- skills/multi-angle-orchestration/SKILL.md: SOP for single-source fan-out
(delegate_task leaf workers, NO external AIs) of >=5 angles + converge.
- brain-core: ingest orchestration-runs/*.json as source 'orchestration'
so past runs are searchable in Brain Search.
- server.py: POST /api/orchestrate (optional headless fan-out of the 3
built-in agents) + GET /api/orchestrate/runs; persist runs to
data/orchestration-runs/<id>.json and upsert to brain.
- dashboard/pages/orchestration.js: launch runs + review past runs.
- skills/multi-agent-run/SKILL.md + api helpers + nav entry.
- start.ps1 fallback 8080 -> 8081 (reads data/settings.json which is now 8081)
- server.py argparse --port default 8080 -> 8081 (bare server.py lands on 8081)
- (desktop Agentic OS.url already points at 127.0.0.1:8081; settings.json
change kept local since it's git-ignored with api_keys)
Single consistent port story across WSL start.sh, Windows start.ps1,
bare server.py, and the desktop shortcut.
1) Brain Search page (dashboard/pages/brain-search.js): mirrors Agent
Insights; queries /api/brain-index/search with source filter + live
index stats; nav entry + PAGE_TITLES added.
2) Scheduled auto-ingest: scheduler gains endpoint-style jobs
(run_endpoint_via_api) so non-skill maintenance tasks can be cron'd;
new job brain-index-ingest-job.json re-indexes every 30 min.
3) Agents auto-upsert learnings: run_skill wrap-up now calls
record_brain_learning() so each skill run is immediately searchable
in the unified index (best-effort, never breaks the run).
Fixed a regression where AGENT_STATS_FILE def was dropped during the
brain-learning helper insertion (would have broken all skill runs).
Verified live: brain search returns ranked results; run_skill wrap-up
upserts a searchable doc; scheduler imports + job validates.
- brain-core/brain_index.py: SQLite+FTS5 index over brain/** markdown,
skills/*/learnings.md, and data/chat-history.json. Idempotent upsert
keyed by (source, path, agent, #msg) so re-ingest never duplicates.
- brain-cli.py: ingest / stats / search CLI.
- server.py: /api/brain-index/{ingest,upsert,search,stats} so agents can
WRITE to the brain (upsert) and you/agents can QUERY it (search).
- .gitignore: exclude the regenerated *.db.
Verified live: ingest 66 docs (16 brain, 32 chat, 18 skill-learnings),
cross-source FTS search, and agent upsert is immediately searchable.
New /api/agent-insights endpoint aggregates real data (chat-history,
cost-history, agent-time, agent-registry, router-keywords) to show:
- which agent you chat with most (user turns)
- which agent spends the most time on tasks (derived agent-time)
- which agent is best suited for what (registry desc + router keywords + roles)
- which models you use most per agent (cost-history)
Dashboard page (agent-insights.js) with stat cards, two Chart.js bar
charts, suite cards, models-per-agent table, and full per-agent
breakdown. Verified live in browser against real data.
- check_agent got mangled during conflict resolution (main's builtin logic
stitched to branch's agent-referencing return -> NameError). Replaced with
the correct dynamic, registry-based implementation (binary/oauth_file/http/
custom) including the shlex injection-safe custom check.
- Removed the redundant standalone WebSocket terminal server on port 8082;
the in-app /ws/terminal PtySession is canonical.
- Verified live: status/agents/integrations/agent-time all 200; double-launch
guard and shell-injection fix both hold.
- check_agent() custom check_type: replace subprocess shell=True with
shlex.split + shell=False (prevents agent-registry config from injecting
arbitrary shell). Verified malicious '; touch' no longer executes.
- Add startup port guard in main: probe API port before uvicorn.run;
exit 1 with clear message if already bound (stops the double-instance
collision that caused phantom 'register does not persist' bugs).
- start_terminal_server (8082): pre-bind probe + try/except so a taken
port logs a warning instead of crashing the whole process.
* Fix path traversal in /api/backup/restore
Validate the restore filename stays within backups/ and refuse tar members
(and symlinks) that escape the extraction root (CVE-2007-4559 class), using
tarfile's data filter.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* Resolve backup filename from directory listing (satisfy CodeQL path-injection)
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: zumayaaustin <zumayaaustin@gmail.com>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Resolve install.ps1 conflict by adopting main's complete installer (Resolve-Python,
agent checks, desktop shortcut, browser auto-open) and folding in this branch's
winget install hints and optional agent-CLI reminders. Clean up README duplicate
Python/Node prerequisite rows and duplicate Windows Quick Start left over from the
earlier merge.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Addresses findings from Devin and Codex reviews on the merged PR #10:
- CRITICAL: /ws/terminal accepted connections from any origin - Starlette's
CORSMiddleware does not protect WebSocket handshakes, so any webpage could
open a socket to the dashboard's terminal and get an interactive shell on
the user's machine. Now validates the Origin header against the same
allowed-origins list used for CORS before accepting.
- PtySession.close() sent SIGKILL to the shell's PID but never reaped it via
os.waitpid(), leaking a zombie process per closed terminal session.
- hermes_available() ran a real subprocess (possibly bridged through WSL) on
every /api/status poll, which the dashboard hits every 15s. Added a 60s
TTL cache.
- create_skill() now also creates learnings.md and the context/ directory,
matching the standard skill template (_template/) instead of only
writing SKILL.md.
- The '+ New Skill' button stayed visible in the Skills Hub detail view
since only its sibling filter input was hidden; both now live under a
shared #skillActions container that's hidden/shown together.
Verified: malicious/missing-origin WebSocket connections are rejected at
the handshake (HTTP 403) before any shell spawns; a valid dashboard origin
still connects and works; closing a session leaves no zombie/orphaned
process; the hermes availability cache avoids repeat subprocess spawns.
Validate the restore filename stays within backups/ and refuse tar members
(and symlinks) that escape the extraction root (CVE-2007-4559 class), using
tarfile's data filter.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
hermes_cli_args() previously trusted any 'hermes' found on native PATH
without checking what it actually was. Windows machines can have an
unrelated tool also named 'hermes' (softwarepub/HERMES, an academic
software-publication tool with harvest/process/curate/deposit
subcommands - confirmed to be what was actually on this machine's
PATH), which would silently get used instead of the real NousResearch
agent installed in WSL, producing the misleading 'Hermes needs setup'
message.
Now check that a native 'hermes' actually exposes the agent's 'chat'
subcommand before using it directly, falling back to the WSL bridge
otherwise.
Two real gaps: the Skills Hub had no way to create a new skill (with
a SKILL.md) at all - the only 'Install' flow was the Plugin registry,
which just records a name in a JSON file, not an actual skill folder.
And the Context Files panel was read-only, just listing filenames
with no way to view, edit, add, or delete their contents.
Adds POST /api/skills (create, with SKILL.md content), PUT
/api/skills/{name} (edit SKILL.md), and GET/PUT/DELETE
/api/skills/{name}/context/{filename} for context files - all
validated through the same regex-allowlist + resolved-path
containment pattern already used for kanban tasks. Dashboard gets a
'+ New Skill' button, an editable SKILL.md view, and per-file
edit/delete plus 'Add File' in the Context Files panel.
The dashboard runs as a native Windows process, but Hermes' official
installer is Bash-only and typically only gets set up inside WSL -
a plain PATH lookup for 'hermes' on Windows will never find it there.
Add hermes_cli_args(), which checks the native PATH first (so Mac/
Linux/WSL-native setups are unaffected) and falls back to routing
through 'wsl -e bash -lc' (a login shell, so PATH additions like
uv's ~/.local/bin are sourced) only when hermes isn't found natively
but wsl.exe is available. Wire both the chat/dispatch invocation and
the agent-health check through it, replacing the plain shutil.which
check that always reported Hermes offline in this setup.
The Kanban detail modal's Delete button called api.deleteKanbanTask(),
which didn't exist on the client, and there was no backend route for
it either - clicking Delete just threw 'api.deleteKanbanTask is not
a function'. Add both the client method and the backend endpoint.
The previous Terminal ran one command at a time via subprocess.run
and returned its output - it couldn't run interactive programs
(colors, live input, TUIs like gemini's chat/auth flow), which is
what a terminal actually needs to do.
Backend: new /ws/terminal WebSocket endpoint spawns a real shell
attached to a pseudo-terminal (stdlib pty on POSIX, pywinpty/ConPTY
on Windows) and streams raw I/O bidirectionally, with resize support.
Replaces the old POST /api/terminal/run and GET /api/terminal/session
endpoints entirely.
Frontend: terminal.js now loads xterm.js + the fit addon from CDN
and renders a real terminal emulator wired to the WebSocket, instead
of a scrollback div with a single input line.
Verified on this Linux sandbox via raw WebSocket tests: shell spawns
correctly, commands execute and echo real output, resize propagates
to the PTY (confirmed via ), and closing the connection
cleanly kills the shell process with no orphans (interactive bash
ignores SIGTERM by default, so cleanup uses SIGKILL). Could not
visually verify the xterm.js browser rendering in this sandbox since
its egress policy blocks the CDN (cdn.jsdelivr.net) outright - same
CDN this app already uses for chart.js, so expected to work on a
normal machine; please confirm on Windows.
subprocess.run(..., shell=True) uses cmd.exe on Windows by default,
which doesn't understand PowerShell syntax like $env:VAR or
$env:USERPROFILE - commands using it failed with 'cannot find the
file specified' since cmd took it as a literal filename. Invoke
powershell.exe explicitly on Windows instead; POSIX behavior is
unchanged.
The regex allow-list alone wasn't enough for CodeQL's path-injection
sanitizer recognition. Resolve the candidate path and verify it's
still a direct child of the resolved kanban directory before
returning it, which is the pattern CodeQL's py/path-injection query
recognizes as clearing taint.
CodeQL flagged path-traversal risk (uncontrolled data used in path
expression) across the kanban endpoints: task_id/parent_id/child_id
path parameters were spliced directly into KANBAN_DIR paths with no
validation. Task ids are always server-generated 8-char hex strings,
so add kanban_task_path() which validates against that shape and
raises 400 otherwise, and route every kanban file path through it.
Assigning a task to opencode, hermes, or gemini now actually runs it:
creating or PATCHing a task with one of those assignees moves it to
in_progress and hands the title+body to execute_agent() in a
background thread. On completion the agent's response is appended as
a task comment and the task is marked done (success) or blocked
(timeout/error), matching the existing block/complete state machine.
Adds POST /api/kanban/tasks/{id}/dispatch for manual dispatch/retry,
and makes the previously-stubbed POST /api/kanban/dispatch actually
scan todo/ready tasks with an agent assignee and dispatch each one.
Dashboard: the assignee field is now a select of the three agents,
the task detail modal shows an activity log of past agent runs and a
Dispatch button, and polls every 3s while a task is in_progress. Also
fixed a pre-existing bug where the detail modal read a nonexistent
kanbanData.tasks field (the board API only returns columns), so
clicking a card silently did nothing before this fix.
Adds a real shell terminal to the dashboard sidebar: POST
/api/terminal/run executes a command via subprocess in a
server-tracked working directory (with cd support), and GET
/api/terminal/session returns the current cwd. The frontend renders
a scrollback panel with command history (up/down arrows) styled to
match the existing chat UI.
Local-only power feature: it executes arbitrary shell commands, same
trust model as the existing agent CLIs the dashboard already shells
out to.
New features:
- Kanban Board: Visual task management with drag-and-drop, filter, priority, block/unblock
- Goals: Project targets with progress tracking, auto-sync to brain/active-projects.md
- Journal: Daily entries stored as brain/journal/YYYY-MM-DD.md with search
- Agent Health: Real-time monitoring of 3 agents (opencode/hermes/gemini)
- Smart Router: Keyword-based task routing with confidence scoring
- Learning Analytics: Skill evaluation scores and performance trends
- Session Replay: Browse and replay past opencode sessions
Technical changes:
- 30 new API endpoints across all 7 features
- api.patch() method added for PATCH support
- UI modernization: glass morphism cards, gradients, glow effects, skeleton loaders, empty states
- New CSS section (v0.2.0) with kanban board, goal cards, journal entries, agent health cards, smart router suggestions, chart cards, session messages
- 7 new sidebar nav items under Workflow and Monitoring sections
- All changes are additive (zero breaking changes to existing 13 pages / 28 endpoints)
- Full smoke test verified: every endpoint returns correct data
Austin
zumayaaustin-creator
Claude
devin-ai-integration[bot]
modimihir07