feat: add work system (forms, rules, engine) and CLI enhancements

docs/adrs/ADR-0003-claude-code-marketplace.md

@@ -14,7 +14,7 @@ Claude Code's plugin marketplace provides a tested distribution primitive: a roo
 
 Adopt the Claude Code marketplace pattern inside the lionagi repository. The structure is:
 
-```
+```text
 .claude-plugin/marketplace.json          # root manifest
 marketplace/<plugin>/.claude-plugin/plugin.json  # per-plugin manifest
 marketplace/<plugin>/skills/             # bundled skills (added in later plays)
@@ -75,12 +75,14 @@ shipped in Phase 0 and why the remaining four were removed or deferred.
 ## Consequences
 
 **Positive**
+
 - Users install only the capability slices they need, keeping Claude Code context lean.
 - Each plugin can version independently; `studio` can ship a breaking MCP config change without bumping `devx`.
 - MCP server configuration can be bundled per plugin (`studio`, `mcp-bundle`) once the FastAPI backend route set is stable.
 - Clear ownership boundary: each plugin directory is a self-contained unit that external contributors or downstream forks can understand and extend.
 
 **Negative**
+
 - More manifests to maintain: root `marketplace.json` plus four `plugin.json` files (v2) must stay in sync as plugin names or descriptions change.
 - Skills authored in `firm/resources/skills/` (canonical) must be copied or symlinked into `marketplace/<plugin>/skills/` for external installs — two places to update per skill change.
 - The `plugin.json` schema is not yet finalized by Anthropic; field names or required keys may shift before GA, requiring a sweep across all four manifests.

docs/adrs/ADR-0004-filesystem-data-layer.md

@@ -3,6 +3,12 @@
 **Status**: Accepted
 **Date**: 2026-05-19 (revised 2026-05-20)
 
+---
+
+> **Related update**: [ADR-0033](ADR-0033-unified-entity-state-model.md) and [ADR-0034](ADR-0034-frontend-data-and-state-architecture.md) formalize how data flows from the filesystem layer through SQLite into the frontend. The hybrid filesystem-as-source-of-truth model here remains accurate; ADR-0033 adds `NormalizedState` as the read-side projection, and ADR-0034 specifies the sync contract for frontend consumption.
+
+---
+
 ## Context
 
 Lion Studio's backend serves runs, agents, playbooks, plugins, shows, and sessions.
@@ -100,12 +106,14 @@ into SQLite at any time.
 ## Consequences
 
 **Positive**
+
 - Authored content stays in git-friendly files editable by any tool.
 - Operational queries are fast (SQLite indexes, JOINs, aggregates).
 - No external database process — SQLite is embedded.
 - CLI and Studio share the same data without coordination protocol.
 
 **Negative**
+
 - Two data sources means two things to keep in sync.
 - Import commands needed for historical data migration.
 - Contributors must know which data lives where.

docs/adrs/ADR-0006-sse-live-streaming.md

@@ -3,6 +3,12 @@
 **Status**: Accepted
 **Date**: 2026-05-19 (revised 2026-05-20)
 
+---
+
+> **Related update**: This ADR defines the SSE transport. [ADR-0034](ADR-0034-frontend-data-and-state-architecture.md) §"SSE architecture" defines the higher-level event taxonomy, dispatch strategy, and integration with TanStack Query cache. ADR-0034 is the canonical owner of event names and envelope structure; this ADR remains authoritative for the wire protocol and reconnect mechanics.
+
+---
+
 ## Context
 
 Lion Studio needs live updates for three surfaces: session message streams
@@ -81,12 +87,14 @@ stop reconnecting.
 ## Consequences
 
 **Positive**
+
 - `EventSource` is native in all modern browsers; no client library.
 - Starlette `StreamingResponse` integrates with the existing FastAPI stack.
 - One-way constraint prevents bidirectional state complexity.
 - No WebSocket server, connection upgrade, or ping/pong management.
 
 **Negative**
+
 - SSE is strictly server-to-client. Any client-to-server action requires a
   separate REST endpoint.
 - HTTP/1.1 browser connection limits (6 per origin) constrain concurrent SSE

docs/adrs/ADR-0007-plugin-auto-discovery.md

@@ -19,7 +19,7 @@ discovers those by scanning `skills/` and `agents/` subdirectories.
 `plugin.json` for every lionagi marketplace plugin does NOT enumerate `skills` or `agents` arrays.
 The Claude Code plugin loader auto-discovers skill and agent files from the directory layout:
 
-```
+```text
 marketplace/<plugin>/
   .claude-plugin/plugin.json   # metadata only — no skills/agents arrays
   skills/                      # auto-discovered by CC loader
@@ -32,11 +32,13 @@ arrays) would create a redundancy that lags the actual directory contents.
 ## Consequences
 
 **Positive**
+
 - Adding a skill file to `skills/` is immediately effective; no manifest update required.
 - Manifest stays small and readable — metadata only.
 - Consistent with the only production reference implementation available at time of decision.
 
 **Negative**
+
 - Auto-discovery behavior is undocumented by Anthropic and may change before GA. If the CC
   loader stops auto-discovering from directory layout, all four plugins break silently — there
   is no explicit manifest to fall back on.

docs/adrs/ADR-0009-sqlite-state-layer.md

@@ -3,6 +3,12 @@
 **Status**: Accepted
 **Date**: 2026-05-20
 
+---
+
+> **Related update**: This ADR establishes SQLite as the current persistent state layer. [ADR-0033](ADR-0033-unified-entity-state-model.md) introduces `NormalizedState` (lifecycle × health × delivery) as the contract for entity state; the SQLite schema here is the current materialization of that contract. [ADR-0034](ADR-0034-frontend-data-and-state-architecture.md) §"Backend-frontend sync contract" specifies how this layer's reads flow to the frontend. The schema extensions for the new dimensions are in ADR-0033 Appendix A. SQLite-specific details remain authoritative; future stores (Postgres, distributed) will materialize the same NormalizedState model differently.
+
+---
+
 ## Context
 
 Lion Studio's filesystem-only backend (ADR-0004) works for post-hoc review but
@@ -129,7 +135,7 @@ Branch config (provider, model, system_prompt, tools, effort) lives in
 > stored in `node_metadata` JSON on the branch row, not as top-level columns. A
 > dedicated branch fork protocol may promote these to first-class columns in a
 > future migration if query patterns require it.
-
+>
 > **`system_msg_id` is first-class.** Unlike fork pointers, the system prompt
 > reference is promoted to a top-level column because (a) every live persist
 > path writes it during branch initialization, (b) the Studio inbox/branch view
@@ -174,6 +180,7 @@ grep, git, symlinks).  ADR-0004 remains valid for these.
 ### Supporting code changes (same session)
 
 **`Branch.to_dict()` streamlined** (commits 7ac31109, a36b5542):
+
 - Added `mode` parameter: `"python"` (default), `"json"`, `"db"`.
 - `mode="db"` renames `metadata` → `node_metadata`.
 - Optional flags: `include_logs`, `include_log_config`,
@@ -182,6 +189,7 @@ grep, git, symlinks).  ADR-0004 remains valid for these.
 - `parse_model` only included when it differs from `chat_model`.
 
 **`create_message()` extracted** (commit 8f6c6d94):
+
 - `MessageManager.create_message()` is now a `@staticmethod` — can be used
   without a MessageManager instance.
 - Standalone `create_message()` function exported from
@@ -190,6 +198,7 @@ grep, git, symlinks).  ADR-0004 remains valid for these.
   insertion and system message replacement.
 
 **`iModel.to_dict()` slimmed** (commits 7ac31109, a36b5542):
+
 - `request_options` excluded by default (bulk, not useful for persistence).
 - `processor_config` excluded by default.
 
@@ -218,6 +227,7 @@ on legacy versions) can drop the autocheckpoint setting or run
 ## Consequences
 
 **Positive**
+
 - Session/branch/message model has a persistent representation that mirrors the
   runtime 1:1. Round-trip `to_dict(mode="db")` → INSERT → SELECT → `from_dict()`
   is lossless.
@@ -230,6 +240,7 @@ on legacy versions) can drop the autocheckpoint setting or run
 - Agent/playbook file-based contract preserved.
 
 **Negative**
+
 - `aiosqlite` is now a mandatory dependency (promoted from `lionagi[sqlite]` optional).
 - JSON array for progression ordering limits query-side operations (no
   `WHERE message_id IN progression` without JSON parsing).

docs/adrs/ADR-0010-plugin-aware-studio.md

@@ -33,6 +33,7 @@ A new service scans two locations for plugins:
 | Third-party | `~/.claude/plugins/cache/*/*/` | No (read-only) |
 
 For each plugin directory, the service:
+
 - Reads `.claude-plugin/plugin.json` for metadata (name, version, description)
 - Falls back to directory name if no manifest exists
 - Enumerates `skills/*/SKILL.md`, `agents/*.md`, `hooks/hooks.json`, `.mcp.json`
@@ -43,7 +44,7 @@ list for repo plugins. Third-party plugins are discovered by directory scan.
 
 ### 2. API surface
 
-```
+```text
 GET  /api/plugins/              → list all plugins with component counts
 GET  /api/plugins/{name}        → plugin detail: metadata + component lists
 GET  /api/plugins/{name}/skills/{skill}  → skill content (read-only)
@@ -62,6 +63,7 @@ Navigation changes from `Playbooks | Agents | Skills | Runs | Shows` to
 name or `Lion Marketplace`), component counts (N skills, M agents).
 
 **Plugin detail page**: tabbed layout:
+
 - **Skills** tab: two-pane list+detail, read-only SKILL.md viewer
 - **Agents** tab: agent names/descriptions with `Open in Agents →` cross-links
 - **Hooks** tab: `hooks.json` viewer (read-only)
@@ -94,11 +96,13 @@ Plugin components need explicit navigation to their editable counterparts and
 clear source/editability indicators:
 
 **Cross-links from plugin detail tabs**:
+
 - Agent tab: `Open in Agents →` link per agent (routes to `/agents/{name}`)
 - Skill tab: `View source` link, `Copy path` action
 - All tabs: `Open raw` for filesystem path
 
 **Source badges on plugin list and detail**:
+
 | Badge | Meaning |
 |-------|---------|
 | `Lion Marketplace` | From `marketplace/`, writable |
@@ -119,19 +123,22 @@ font with code fences in monospace, not the current all-monospace treatment.
 ### 6. Deferred: CLI sync, monitors, themes, LSP
 
 These are not part of this ADR:
+
 - `li plugin sync` CLI (marketplace → CC plugin install) — separate feature
 - Monitors, themes, LSP components — surface in Studio when usage warrants
 - Model-agnostic abstraction — current format is CC-native; abstract later
 
 ## Consequences
 
 **Positive**
+
 - Skills and agents gain context: users see which plugin provides each capability.
 - Editing flows through existing definitions versioning — no new persistence layer.
 - Third-party plugin visibility without management complexity.
 - Direct alignment with CC's plugin model — what you see in Studio is what CC loads.
 
 **Negative**
+
 - Plugin discovery adds startup scan cost (~22ms for 4 plugins, negligible).
 - Two navigation paths to the same skill (Plugins → devx → commit vs. direct URL).
 - `plugin.json` may not exist for all marketplace plugins yet — fallback to dir name.

docs/adrs/ADR-0011-shows-data-model.md

@@ -4,6 +4,12 @@
 **Date**: 2026-05-20
 **Extends**: ADR-0009 (SQLite state layer), ADR-0010 (plugin-aware Studio)
 
+---
+
+> **Related update**: The shows/plays schema defined here stores structural and operational state. [ADR-0033](ADR-0033-unified-entity-state-model.md) adds `NormalizedState` columns (health, delivery, reason_codes) additively to these tables. The lifecycle vocabularies for Shows and Plays in ADR-0033 §"Per-entity lifecycle vocabularies" are derived from this ADR's status enums. This ADR remains authoritative for the table structure and cardinality; ADR-0033 governs the meaning of state values written to those tables.
+
+---
+
 ## Context
 
 Shows (multi-play DAGs orchestrated by the `show` skill) currently exist only
@@ -140,7 +146,7 @@ accessible: `plays.session_id` → `sessions` → `branches` → `progressions`
 
 ### Read path (Studio API)
 
-```
+```text
 GET /api/shows/           → SELECT from shows table (fast, no filesystem)
 GET /api/shows/{topic}    → shows row + plays rows + _show.md from disk
 GET /api/shows/{topic}/plays/{name}
@@ -250,7 +256,7 @@ ADR-0012), not `/sessions/{session_id}`.
 The plays table uses a single **State** column with a primary lifecycle pill
 plus optional secondary badges for gate and integration:
 
-```
+```json
 [completed] [passed] [merged]      ← full provenance
 [completed]                        ← no gate, no merge
 [awaiting gate]                    ← running_complete
@@ -263,6 +269,7 @@ a single compound pill (hard to scan because dimensions aren't visually
 separable).
 
 **Badge colors**:
+
 - Lifecycle: pending (amber), running (blue), awaiting_gate (amber),
   completed (green), failed (red), aborted (gray)
 - Gate: passed (green outline), failed (red outline), skipped (gray outline)
@@ -279,7 +286,7 @@ lifecycle + gate + review badges per ADR-0012's status vocabulary.
 The PlayDag renders as a **compact dependency graph strip above the plays
 table**, not a competing primary view:
 
-```
+```text
 Dependency graph                                        [Hide graph]
 ┌────────────────────────────────────────────────────────────────────┐
 │  compact PlayDag, fitView, 150–170px tall                          │
@@ -290,6 +297,7 @@ Dependency graph                                        [Hide graph]
 ```
 
 Height scales with play count:
+
 - ≤4 plays: 112px
 - 5–20 plays: 160px
 - 21+ plays: 220px with scroll/pan
@@ -301,6 +309,7 @@ a plain status table and loses the orchestration identity of the page.
 ### Show status provenance (added post-review)
 
 Show status includes provenance to address trust concerns:
+
 - `active` — at least one play running or pending
 - `completed` — all plays merged and final gate passed (or inferred from filesystem)
 - `aborted` — `_ABORT` file exists
@@ -317,6 +326,7 @@ indicate confidence level.
 ## Consequences
 
 **Positive**
+
 - Show → play → session → message drill-down via foreign keys.
 - Fast list/filter queries without filesystem scanning.
 - Structural queries: blocked plays, critical path, cross-show stats.
@@ -326,6 +336,7 @@ indicate confidence level.
 - Status normalization eliminates contradictory displays across pages.
 
 **Negative**
+
 - Dual write: every `_meta.json` update must also update SQLite. If the
   show skill crashes between writes, they can drift. Mitigation: the
   `li state import-shows` command can re-sync from filesystem at any time.