@turmind/halo 0.1.9 → 0.2.0

package/bundled-docs/test/session.md

@@ -1,374 +0,0 @@
-# Session & Channel Test Cases
-
-Manual smoke + regression checklist before a release. Last refreshed for the 2026-05 schema/refactor wave.
-
-**Test environment** (matches `reference_test_env.md` memory):
-- Local URL: `http://localhost:9527/?folder=<absolute-workspace-path>`
-- Password: `HALO_PASSWORD` env var (or the random one printed on first startup)
-- Workspace under test: `/path/to/test-workspace`
-- Required workspace agents: `default`, `sleeper`, `test-agent`
-- **Delegation gating (since the team-driven switch):** the agent that *initiates* a delegation must list the target in its `agent.yaml` `team`. The seed `default`'s team is `[default, executor, deep-executor]`, which does **not** include `sleeper` / `test-agent` — so the delegation/inspection cases (A6 `query_agent sleeper`, D1/D5 `start_session sleeper`, G3 `test-agent → sleeper`) require the initiating agent's `team` to contain `sleeper` (and `test-agent`). Add them to the test agent's `team`, or these cases fail with "not in your team / no session tools".
-- Required env vars for the LLM-touching cases: at least one of `KIMI_API_KEY` / `DEEPSEEK_API_KEY` / AWS credentials, configured via Settings page.
-
-**How to use this doc**:
-- Run the **automated regression script first** (covers ~25% of cases below): `HALO_TEST_PASSWORD=... HALO_TEST_PROJECT=/path/to/test-workspace node packages/server/tests/test-session-system.mjs`
-- Then walk the table top-to-bottom for full release validation.
-- 🟢 = automated covers it; 🟡 = automated partial; ⬜ = manual only.
-
----
-
-## A. Agent system (REST + UI, no LLM)
-
-| # | Coverage | Scenario | Method | Expected |
-|---|---|---|---|---|
-| A1 | 🟢 | Default agent loads from global | `GET /api/agent-configs` | `default` present, `scope=global`, `priority=99` |
-| A2 | 🟢 | Workspace agents visible | `GET /api/agent-configs?projectId=...` | `sleeper` + `test-agent` listed alongside global ones |
-| A3 | 🟢 | Create workspace agent via API | `POST /api/agent-configs` then `GET .../yaml` | YAML round-trips name/description/model |
-| A4 | 🟢 | Delete workspace agent | `DELETE /api/agent-configs/:id?scope=workspace` | 200; gone from list |
-| A6 | 🟢 | `query_agent` tool | Chat: "Use query_agent to inspect 'sleeper'" | Full fields (model/tools/skills) |
-| A8 | ⬜ | Provider switch carries over thinking effort if still valid | UI: change Bedrock → Bedrock other model with same effort presets | Effort label stays highlighted |
-| A9 | ⬜ | Provider switch resets effort when vocabulary differs | UI: Bedrock (effort=medium) → Kimi | Effort resets to Kimi's `defaultEnabled` config |
-| A10 | ⬜ | Provider switch turns prompt-caching default on for Bedrock | UI: switch any provider → Bedrock | Prompt caching toggle on, value = `1h` |
-| A11 | ⬜ | Provider switch clears prompt caching when target doesn't support it | UI: Bedrock (caching=1h) → DeepSeek | Caching toggle off, no stale `promptCaching:` field saved |
-| A12 | ⬜ | Agent form Context fields reflect settings.yaml `general.compact.compress_at` default | New agent without override | `compressAt` placeholder shows `0.8` |
-
-## B. Default agent basics (WS + LLM)
-
-| # | Coverage | Scenario | Input | Expected |
-|---|---|---|---|---|
-| B1 | 🟢 | Multi-turn context | "Say hello" → "What did I just say?" | Context preserved; recalls |
-| B2 | ⬜ | Stop + resume | "Write a 500 word essay" → Stop → "Continue" | Clean stop; resume produces continuation |
-| B3 | ⬜ | Refresh during tool | "sleep 10, then tell a joke" → refresh mid-sleep | Reconnect shows in-progress; completes |
-| B4 | ⬜ | Stop during tool | "sleep 10, then tell a joke" → Stop mid-sleep | Clean abort, no stray "stopped" badges on unrelated sessions |
-| B5 | ⬜ | New message interrupts | "sleep 10, then tell a joke" → new msg mid-sleep | New msg handled after current tool completes (graceful interrupt) |
-| B6 | ⬜ | Agent selector switch | Agent dropdown → `sleeper` → send msg | Session uses sleeper |
-| B7 | ⬜ | Session lock during conversation | Start chat then try switching agent | Dropdown locked |
-
-## C. Agent capabilities (LLM + tools + features)
-
-| # | Coverage | Scenario | Input | Expected |
-|---|---|---|---|---|
-| C1 | ⬜ | Tool call basic | "Read /tmp/test.txt" | `file_read` tool fires; result rendered |
-| C2 | ⬜ | Multiple tools in one turn | "List files in /tmp and read /tmp/test.txt" | All tool calls visible |
-| C3 | ⬜ | Long shell output | `find /path/to/halo -name '*.ts' -type f` | Truncation marker shown only past `toolResultMax` |
-| C4 | 🟢 | Prompt caching 5min | Two messages in succession | Second usage event has `cacheReadInputTokens > 0` |
-| C5 | ⬜ | Prompt caching 1h | Agent.yaml `promptCaching: 1h` | First message creates cache, subsequent reads it; `cacheWriteInputTokens` only on write |
-| C6 | ⬜ | Thinking adaptive low | Agent `thinking.effort: low` | Reply contains thinking block; small char count |
-| C7 | ⬜ | Thinking adaptive medium | `thinking.effort: medium` | Larger thinking block |
-| C8 | ⬜ | Thinking adaptive high | `thinking.effort: high` | Largest thinking block |
-| C9 | ⬜ | Thinking manual mode (Haiku 4.5) | Agent on `claude-haiku-4-5` w/ `thinking.budget_tokens: 8192` | Usage badge shows `think 8192` (NOT `think medium`) |
-| C10 | ⬜ | Streaming completes after error | Bad credentials in agent.yaml provider | Red "Error: ..." system message rendered; `isStreaming` reverts to false (no spinner stuck) |
-| C11 | ⬜ | Image input passthrough | Paste image into chat → "describe this" | Image saved under `<workspace>/.halo/assets/web/inbound/`, agent sees and describes |
-| C12 | ⬜ | Image filtered for non-vision provider | Paste image while on DeepSeek | Image dropped with `[图片已保存]` marker; agent still answers |
-
-**C4-C5 verification (raw usage):**
-```bash
-python3 -c "
-import json, glob
-files = sorted(glob.glob('/path/to/test-workspace/.halo/sessions/default/*.json'), key=lambda f: -__import__('os').path.getmtime(f))
-if files:
-    d = json.load(open(files[0]))
-    for msg in d.get('rawMessages', []):
-        u = msg.get('usage', {})
-        if u.get('cache_read_input_tokens', 0) > 0:
-            print(f'Cache hit: read={u[\"cache_read_input_tokens\"]} create={u.get(\"cache_creation_input_tokens\", 0)}')"
-```
-
-## D. Session tools (sub-agent delegation)
-
-| # | Coverage | Scenario | Input | Expected |
-|---|---|---|---|---|
-| D1 | 🟢 | Basic delegation | "Ask sleeper to sleep 10s then tell its height" | start_session → wait → auto-report |
-| D2 | ⬜ | query_session continues | After D1: "Ask it again, what's its weight?" | Same session reused (no second start_session) |
-| D3 | ⬜ | interrupt_session | "Ask sleeper to sleep 30s then tell story" → "Interrupt and ask weight" | interrupt fires, new task runs |
-| D4 | ⬜ | stop_session | "Ask sleeper to sleep 30s" → "Stop it" | clean stop |
-| D5 | ⬜ | New session | After D1: "Start another sleeper" | Fresh start_session (different sessionId) |
-| D6 | ⬜ | Parallel delegation | "Start two sleepers concurrently" | Both auto-report |
-| D7 | 🟢 | session_list inside agent | Agent calls session_list mid-conversation | Returns active children |
-| D8 | ⬜ | get_session_output | After D1: "Read full output from sleeper" | Returns full last-turn text |
-| D9 | ⬜ | archive_session | Agent calls archive_session on a sub-tree | Cascade archives, no longer in session_list |
-| D10 | ⬜ | Tool **only sees agent's declared params** | `test-agent` tries `{{tavily-web-search.params.api_key}}` in shell_exec but doesn't list tavily as a skill | Placeholder kept literal, curl 401 — verifies new namespace whitelist |
-
-## E. Session ops (REST + WS lifecycle)
-
-| # | Coverage | Scenario | Method | Expected |
-|---|---|---|---|---|
-| E1 | ⬜ | start_session emits sessionId | Chat triggers start_session | UI shows sub-session in tree, sessionId returned |
-| E2 | ⬜ | session_list during running | Agent calls during another's run | Status `running` |
-| E3 | ⬜ | query_session restores idle | Query a completed session | Restored from disk; reply produced |
-| E4 | ⬜ | query_session queues busy | Query a running session | Message queued, handled after current turn |
-| E5 | ⬜ | stop_session sets stoppedAt | Agent stops a peer | SQLite has stoppedAt; UI shows stopped badge |
-| E6 | 🟢 | delete_session cascades | WS `session:delete` | SQLite + JSON files for parent + descendants gone |
-| E7 | 🟢 | delete via REST goes through SessionManager | `DELETE /api/sessions/logs/:id` | Tombstone set; pending in-flight saves don't resurrect file |
-| E8 | ⬜ | Resurrection guard | Active session being subscribed; delete via REST while a `chat:complete` save is pending | File stays gone after delete (regression test for the WS save resurrection bug) |
-| E9 | ⬜ | UI Reset button hidden when value = default | Settings page general field at default | "Reset" link not visible |
-| E10 | ⬜ | UI Reset reverts to schema default | Settings page general field overridden, click Reset | Field shows muted-italic placeholder = schema default; source badge `unset` |
-
-**E6-E8 verification:**
-```bash
-# Before delete
-sqlite3 /path/to/test-workspace/.halo/halo.db "SELECT id, agent_id, parent_id FROM agent_sessions ORDER BY created_at DESC LIMIT 5;"
-ls -lt /path/to/test-workspace/.halo/sessions/*/ | head
-
-# After delete — both should be empty for the deleted id
-sqlite3 /path/to/test-workspace/.halo/halo.db "SELECT id FROM agent_sessions WHERE id = '<deleted-id>';"
-find /path/to/test-workspace/.halo/sessions -name '<deleted-id>*'
-```
-
-## F. Background & restore
-
-| # | Coverage | Scenario | Input | Expected |
-|---|---|---|---|---|
-| F1 | ⬜ | Detached run | Start sub-agent → `/session new` → switch back later | Sub-agent finishes in background; result visible |
-| F2 | ⬜ | WS disconnect/reconnect | Start sub-agent → close tab → reopen | Sub-agent finishes; reconnect shows report |
-| F3 | ⬜ | Server restart restore | Active session → restart server → reconnect | UIState restored from disk; conversation visible |
-| F4 | ⬜ | UIState flush on release | Start agent, immediately interrupt mid-stream | Disk file `.json` `messages` stays in sync with `rawMessages` (regression for releaseSession flushPersist) |
-
-## G. Loop & depth
-
-| # | Coverage | Scenario | Input | Expected |
-|---|---|---|---|---|
-| G1 | ⬜ | Ping-pong | "Play ping-pong with sleeper, 3 rounds" | Two-way query_session; LLM stops naturally |
-| G2 | ⬜ | 3-level chain | "Start test-agent, have it start another, depth 3" | 3-level parent chain; reports bubble up |
-| G3 | ⬜ | Cross-agent delegation | "Have test-agent ask sleeper its name" | test-agent → start_session(sleeper) |
-| G4 | ⬜ | Max nesting depth guard | Manually trigger 17-deep delegation chain | start_session returns `code 1` "Maximum nesting depth (16) reached" |
-
-## H. Context window & compact
-
-| # | Coverage | Scenario | Input | Expected |
-|---|---|---|---|---|
-| H1 | ⬜ | Auto-compact at threshold | Read large files until 80% (default `compress_at: 0.8`) | Auto-compact fires; tokens drop |
-| H2 | ⬜ | `general.compact.compress_at` honored | Settings → set 0.5 → start session, fill | Compact fires earlier (50% instead of 80%) |
-| H3 | 🟢 | Manual `/session compact` | Type `/session compact` or click TokenRing | `compact:progress` → `session:compacted` → `compact:done` |
-| H4 | ⬜ | Compact disabled while streaming | Click TokenRing while agent runs | Button disabled (isStreaming guard) |
-| H5 | ⬜ | Stop during compact | Click Stop during compact | Compact cancelled; messageLog unchanged |
-| H6 | ⬜ | **Message during compact drains after** | Send msg during compact | "Queued" notification → after compact ends, message **immediately runs** (regression for `endCompact` drain fix) |
-| H7 | ⬜ | Refresh after compact | Reload after compact | Message order correct; tool calls render in their turns |
-| H8 | ⬜ | Sub-agent context overflow recovery | Sub-agent reads huge files, hits overflow | Local truncation fallback fires + retry; no crash |
-| H9 | ⬜ | Rate-limit retry with backoff | Hammer fast messages | "Rate limited, retrying in Xs…" system msg; eventually succeeds |
-
-## I. Slash commands
-
-| # | Coverage | Scenario | Method | Expected |
-|---|---|---|---|---|
-| I1 | 🟢 | `/session new` | WS `session:clear` | Old session saved; new conversation empty; old in sidebar |
-| I2 | ⬜ | `/session new` during sub-agent | `/session new` while child running | Sub-agent continues background; new conversation independent |
-| I3 | 🟢 | `/session compact` | WS `command:compact` | Summary generated; older messages compressed |
-| I4 | ⬜ | `/session context` | UI command | Shows tokens, model, agent's available tools, system prompt |
-| I5 | ⬜ | `/session list` (full access) | UI command on web channel (admin-equivalent) | All workspace sessions visible with tags |
-| I6 | ⬜ | `/session list` (readonly/workspace user) | wechat/tg readonly user `/session list` | **Only own-prefix sessions** (regression for cross-user privacy) |
-| I7 | ⬜ | `/session switch` matches `/session list` indices | After I6, `/session switch 1` for readonly user | Switches to user's first own session |
-| I8 | ⬜ | `/session switch` to non-own blocked | readonly tries `/session switch` to a global admin session via crafted index | Returns `switch.readonly` rejection (defense-in-depth even though list filtered) |
-| I9 | ⬜ | All builtin `/help`/`/workspace`/`/session stop` reachable from each channel | Run each through web/tg/wx | Same response shape |
-| I10 | ⬜ | Server startup descriptor↔dispatch sanity check | Restart server | Throws if a builtin server descriptor is missing a dispatch case |
-
-## J. Session viewer & Debug
-
-| # | Coverage | Scenario | Input | Expected |
-|---|---|---|---|---|
-| J1 | ⬜ | Debug rendering | Multi-tool conversation | Debug: exchange layout, tool_call cards, usage badges |
-| J2 | ⬜ | Non-Debug filter | Same conversation | Only assistant text + inline tool cards |
-| J3 | ⬜ | Usage fields | After multi-tool | in / out / ctx / cache% / ttft / e2e / model shown |
-| J4 | ⬜ | System prompt view | Click Prompt button | Full system prompt rendered |
-| J5 | ⬜ | Parent/child prompts | Inspect parent then child sub-session | Both have Prompt buttons; each shows its own |
-
-## K. Sidebar & cleanup
-
-| # | Coverage | Scenario | Expected |
-|---|---|---|---|
-| K1 | ⬜ | Session tree render | Users' sessions with children show hierarchy |
-| K2 | ⬜ | Depth indentation | After G2: 3-level indentation correct |
-| K3 | ⬜ | Cascade delete | Delete main session → all sub-sessions + SQLite + files cleaned |
-| K4 | ⬜ | Sub-session preview | Click sub-session → full agent conversation |
-| K5 | ⬜ | Stopped badge | After D4: stopped session shows StopCircle |
-| K6 | ⬜ | Refresh button shows spinning animation | Click refresh on agent-sessions sidebar | RefreshCw rotates ≥350ms even on fast responses |
-| K7 | ⬜ | No spinner flash on bus-driven refresh | Delete session, observe refresh | Other consumers reconcile silently, no extra spinner stutter |
-
-## L. Settings UI (post-2026-05 schema refactor)
-
-| # | Coverage | Scenario | Expected |
-|---|---|---|---|
-| L1 | ⬜ | Sections grouped by declarer | Open Settings | "System" / "Model Providers" / "Agents" / "Skills" / "Orphans" navigation |
-| L2 | ⬜ | Provider yaml drives section | Add a new `models/myprov.yaml` with `secrets:` | Section appears for it after server restart |
-| L3 | ⬜ | Skill yaml drives section | Add `skills/X/config.yaml` with `params:` | Section visible |
-| L4 | ⬜ | Agent yaml drives section | Add `agents/X/agent-config.yaml` with `params:` | Section visible |
-| L5 | ⬜ | Field type renders right control | Schema int/float/boolean/enum/string fields | Number / toggle / dropdown / text inputs |
-| L6 | ⬜ | YAML scalar preserved on save | Save `compress_at: 0.8` then read settings.yaml | `compress_at: 0.8` (unquoted, real number) |
-| L7 | ⬜ | Mtime-cached config picks up edits live | Edit settings.yaml externally, send a chat | Server picks new value on next call without restart |
-| L8 | ⬜ | Secret masking on transport | Settings page secret field with real value | API returns `xx****yy`; env-var refs `<<KEY>>` not masked |
-| L9 | ⬜ | Source badges | Workspace scope, value comes from global | Badge "inherited from global" |
-| L10 | ⬜ | Orphan listing + remove | Insert junk namespace into settings.yaml | Appears in Orphans tab; Remove deletes it |
-| L11 | ⬜ | Schema default vs config.ts callsite consistency | Boot server | No `[config] default mismatch for "..."` warnings in log |
-| L12 | ⬜ | Missing env var keeps `<<NAME>>` literal | Set `kimi.secrets.api_key: <<UNSET_VAR>>`, chat with kimi | API call returns 401 with literal `<<UNSET_VAR>>` (regression for "no silent env fallback") |
-
-## M. Channel — wechat/telegram/web (multi-user)
-
-| # | Coverage | Scenario | Method | Expected |
-|---|---|---|---|---|
-| M1 | ⬜ | Inbound text | Send chat from each channel | Routes to its own prefix session |
-| M2 | ⬜ | Inbound image (telegram) | Send photo | Saved under `assets/telegram/inbound/`; agent describes |
-| M3 | ⬜ | Inbound image (wechat) | Send image via wechat-bot | Same: saved + described |
-| M4 | ⬜ | Inbound voice/audio (web) | Paste audio file | Saved under `assets/web/inbound/`; `[语音已保存]` marker |
-| M5 | ⬜ | `inferImageMime` magic byte detection | Various jpg/png/webp/gif | All correct after the dedupe to shared/media-store |
-| M6 | ⬜ | `resolveAccountWorkspace` after dedupe | Disable a workspace path on disk, send msg | All channels return early, no broken sm calls |
-| M7 | ⬜ | Two wechat users in same workspace, isolation | User A `/session list` doesn't show user B's session titles | Confirms post-fix `/session list` filtering |
-| M8 | ⬜ | Compact-during-tg-message drain | Trigger compact while tg user sends a message | Message processed right after compact ends (regression for endCompact) |
-| M9 | ⬜ | tg compaction status feedback | Send message during compact | "compacting" reply (parity with web/wechat) |
-
-## N. CLI / TUI
-
-The CLI ships an embedded agent loop and **bypasses the WS server** by design (lifestyle separation: CLI/TUI for offline scripts, WS for browser/admin). Tests here live separately.
-
-**Setup**:
-```bash
-# Build CLI (if not done)
-pnpm --filter @turmind/halo-cli build
-
-# Smoke
-node packages/cli/bin/halo.js --help
-```
-
-| # | Coverage | Scenario | Method | Expected |
-|---|---|---|---|---|
-| N1 | ⬜ | CLI prints version | `halo --version` | Matches package.json |
-| N2 | ⬜ | TUI launch | `halo tui` | Starts, shows agent picker, accepts input |
-| N3 | ⬜ | TUI single turn | `halo tui` → say "what is 2+2" | Reply rendered with usage badge `think medium` (or model default) |
-| N4 | ⬜ | TUI session id has `cli_` prefix | After N3 | `~/.halo/sessions/default/cli_<...>.json` exists |
-| N7 | ⬜ | TUI `/session new` | `/session new` | New session, default-scoped |
-| N8 | ⬜ | TUI `/session stop` | "sleep 30 then say hi" → Ctrl-C or `/session stop` | Clean abort |
-| N9 | ⬜ | TUI `/session compact` | Build up 8+ messages → `/session compact` | Compact runs; ctx tokens drop in next prompt |
-| N10 | ⬜ | TUI tool call rendering | Ask file_read | Tool card renders inline in tui |
-| N11 | ⬜ | TUI thinking display (manual mode) | Switch to Haiku 4.5 agent | Usage badge shows `think 8192` (or whatever budget_tokens), not `think medium` |
-| N12 | ⬜ | TUI session resume across restart | N3 then exit + restart `halo tui --resume` (or whatever the resume cmd is) | History visible |
-| N13 | ⬜ | TUI graceful interrupt on new message | "sleep 10 then say hi" → press Esc + new message | New message handled after current tool, no abort error |
-| N14 | ⬜ | CLI `halo server start` single-instance lock | Start two instances | Second exits with "another instance running" |
-| N15 | ⬜ | CLI `halo setup` rewrites config.yaml | Run with `--password new-pw` | `~/.halo/secrets/config.yaml` has new scrypt hash + new jwt_secret |
-| N16 | ⬜ | TUI sandbox enforced | `tui --access-level readonly` (if flag exists) → try `file_write` | Tool refused / sandbox blocks |
-| N17 | ⬜ | TUI uses **same** SessionManager.persistSessionFile path | Run a session in TUI, then start the WS server pointing at same workspace | UI shows the TUI session in sidebar (single source of truth) |
-
-> **Note on TUI commands**: I've listed commands by name above (`/session new`, `/session stop`, `/session compact`, etc.). If your TUI uses different keybindings instead, swap the trigger column accordingly — the **expected behavior** is what we're checking.
-
----
-
-## Detailed steps & verification (selected cases)
-
-### D1 — Basic delegation
-
-**Steps:**
-1. Send: "Ask sleeper to sleep 10s then tell its height"
-2. Default agent calls `start_session` (agent: sleeper)
-3. Sleeper completes; `[Message from ...]` auto-delivered
-4. Default agent activates, reports
-
-**Verify:**
-- [ ] No unnecessary polling (no extra `session_list` / `get_session_output`)
-- [ ] Single final reply with sleeper's height
-- [ ] Sidebar: user session shows 1 sub-session
-
-**Data:**
-- [ ] SQLite `agent_sessions`: child row has `parent_id = user's session id`
-- [ ] `.halo/sessions/sleeper/`: JSON file exists
-- [ ] JSON: `parentSessionId` = user's id; `rawMessages` non-empty
-
-### D10 — Skill namespace whitelist (regression)
-
-**Steps:**
-1. Configure `tavily-web-search` skill with a real `api_key` in Settings.
-2. Edit `test-agent/agent.yaml` to include `skills: [tavily-web-search]` — do NOT include it.
-3. Have the agent run: `shell_exec('curl -H "Authorization: Bearer {{tavily-web-search.params.api_key}}" https://api.tavily.com/search')`
-
-**Expected:**
-- Server log: `[workspace-tools] {{tavily-web-search.params.api_key}} rejected — namespace not in allowed list`
-- The literal `{{tavily-web-search.params.api_key}}` is sent to Tavily → 401.
-- Re-add `tavily-web-search` to `skills:` list, retry → real key resolves → 200.
-
-### E7 — REST DELETE goes through SessionManager
-
-**Steps:**
-1. Open a session in the browser, send a couple of messages.
-2. Without closing the browser tab, run from another shell:
-   `curl -b cookie.txt -X DELETE "http://localhost:9527/api/sessions/logs/<sid>?projectId=/path/to/test-workspace"`
-3. Send another message in the browser.
-
-**Expected:**
-- Server log: `[SessionManager] Deleted N sessions from SQLite (root: <sid>)`
-- The next `chat:complete`-driven save in WS handler is dropped (tombstone hit).
-- File on disk stays gone (don't see the file size grow back).
-
-### H6 — Message during compact
-
-**Steps:**
-1. Build a session with 9+ messages.
-2. Trigger `/session compact`.
-3. While `compact:progress` event is in flight, send a new message.
-4. Observe the order of events.
-
-**Expected:**
-- WS receives `chat:queued` (message stashed in `pendingUserMessages`).
-- `session:compacted` arrives.
-- **Within 1–2 seconds**, the queued message starts running automatically (no need to send anything else).
-- This is the regression test for the recent `endCompact` drain fix — before the fix, the queued message would sit indefinitely.
-
----
-
-## Automated regression script
-
-Path: `packages/server/tests/test-session-system.mjs`
-
-What it covers (🟢 cells in the tables above): A1–A6 (tool calls), B1, C4, D1, D7, E6, E7, I1, I3.
-
-Run:
-```bash
-HALO_TEST_PASSWORD=$HALO_PASSWORD \
-HALO_TEST_PROJECT=/path/to/test-workspace \
-node packages/server/tests/test-session-system.mjs
-```
-
-Exits 0 on all-green. The script does **not** start/stop a server — bring your own. It logs in via REST, opens a WS, runs the suite, prints a final pass/fail count.
-
-**What's NOT in the regression script** — handled by the manual table above:
-- B2–B7 (UI-driven flows: stop+resume, refresh during tool, agent dropdown lock)
-- C5–C12 (caching variants, thinking effort tiers, image filtering)
-- D2–D6, D8–D10 (continued conversations, parallel delegation, namespace whitelist)
-- E8–E10 (resurrection guard, settings reset visuals)
-- F, G, H (background, depth chains, full compact flow)
-- L (settings UI rendering)
-- M (channel inbound)
-- N (CLI/TUI)
-
-If you change session lifecycle code, **at minimum**: re-run the regression script + walk D, E, F, H by hand.
-
----
-
-## Quick data-verification pipeline
-
-After any test case that should produce / mutate session state:
-
-```bash
-PROJECT=/path/to/test-workspace
-
-# SQLite — agent_sessions index
-sqlite3 $PROJECT/.halo/halo.db \
-  "SELECT id, parent_id, agent_id, stopped_at IS NOT NULL AS stopped, archived_at IS NOT NULL AS archived
-   FROM agent_sessions ORDER BY created_at DESC LIMIT 10;"
-
-# Session files (all agents)
-for dir in $PROJECT/.halo/sessions/*/; do
-  echo "=== $(basename $dir) ==="
-  ls -lt "$dir" 2>/dev/null | head -5
-done
-
-# parentSessionId integrity check
-for f in $PROJECT/.halo/sessions/*/*.json; do
-  python3 -c "
-import json
-d = json.load(open('$f'))
-parent = d.get('parentSessionId', 'NONE')
-print(f\"{d['agentId']}/{d['id']}  parent={parent}\")
-" 2>/dev/null
-done
-
-# Settings schema sanity
-curl -sb /tmp/halo-cookie.txt http://localhost:9527/api/settings/schema | \
-  python3 -c "import sys,json; r=json.load(sys.stdin); print(f'sections={len(r[\"sections\"])} orphans={len(r[\"orphans\"])}')"
-```