-21

Added

• Added a public custom trace folder flow asset to the custom trace guide.

Fixed

• Updated the release workflow, Docker runtime pull target, OCI source label, PyPI metadata, README links, and docs repo links after the GitHub repository rename from lerim-cli to lerim.
• Changed GHCR publishing and lerim up defaults to ghcr.io/lerim-dev/lerim after the old CLI package name was retired.
• Clarified that MCP config support provides context retrieval and explicit lerim_trace_submit, not automatic completed-session capture, and documented the OpenClaw/custom clean trace boundaries.

[0.3.11] - 2026-05-21

Fixed

• Made lerim answer report real internal answerer failures as HTTP 500 while reserving HTTP 504 for the configured five-minute answer deadline.
• Aligned CLI, server, and BAML model-call timeout contracts so complicated answer queries can run up to the five-minute budget without being cut off by a shorter layer.
• Added semantic retries for invalid context-answer retrieval plans and placeholder-only final answers, preventing bad model outputs from becoming successful CLI responses.

[0.3.1] - 2026-05-20

Changed

• Generalized imported market-baseline benchmark artifacts and docs so public benchmark structure is not organized around one competitor-specific name.
• Kept release benchmark pages on the clean public artifact numbers while preserving source/provenance boundaries for imported competitor rows.

Fixed

• Hardened coding trace ingestion against one-way implementation runs that should be archived without durable records.
• Improved coding extraction retention for user strategy, project identity, and supported coding-context records before persistence.

[0.3.0] - 2026-05-20

Added

• Added a Python-first MCP server with context brief, context answer, context search, deterministic records listing, trace submission, and ingest/status tools.
• Added lerim connect --mode mcp config writers for Codex CLI, Claude Code, Cursor, OpenCode, Gemini CLI, Cline, Claude Desktop, OpenClaw, Hermes, Goose, Roo Code, Kilo Code, Windsurf, and OpenHuman, with dry-run, backup, and verification support.
• Added a native pi JSONL session adapter for ~/.pi/agent/sessions/.
• Added generic trace submission and retry plumbing for JSONL, JSON arrays, message wrappers, and plain-text transcripts through lerim trace import and MCP lerim_trace_submit.
• Added public benchmark scripts, raw artifacts, generated reports, and docs for LongMemEval-S retrieval, context budget, retrieval latency, MCP integration, trace-ingestion cost/performance, and aggregate extraction diagnostics.
• Added public integration docs, source-session context compiler docs, MCP versus native-adapter boundary docs, custom trace examples, and a commercial boundary document.

Changed

• Moved the core project metadata, license text, and contribution guidance to Apache-2.0.
• Reworked retrieval indexing to store compact public records plus hidden source-session index text, then rank with weighted reciprocal-rank fusion over semantic and lexical signals.
• Updated README and docs to position Lerim as a source-session context compiler for agent workflows, not only coding-agent memory.
• Updated MCP client config generation to use an absolute Python executable with -m lerim.mcp_server, avoiding client startup failures when lerim is not on the MCP client's PATH.
• Split benchmark documentation into Lerim-only results and market-wide comparison pages, with source/provenance boundaries for competitor rows.

Fixed

• Hardened generic trace loading so empty trace files fail clearly instead of flowing into extraction.
• Preserved validated supporting record IDs in lerim answer output so CLI and MCP users get explicit Sources: rec_... evidence pointers.
• Corrected support wording so MCP config support is not presented as native completed-session capture.
• Removed stale public launch visuals and regenerated current public/private release assets from checked source assets and transcripts.
• Hardened release workflows with strict docs build, clean/tracked public benchmark validation, PEP 517 build, twine check, clean install checks, and MCP stdio probes before publishing.

[0.2.1] - 2026-05-16

Added

• Added custom trace-folder projects with lerim project add <path> --type custom.
• Added direct custom JSONL session discovery with agent_type=custom, without platform adapters or compaction.
• Added custom-agent integration docs with a pasteable cleaner prompt for generating customer-owned trace cleaning scripts.
• Added per-agent architecture docs with Mermaid flowcharts generated from the compiled LangGraph graphs.
• Added integration coverage that registers a custom project, writes synthetic traces, runs ingest, and verifies custom sessions/jobs use the original clean trace files.

Changed

• Updated README, docs, bundled skill text, and configuration reference around the broader context-compiler positioning and custom-agent trace flow.
• Extended project config/API/CLI payloads with supported and custom source types while preserving existing project defaults.

[0.2.0] - 2026-05-16

Added

• Expanded Lerim from a coding-agent memory layer into a general trace-to-context architecture for AI agent workflows.
• Added BAML/LangGraph context curation, context answering, and context-brief compilation alongside trace ingestion.
• Added layered durable-signal filtering, source-session review, synthesized record updates, and a final context quality gate.
• Added integration coverage for each BAML-backed agent role with real LLM calls.

Changed

• Renamed the live agent roles around the new context architecture: trace ingestion, durable-signal filtering, context writing, context curation, context answering, and context-brief compilation.
• Refreshed prompts to avoid fixture-specific or coding-only assumptions and to classify durable signal from broader agent activity traces.
• Updated README, docs, bundled skill text, and landing-page positioning around agent traces, durable signal, and future-agent context.

Fixed

• Hardened SQLite migrations for foreign-key-safe table rebuilds.
• Tightened answerer retrieval so topical words are not misread as record-kind filters.
• Improved context-brief validation by dropping unsupported generated lines before rendering.

Removed

• Removed the remaining PydanticAI agent runtime and retired fallback/dead configuration paths.

[0.1.83] - 2026-05-14

Added

• Packaged the BAML source and generated client under src/lerim/agents/ so future agents can share the same BAML/LangGraph layout.
• Added the production BAML/LangGraph extract package with deterministic trace windowing, typed BAML scans, record synthesis, context-store persistence, and structured graph events.

Changed

• Replaced sync extraction with the BAML/LangGraph harness while keeping maintain, ask, and working-memory on PydanticAI.
• Updated extraction evals, integration tests, docs, and run artifacts to use graph events instead of PydanticAI extract messages.
• Tuned extraction prompts to avoid storing incidental personal names unless identity itself is the durable context.

Fixed

• Hardened session catalog/API status paths so catalog storage issues degrade status responses instead of crashing status or maintain.
• Made extraction persistence idempotent when a rebuilt session catalog replays a session whose episode already exists.
• Improved long-running extraction queue handling so transient SQLite heartbeat write failures and sequential processing do not create false stale-running jobs.

Removed

• Removed the legacy PydanticAI extract agent, extract-only trace tools, history processors, and the experimental baml_agents/ sidecar.

[0.1.81] - 2026-04-29

Fixed

• Unit docs-contract tests no longer require a repo-local AGENTS.md; the authoritative agent tool contract check now only reads committed public docs.

[0.1.80] - 2026-04-29

Added

• Local-first context runtime built around the canonical SQLite context store, project-scoped records, version history, FTS, and local embedding-backed retrieval.
• Semantic agent toolsets for sync, maintain, and ask flows, including context listing, search, fetching, writing, revising, archiving, superseding, counting, trace-note, and pruning tools.
• End-to-end MLflow observability for sync, maintain, and ask runs using Lerim-owned root/tool/event spans that continue through controlled PydanticAI retries.
• Expanded runtime artifacts, queue/status metadata, Docker runtime helpers, and CLI/API coverage for context operations.
• Larger unit, smoke, integration, and end-to-end test suites for context extraction, maintenance, retrieval, queueing, cloud sync, scope handling, and CLI behavior.

Changed

• Reworked memory terminology and docs around durable context records instead of legacy per-project markdown memory files.
• Updated provider, tracing, configuration, logging, and run-artifact behavior for the DB-backed context architecture.
• Refreshed README and documentation for setup, commands, storage layout, context model, tracing, and operational workflows.

Fixed

• Release-readiness cleanup for provider fallback parsing, strict TOML string/path validation, SPDX license metadata, and docs accuracy around sync ordering, tracing, semantic search config, and query sessions.
• MLflow traces now preserve a successful Lerim root run even when a tool attempt raises a controlled retry before later recovery.

[0.1.72] - 2026-04-13

Fixed

• CI unit-tests pipeline now passes again after restoring missing test compatibility helper build_test_ctx in lerim.agents.tools.
• Addressed Ruff failures in unit tests (unused imports and ambiguous loop variable names), so release branches no longer fail lint before tests run.
• Kept queue project filter fallback behavior while removing unused exception binding in API path resolution.

[0.1.71] - 2026-04-13

Added

• lerim status --live now uses the same data and renderer as snapshot mode, with periodic refresh only as the difference.
• New status payload fields for richer operations visibility:
  • projects[] per-project memory/queue/blocker summary
  • recent_activity[] timeline including sync and maintain
  • unscoped_sessions totals by agent
• New lerim unscoped command to inspect indexed sessions that do not map to a registered project.
• Queue filters now support exact project matching (--project) and explicit substring matching (--project-like).

Changed

• Status UI redesigned for clarity:
  • project stream table (blocked / running / queued / healthy)
  • explicit “What These Terms Mean” section
  • actionable “What To Do Next” section with full lerim ... commands
  • activity panel (sync + maintain)
• Read/query defaults now use all registered projects unless explicitly narrowed:
  • lerim status --scope all|project --project ...
  • lerim ask --scope all|project --project ...
  • lerim query records list --scope all|project --project ...
• Canonical run telemetry is now written in service_runs.details_json with normalized keys (metrics_version=1, sync/maintain totals, per-project metrics, events) while preserving legacy compatibility fields.

Fixed

• Live status activity no longer appears stale during long in-flight sync runs; running queue jobs are now surfaced in recent_activity.
• Fixed maintain runtime error (name 'index_path' is not defined) that caused maintain runs to fail.

[0.1.70] - 2026-03-28

Quality Improvements

• +41% composite quality score via Layer 1 AutoResearch optimization
• ChainOfThought for the prior extraction pipeline (biggest single improvement)
• Explicit dedup classification thresholds (0.7/0.4) in sync prompt
• Improved MemoryCandidate schema field descriptions for better output consistency
• Tighter post-extraction body filter (30→50 chars minimum)

Evaluation Infrastructure

• 4 new eval runners: dedup accuracy, maintain quality, search relevance (NDCG@5), tool selection
• LerimBench 7-dimension composite scoring with configurable weights
• Fuzzy title matching for dedup accuracy (substring + Jaccard similarity)
• Golden dataset support via --golden-dir flag
• Deterministic extraction and summarization assertion checkers

Dashboard

• Local bundled dashboard removed — web UI moving to https://lerim.dev
• lerim dashboard shows transition message with CLI alternatives
• API server remains for Docker container health checks

Cleanup

• Removed stale Codex tool references from ask prompt
• Cleaned up stale OAI SDK / ResponsesProxy references in internal docs

[0.1.69] - 2026-03-25

Breaking

• Removed PydanticAI dependency -- all agent operations now use a ReAct runtime.
• Removed explorer subagent — replaced by Codex filesystem sub-agent.
• Removed custom filesystem tools (read, write, edit, glob, grep) — Codex handles all filesystem ops.
• Removed [roles.explorer] config section (kept in default.toml for compatibility but unused).

Added

• ReAct runtime with a provider-agnostic LM wrapper for multi-provider support (MiniMax, ZAI, OpenRouter, OpenAI, Ollama, MLX).
• Codex tool as intelligent filesystem sub-agent with kernel-level sandboxing.
• Unified providers.py -- all providers use the same LM wrapper path (no proxy layer needed).
• Cross-session intelligence in maintain: signal amplification, contradiction detection, gap detection.
• Cross-agent knowledge synthesis: detects patterns across Claude, Cursor, Codex, OpenCode sessions.
• Context curation with Active Decisions, Key Learnings, Recent Context, and Watch Out sections.
• Memory outcome field (worked/failed/unknown) for feedback tracking.
• Docker container hardening: read_only root, cap_drop ALL, seccomp profile, mount only .lerim/ dirs.
• Dashboard Intelligence tab: memory health score, contradictions, signals, gaps, cross-agent insights.

[0.1.68] - 2026-03-21

Added

• Server readiness check on lerim up: cli.py now polls /api/health for up to 30 seconds after starting the container, printing a clear warning if the server never responds.
• pytest-timeout added to the [test] optional dependency group for controlled test execution time.
• MINIMAX_API_KEY added to the environment-variable look-up list in the HTTP API.

Fixed

• Docker dashboard path: dashboard.py resolves the dashboard directory correctly inside containers by falling back to /opt/lerim/dashboard when the repo-relative path does not exist. A corresponding COPY dashboard/ step is added to the Dockerfile.
• Test pythonpath: [tool.pytest.ini_options] now includes pythonpath = ["."] so from tests.helpers import ... resolves when running uv run pytest.
• HTTP API key list: _API_KEY_ENV_NAMES is sorted alphabetically and now includes ANTHROPIC_API_KEY and MINIMAX_API_KEY.

[0.1.66] - 2026-03-15

Added

• Parallelism support: three config knobs control concurrent execution:
  • max_workers in [roles.extract]: parallel extraction window processing via ThreadPoolExecutor (each thread gets its own LM instance for thread safety).
  • max_explorers in [roles.explorer]: concurrent explorer subagent calls per lead turn.
  • parallel_pipelines in [server]: run extract + summarize pipelines in the same tool turn.
• Async explore tool: explorer subagent changed from sync (run_sync) to async (await agent.run), enabling true concurrent dispatch via PydanticAI's asyncio.create_task.
• Adaptive prompts: sync and maintain prompts now emit parallel or sequential instructions based on config values. Set knobs to 1/false for local models.

Removed

• max_workers from [roles.summarize] — summarization uses sequential refine/fold, so window parallelism does not apply.

[0.1.60] - 2026-03-05

Added

• Ollama lifecycle management: automatic model load/unload around sync and maintain cycles. Models are warm-loaded into GPU/RAM before each cycle and unloaded after (keep_alive: 0) to free 5-10 GB of memory between runs. Controlled by auto_unload = true in [providers].
• Proxy bridge support: new proxy provider base URL in [providers] for routing PydanticAI OpenAI-format calls to Ollama's native API (enables thinking mode control).
• Eval framework: four eval pipelines (extraction, summarization, sync, maintain) with LLM-as-judge scoring, config-driven model comparison, and bench_models.sh multi-model benchmarking script.
• Eval configs for Ollama models (Qwen3.5 4B/9B, thinking/non-thinking) and MiniMax-M2.5 cloud baseline.
• Synthetic eval traces and judge prompt templates for all four pipelines.
• evals/compare.py for cross-config result comparison.
• lerim skill install command to copy skill files into agent directories.

Fixed

• Docker networking: generated docker-compose.yml now includes extra_hosts: host.docker.internal:host-gateway so containers can reach Ollama running on the host.

Changed

• Evals folder reorganized: active configs moved to evals/configs/, stale MLX configs removed.
• Default provider switched to MiniMax-M2.5 with Z.AI fallback.

[0.1.53] - 2026-03-01

Fixed

• Daemon loop: maintain never triggered on startup in Docker containers where time.monotonic() reflected VM uptime smaller than the maintain interval (60 min).
• Daemon loop: sync/maintain cycles produced zero log output, making lerim logs appear idle. Added per-cycle status logging.
• Session queue: NULL repo_path jobs clogged the claim queue, preventing valid sessions from being extracted. Added filter in claim_session_jobs and guard in enqueue_session_job.
• DB migration: orphaned NULL repo_path pending/failed jobs are now purged on schema init.
• Explorer subagent: switched from structured ExplorerEnvelope output to plain str to avoid repeated output-validation failures with models that return empty responses after tool calls.
• Explorer failures no longer crash the lead agent; the explore tool returns empty evidence and logs a warning.
• Maintain action path validation: handle list-valued source_path/target_path from LLM output (model sometimes returns multiple paths per action).
• run_maintain_once now accepts a trigger parameter instead of hardcoding "manual" for all service-run records.

[0.1.5] - 2026-03-01

Added

• Per-run LLM cost tracking via OpenRouter's usage.cost response field. Cost (USD) logged in activity.log and returned in sync/maintain/ask result payloads.
• Chronological (oldest-first) session processing for correct memory ordering.

Changed

• Structured context-writing tool replaces raw file writes.
• _process_claimed_jobs runs sequentially (was parallel) for chronological memory consistency.
• Activity log format now includes cost column.

[0.1.0] - 2026-02-28

Added

• Docker service architecture: always-on daemon + HTTP API + dashboard in a single container.
• lerim init interactive setup wizard for first-time configuration.
• lerim project add/list/remove for incremental project registration.
• lerim up/down/logs for Docker container lifecycle management.
• lerim serve command — combined HTTP API + dashboard + daemon loop (Docker entrypoint, also usable directly without Docker).
• Service commands (ask, sync, maintain, status) are thin HTTP clients that talk to the running server.
• HTTP API: /api/health, /api/ask, /api/sync, /api/maintain, /api/status, /api/connect, /api/project/*.
• [agents], [projects], and [providers] config sections in config.toml.
• Provider API base URLs configurable via [providers] section (no more hardcoded URLs).
• Dockerfile with Python 3.12, health check, lerim serve entrypoint.
• Same-path volume mounting for zero path translation between host and container.
• Continual learning layer for coding agents and projects.
• Platform adapters for Claude Code, Codex CLI, Cursor, and OpenCode.
• Memory extraction pipeline using ChainOfThought with transcript windowing to extract decisions and learnings from coding session traces.
• Trace summarization pipeline using ChainOfThought with transcript windowing to produce structured summaries.
• PydanticAI lead agent with a read-only explorer subagent for memory operations.
• Three CLI flows: sync (extract, summarize, write memories), maintain (merge, archive, decay), and ask (query memories).
• Daemon mode for continuous sync and maintain loop.
• Local read-only web dashboard with HTTP API.
• Session catalog with SQLite FTS5 for session search.
• Job queue with stale job reclamation.
• TOML-layered configuration: shipped defaults, global, project, and env var override.
• OpenTelemetry tracing via Logfire with PydanticAI and runtime instrumentation.
• Multi-provider LLM support: OpenRouter (with Nebius routing), Ollama, ZAI, OpenAI, Anthropic.
• SQLite-backed context model for durable records and derived indexes.
• Project-scoped context records in the global context DB.
• Context record kinds: decisions, facts, procedures, preferences, and episodes.
• Comprehensive test suite with 290 tests across unit, smoke, integration, and e2e layers.
• Skills distribution via npx skills add lerim-dev/lerim.