How to enable and configure AgentWatch

In brief: Enable AgentWatch in Studio Settings, choose cadence and agents, clear the QA baseline gate, then review and approve proposed story and scenario updates.

Enable AgentWatch

1. Open TestChimp Studio and select the project whose local folders are mapped.
2. Open Settings → AgentWatch.
3. Check Enable AgentWatch.

Studio starts the local AgentsView daemon (provisioned and pinned with Studio). The first start may take a few minutes while sessions sync. Daemon status appears at the bottom of the settings pane—use Refresh status or Restart daemon if needed.

Settings → AgentWatch — enable, cadence, actions, and agents to watch

Settings reference

Setting	What it does	Default
Enable AgentWatch	Starts/stops the AgentsView daemon and BAU sync loop	Off
Sync cadence (hours)	How often Studio polls mapped agent sessions for new product decisions	1 (minimum 1)
Daemon watch port	Preferred local port for the AgentsView HTTP API	27451
Allowed actions	Whether plans may add / update stories and add / update scenarios	All on
Agents to watch	Which coding-agent providers to include (Cursor, Claude Code, Codex, and many more)	All supported agents

Settings are global on the Studio install. Eligibility (baseline, folder mapping) is evaluated per project on each cadence tick.

Agents to watch

Uncheck providers you do not want indexed. Supported providers come from the pinned AgentsView release (Cursor, Claude Code, Codex, Copilot, Gemini, Windsurf, OpenHands, and dozens more). Empty selection means “all supported.”

Allowed actions

Tighten AgentWatch if you only want scenario upkeep, or only story adds:

• Add stories / Update stories
• Add scenarios / Update scenarios

Disabled actions are not proposed in the generated plan.

Baseline

Day-to-day AgentWatch sync (checkpoints, Jev filtering, plan authoring) runs for a project only after a QA baseline exists:

• ≥ 10 stories or ≥ 20 scenarios (not both required)

Until then, enabling AgentWatch does not process that project’s sessions.

Getting a baseline

Path	When to use it
Import from Jira / Linear / TestRail (Plans tab)	Recommended for an existing backlog
Bootstrap from local specs	Studio can author an initial plan from local .md specs / agent-authored plans when no baseline exists and a folder is mapped
Thin corpus after bootstrap execute	If counts remain below threshold after you execute a bootstrap plan, BAU sync is still allowed

Import and bootstrap details live in Studio’s enable flow and the Plans → Import path (“Bootstrap from local specs” on desktop).

What happens on each sync tick

Once enabled and baseline-ready for the selected / mapped projects:

1. Studio lists recent agent sessions (activity in roughly the last 90 days).
2. Sessions are mapped to a TestChimp project via workspace path / cwd affinity with your mapped folders.
3. Ongoing chats are skipped until quiet (heuristic: last message older than ~5 minutes), unless a large unprocessed backlog needs draining.
4. New chat content since the last checkpoint is chunked and scored.
5. High-signal sessions are batched into an AgentWatch plan (or an existing pending plan is updated).
6. You get a Studio notification when a plan is ready.

At most one pending AgentWatch plan per project on a Studio install. Approve or reject before a new one is authored for that project.

Review and approve

1. Open the pending plan from the Studio notification (or your workflow plans under the AgentWatch / autocontext path).
2. Review proposed story and scenario adds/updates.
3. Approve to execute, or reject to dismiss.
4. Raise a PR with the plans/ changes so the rest of the team—and downstream workflows—see the updated QA surface.

How filtering works

Agent chats are long. Most chunks are implementation chatter, not durable product guidance. AgentWatch therefore:

1. Queries your project’s shared semantic embedding space for story / scenario neighbours closest to the chat chunk (Semantic Canvas is the visual surface of that same space).
2. Sends the chat summary + chunk + neighbour titles/descriptions to Jev—TypeSafe’s System One decision model—to score whether the chunk contains decisions that should hit the QA surface, which neighbours are affected, and whether a new story/scenario should be authored.
3. Only filtered hits go to an LLM to draft the update plan.

Jev’s tighter context budget (~32K for state + longest question) is why neighbour retrieval matters: stuffing every story and scenario into the filter is infeasible; closest neighbours keep the score accurate and cheap enough to run across every developer’s agents.

Related tools (same problem space)

People searching for agent-session memory or history tooling often land on adjacent products. AgentWatch’s job is different—turn agent decisions into governed QA plans—but these are useful neighbours if you are comparing approaches:

Tool	Role
AgentsView	Local-first multi-agent session index (powers AgentWatch’s watch layer)
aise (aise)	Fast local search / MCP over coding-agent sessions
Callimachus	Local hybrid search + recall over agent threads
Prism	Persistent session / knowledge memory for coding agents

Troubleshooting

Symptom	What to check
Daemon not running	Settings → AgentWatch → enable; Restart daemon; confirm port is free
No plans appearing	Baseline met? Folder mapped? Agents selected? Cadence waited? Sessions older than ~5 minutes since last message?
Plans never update a project	Confirm session cwd / workspace path falls under that project’s mapped folders
Too many / too few proposals	Adjust Allowed actions; keep cadence ≥ 1 hour; remember Jev prefers false negatives over noisy updates

Learn more

• AgentWatch overview
• Agent Sessions Explorer
• TestChimp Studio
• Semantic Canvas
• Meeting Bots