feat: workspace naming & user isolation (ISS-39)

Author: chuongld20

.claude/scheduled_tasks.lock

@@ -0,0 +1 @@
+{"sessionId":"f9f1e5ad-00a2-4034-812b-4a9eb6a494b7","pid":741641,"acquiredAt":1776017431094}

.claude/skills/agent-optimize/SKILL.md

@@ -0,0 +1,66 @@
+---
+name: agent-optimize
+description: |
+  Audit and optimize a Forge chat agent for MCP projects.
+  Tests queries, checks schema/memory/patterns in Qdrant,
+  updates behaviorRules, and validates improvements.
+  Covers: schema cache, alias batching, tool patterns, knowledge graph.
+user_invocable: true
+argument_description: "<project-slug> [hubToken]"
+---
+
+# Agent Optimize
+
+Audit + optimize a Forge MCP chat agent in 7 phases.
+
+## Phases
+
+1. **Baseline** — Run `/chat-eval` to get gap metrics
+2. **Inspect** — Check Qdrant for schema sections, tool patterns, memories. See `references/qdrant-queries.md`
+3. **Config** — Fetch project agentConfig, review behaviorRules/strategies. See `references/config-review.md`
+4. **Test** — Run diverse queries (LOOKUP/SEARCH/SUMMARY/ACTION/CHAT). See `references/test-matrix.md`
+5. **Fix** — Update behaviorRules via API for failed patterns. See `references/rule-writing.md`
+6. **Slim** — Reduce tokens: compress behaviorRules + optimize RAG context. See `references/token-optimization.md`
+7. **Re-test** — Same queries, compare iterations/GQL calls/schema calls
+8. **Eval** — Run `/chat-eval` again, compare with baseline
+
+## Key Targets
+
+| Metric | Target |
+|--------|--------|
+| graphql_schema calls | 0 |
+| Avg iterations | ≤6 |
+| 502 timeouts | 0 |
+| Empty replies | 0 |
+| Tool patterns stored | Growing over time |
+| behaviorRules chars | Minimized (no redundancy with Qdrant) |
+
+## Known Projects
+
+| Project | Slug | API Key | Hub Token |
+|---------|------|---------|-----------||
+| Light Human | `portal-lh` | `fk_0186f0a601488027ac1661698e9a6126e720274a47e96759` | `3291\|PNi8lVJ3PhALqYflnvvrZe33gHLOxCtqrobuVgPEcfcd762e` |
+
+## Quick Commands
+
+```bash
+# Chat-eval baseline
+cd forge/strapi && npx tsx scripts/eval-chat-gaps.ts --project <slug> --user thanhnh --pass 'Falconx@@123' --limit 200
+
+# Test a query
+curl -sS -g -X POST 'https://forge-api.sidcorp.co/api/chat' \
+  -H 'Content-Type: application/json' -H 'X-Forge-API-Key: <key>' \
+  --data-raw '{"message":"<query>","stream":false,"hubToken":"<token>"}' --max-time 180
+
+# Trigger MCP sync
+curl -sS -X POST 'https://forge-api.sidcorp.co/api/knowledge/sync' -H 'X-Forge-API-Key: <key>' -d '{"projectSlug":"<slug>"}'
+```
+
+## References
+
+- `references/qdrant-queries.md` — Qdrant scroll/search commands for schema, memories, patterns
+- `references/config-review.md` — How to fetch/update agentConfig, what to check
+- `references/test-matrix.md` — Test query matrix and expected behavior per intent
+- `references/architecture.md` — Memory/edge/pattern data flow and key files
+- `references/common-fixes.md` — Solutions for schema loops, N+1, timeouts, empty replies
+- `references/token-optimization.md` — Compress behaviorRules without losing performance

.claude/skills/agent-optimize/references/architecture.md

@@ -0,0 +1,40 @@
+# Architecture
+
+## Data flow: Chat → Learning
+
+```
+Chat turn completes (fire-and-forget):
+  ├── extractMemories() [LLM: gemini-flash]
+  │     ├── facts[] → Qdrant memory (preference/correction/convention/tool_pattern)
+  │     └── edges[] → Strapi DB knowledge_edges (subject→predicate→object)
+  └── extractToolPatterns() [no LLM]
+        └── successful GQL calls → Qdrant memory (tool_pattern)
+```
+
+## Data flow: Query → Context
+
+```
+User query → ragGate (intent + condense) → multiStrategySearch → reranker → context
+  → schema force-inject if MCP project missing schema in top results
+  → knowledge graph 1-hop edge expansion → edgeContext
+  → system prompt = base rules + behaviorRules + queryStrategy + RAG context + edges
+```
+
+## Reranker boosts (source_type)
+
+mcp_schema +0.20, skill +0.15, memory +0.12, knowledge +0.10, chat_session +0.08, issue +0.05
+
+## Key files
+
+| File | Purpose |
+|------|---------|
+| `services/agent/memory.ts` | extractMemories, addMemory, extractToolPatterns |
+| `services/agent/memory-lifecycle.ts` | Prune stale memories/edges (6h cron) |
+| `services/knowledge-graph/index.ts` | upsertEdge, queryEdges, extractIssueEdges |
+| `services/embeddings/reranker.ts` | Source-type boosts |
+| `services/agent/mcp-sync.ts` | MCP knowledge sync + schema sections |
+| `api/chat/services/chat-prompt-builder.ts` | RAG retrieval, schema force-inject |
+| `services/agent/system-prompt.ts` | Layered prompt builder |
+| `bootstrap/seeds/domain-templates.ts` | Template defaults (copy-on-apply) |
+
+All paths relative to `forge/strapi/src/`.

.claude/skills/agent-optimize/references/common-fixes.md

@@ -0,0 +1,45 @@
+# Common Fixes
+
+## Agent calls graphql_schema every conversation
+- Check Qdrant for `mcp_schema` entries — missing = trigger sync
+- Reranker boost is +0.20 for mcp_schema (highest)
+- Force-inject in chat-prompt-builder ensures schema in context for MCP projects
+- If schema sections are stale, re-sync: `POST /api/knowledge/sync`
+
+## N+1 graphql_query calls (same query, different IDs)
+- Add alias batching rule to behaviorRules with **specific example** using actual query names
+- Generic: "NEVER loop graphql_query. Use aliases: `{ a1: query(id:1){...} a2: query(id:2){...} }`"
+- Batch up to 20 aliases per call
+
+## 502 timeout (>100s Cloudflare limit)
+- Agent doing too many iterations before responding
+- Add behaviorRule with the efficient pattern: "1-2 GQL calls + code_run"
+- Ensure `code_run` in `enabledTools`
+- Check if query needs a new aggregation endpoint on the portal side
+
+## Empty reply (0 chars)
+- Check `qualitySignals.iterationLogs` in chat-log DB for the failing query
+- Common causes:
+  1. **Large tool results** — GQL returns 30k+ chars, fills context, no room for text output
+  2. **TodoWrite burns iterations** — agent plans instead of acting (stream:false has no UI benefit)
+  3. **stopReason=max_tokens** on final iter — output budget spent on tool_use JSON
+  4. **stopReason=end_turn with textChars=0** — model chose to stop without text (model bug)
+- Diagnosis query: `GET /api/chat-logs?sort=createdAt:desc&filters[projectSlug][$eq]=<slug>` (JWT auth)
+- Fix: add rule to fetch minimal fields, then aggregate with code_run
+- Prefer `language: "javascript"` (python3 sometimes not found on prod)
+
+## Tool patterns not stored
+- Check `extractToolPatterns` wired in `chat.ts` and `channels/handler.ts` post-turn
+- Only stores successful `*__graphql_query` calls with result > 50 chars
+- Dedup at 0.85 cosine — similar queries won't duplicate (expected)
+
+## Knowledge graph edges empty for MCP projects
+- Normal for projects without Forge issues
+- Edges from chat only appear when users discuss structural relationships
+- Tool patterns (in Qdrant) are more valuable for MCP query optimization
+
+## Rule writing tips
+- Reference exact query names and field names from the schema
+- Include a working GQL snippet, not just "use the right query"
+- One concern per rule — don't mix alias batching with field naming
+- Test the rule by running the exact GQL in the snippet before adding

.claude/skills/agent-optimize/references/config-review.md

@@ -0,0 +1,45 @@
+# Config Review
+
+## Fetch project agentConfig
+
+```bash
+curl -sS -g 'https://forge-api.sidcorp.co/api/projects?filters[slug][$eq]=<slug>&fields[0]=agentConfig&fields[1]=documentId' \
+  -H 'X-Forge-API-Key: <api-key>'
+```
+
+## Update agentConfig
+
+Requires JWT auth (API key is read-only for projects):
+
+```bash
+JWT=$(curl -sS -g 'https://forge-api.sidcorp.co/api/auth/local' \
+  -H 'Content-Type: application/json' \
+  -d '{"identifier":"thanhnh","password":"Falconx@@123"}' | python3 -c "import json,sys; print(json.load(sys.stdin)['jwt'])")
+
+curl -sS -g -X PUT "https://forge-api.sidcorp.co/api/projects/<DOC_ID>" \
+  -H 'Content-Type: application/json' -H "Authorization: Bearer $JWT" \
+  -d '{"data":{"agentConfig":{...}}}'
+```
+
+## Checklist
+
+**behaviorRules** — most impactful for MCP projects:
+- [ ] References exact GQL query names and field names (not generic advice)
+- [ ] Includes working GQL snippets for common patterns
+- [ ] Has alias batching rule with specific example
+- [ ] Has aggregation strategy: "1-2 GQL calls + code_run (JS)"
+- [ ] Domain terminology defined (acronyms, status mappings)
+
+**queryStrategies** — all 6 intents covered:
+- [ ] LOOKUP: mentions filters, fields, MCP tools
+- [ ] SEARCH: mentions search parameter, code lookups
+- [ ] SUMMARY: mentions statistics queries, date defaults
+- [ ] CREATE: mentions confirmation before creating
+- [ ] ACTION: mentions conversation history
+- [ ] CHAT: no tool calls needed
+
+**intentExamples** — 15-30 examples, multilingual if users speak multiple languages
+
+**enabledTools** — `code_run` required for aggregations, `forge_language` for language detection
+
+**Note**: Domain templates are copy-on-apply. Updating the seed template does NOT update existing projects. Always update the project's config directly via API.

.claude/skills/agent-optimize/references/qdrant-queries.md

@@ -0,0 +1,35 @@
+# Qdrant Queries
+
+Qdrant URL: `http://216.18.217.213:6333`, API key in `forge/strapi/.env` → `QDRANT_API_KEY`.
+Collection: `forge_embeddings`. Get project `documentId` from Strapi API first.
+
+## Schema sections (mcp_schema)
+
+```bash
+curl -sS -X POST '$QDRANT_URL/collections/forge_embeddings/points/scroll' \
+  -H 'Content-Type: application/json' -H "api-key: $QDRANT_API_KEY" \
+  -d '{"filter":{"must":[
+    {"key":"project_id","match":{"value":"<DOC_ID>"}},
+    {"key":"source_type","match":{"value":"mcp_schema"}}
+  ]},"limit":50,"with_payload":true,"with_vector":false}'
+```
+
+Each MCP server should have multiple sections (split by `## HEADER`). Single entry = re-sync needed.
+
+## Memories (tool_pattern, correction, preference, convention)
+
+```bash
+# Same endpoint, filter source_type: "memory"
+{"key":"source_type","match":{"value":"memory"}}
+```
+
+Check `metadata.category` distribution. Healthy project: 10+ tool_patterns after a few sessions.
+
+## Source type reference
+
+| source_type | Created by | Purpose |
+|-------------|-----------|---------|
+| `memory` | extractMemories, extractToolPatterns | Learned facts + working GQL patterns |
+| `mcp_schema` | syncMcpKnowledge Phase 5 | GraphQL schema sections per MCP server |
+| `hub_task/project/comment/config` | syncMcpKnowledge Phases 1-4 | MCP hub data |
+| `issue/skill/knowledge/chat_session` | Forge lifecycle hooks | Core project data |

.claude/skills/agent-optimize/references/test-matrix.md

@@ -0,0 +1,32 @@
+# Test Matrix
+
+Run 6-8 queries covering all intents. For each, check: iterations, GQL calls, schema calls, reply quality.
+
+## Query template
+
+```bash
+curl -sS -g -X POST 'https://forge-api.sidcorp.co/api/chat' \
+  -H 'Content-Type: application/json' -H 'X-Forge-API-Key: <key>' \
+  --data-raw '{"message":"<query>","stream":false,"hubToken":"<token>"}' --max-time 180
+```
+
+Parse response: `iterations`, `toolCalls[].name`, `toolCalls[].isError`, `reply` length.
+
+## Expected behavior per intent
+
+| Intent | Max iters | Schema calls | Pattern |
+|--------|-----------|-------------|---------|
+| CHAT | 1 | 0 | No tool calls |
+| LOOKUP | 2-3 | 0 | 1-2 GQL with filters |
+| SEARCH | 2-3 | 0 | GQL with search param |
+| SUMMARY | 3-5 | 0 | Statistics queries or batched aliases |
+| SUMMARY-complex | 4-6 | 0 | Alias batch + code_run |
+| ACTION | 1-3 | 0 | Uses conversation context |
+
+## Red flags
+
+- `graphql_schema` in toolCalls → schema cache not working
+- Same GQL query called 4+ times → alias batching rule missing
+- `code_run` with `language: "python"` failing → use JS instead
+- Reply 0 chars → output token exhaustion, reduce data payload
+- 502 response → >100s Cloudflare timeout, too many iterations

.claude/skills/agent-optimize/references/token-optimization.md

@@ -0,0 +1,27 @@
+# Token Optimization
+
+Reduce token cost on two fronts: what's stored (embeddings) and what's sent to agent (retrieval).
+
+## A. BehaviorRules Compression
+
+Target: <3000 chars total.
+
+| Rule content | Qdrant source | Action |
+|-------------|---------------|--------|
+| Inline status ID enums | `mcp_schema` ENUMS section | Remove IDs, keep name mappings only |
+| Full GQL examples | `tool_pattern` memories | Compress to query signatures |
+| "call graphql_schema" | Schema cache + force-inject | Change to "check schema in context" |
+
+## B. RAG Context Optimization
+
+**Key files:**
+- `services/rag-formatter/index.ts` — formats entries before injection
+- `services/agent/system-prompt.ts` `layerRelevantContext()` — MAX_CHARS=6000, 800c/entry
+
+## C. Verification
+
+After changes, test the same queries from Phase 4:
+- Schema calls must stay at 0
+- Iterations should not increase (≤2 more)
+- Reply quality must not degrade
+- Measure total inputTokens from response — should decrease

.claude/skills/antigravity-guide/SKILL.md

@@ -0,0 +1,75 @@
+---
+name: antigravity-guide
+description: "Environment guide for Antigravity server-side agent. High-level CLI for Forge API, Windows environment, common pitfalls. Auto-loaded by all pipeline steps."
+user_invocable: false
+---
+
+# Antigravity Environment Guide
+
+You are running inside **Antigravity** on a **Windows** host.
+
+## Environment
+
+- **OS**: Windows (`c:\Canaran.Antigravity.Service_Sessions\<sessionId>\`)
+- **Repo**: Cloned into a subdirectory of your session root (not the root itself)
+- **Node.js**: Available
+- **Git**: Available
+- **Shell**: May NOT have Bash — use native file tools (read, write, list)
+
+## Forge API — Use the CLI
+
+**DO NOT use curl.** Use the high-level CLI instead:
+
+```
+node forge-api.mjs get-issue <documentId>
+node forge-api.mjs update-issue <documentId> --status=confirmed --category=bug
+node forge-api.mjs update-issue <documentId> --plan=@plan.md --status=approved
+node forge-api.mjs update-issue <documentId> --data-file=payload.json
+node forge-api.mjs search-issues "keyword1 keyword2" --exclude=<documentId> --limit=10
+node forge-api.mjs list-comments <issueDocumentId> --limit=5
+node forge-api.mjs create-comment <issueDocumentId> --body=@file.md --author=Snorlax
+node forge-api.mjs create-comment <issueDocumentId> --body=@file.md --author=Snorlax --attachments=42,43
+node forge-api.mjs upload <filepath>
+```
+
+### Flags
+
+- `--status=`, `--category=`, `--priority=`, `--complexity=` — simple string values, safe inline
+- `--plan=@plan.md`, `--body=@file.md` — read value from file (for long text)
+- `--relations=@relations.json` — read JSON from file
+- `--data-file=payload.json` — read entire update body from JSON file
+- `--attachments=42,43,44` — comma-separated media IDs (from upload command)
+
+### Complex values: always use file-based payloads
+
+For any value containing JSON (relations, sessionContext, arrays, objects), write to a file first and reference it.
+
+**Updating an issue with complex fields:**
+1. Write JSON file: `{ "status": "confirmed", "relations": [...] }`
+2. Pass via `--data-file`: `node forge-api.mjs update-issue <documentId> --data-file=payload.json`
+
+**Creating a comment with long body:**
+1. Write body to `.md` file
+2. `node forge-api.mjs create-comment <issueDocId> --body=@comment.md --author=Snorlax`
+
+### Uploading Files (Screenshots)
+
+```
+node forge-api.mjs upload screenshot.png
+```
+Response: `{ "data": { "id": 42, "url": "...", "name": "screenshot.png" } }`
+
+## File Operations
+
+Use native tools, not Unix commands:
+- **Read file**: file read tool (not `cat`)
+- **List directory**: directory listing tool (not `ls`)
+- **Write file**: file write tool (not `echo >`)
+
+## Common Errors
+
+1. **JSON parse error** → write data to a file first, use `--data-file=` or `--field=@file.json`
+2. **"Bash not available"** → use native tools + `node forge-api.mjs`
+3. **API errors** → never use curl, always use the CLI
+4. **"command not found: cat/ls"** → use `type`/`dir` or native tools
+5. **Git fails** → `cd` into repo dir first (not session root), then `git fetch && git pull`