forge-connector

Forge Connector

Builds a graph:connector Forge app that ingests external data into Atlassian's Teamwork Graph so it appears in Rovo Search and Rovo Chat.

Critical Rules

1. Must install in Jira — Apps using Teamwork Graph modules must be installed on a Jira site. Confluence-only installs will not work.
2. Never ask for credentials in chat — Direct users to run forge login in their own terminal.
3. Always run the scaffold script yourself — Do not only give manual instructions; run scripts/scaffold_connector.py to generate the boilerplate.
4. Always ask the user for their Atlassian site URL when install is needed — never discover or guess it.
5. Atlassian deletes data on disconnect — When action = 'DELETED', the app only needs to clean up local state; Atlassian removes the Teamwork Graph data automatically.
6. Handler arguments are passed directly — Forge passes the request object as the first argument to handlers, NOT nested under event.payload. Config values are at request.configProperties, NOT event.payload.config. This is the most common source of TypeError: Cannot destructure property of undefined errors.
7. Use @forge/kvs for storage — Import kvs from @forge/kvs. Do NOT use @forge/storage — its storage export is undefined at runtime in connector functions.
8. Use graph named export from @forge/teamwork-graph — The correct import is const { graph } = require('@forge/teamwork-graph'). Call graph.setObjects({ objects, connectionId }). Do NOT import setObjects as a named export directly.
9. validateConnectionHandler must return { success, message } — Do NOT throw an Error. Return { success: false, message: '...' } to reject, { success: true } to accept.
10. function declarations belong under modules — In manifest.yml, function: is a key under modules:, not a top-level key. Placing it at the top level causes a lint error.
11. formConfiguration uses form array with type: header — Do NOT use fields: or beforeYouBegin:. The correct format uses form: [{ key, type: header, title, description, properties: [...] }].
12. Scopes are read/write/delete:object:jira — Use read:object:jira, write:object:jira, delete:object:jira. The scopes read:graph:teamwork and write:graph:teamwork are invalid and will fail forge lint.
13. Set ATL_FORGE_ATTRIBUTION_SKILL_NAME=forge-connector on forge commands run for this skill — prefix forge invocations with this env var: ones you run in the shell (e.g. forge lint, forge logs, forge deploy) and the interactive forge create command you hand the user as a fallback. The bundled scripts set it automatically; other commands shown in this skill omit it for brevity — add it when you run them. The only exclusions are forge login and forge tunnel (user-run auth / live-dev commands).
14. Never ignore setObjects / setUsers / setGroups responses — Do NOT treat ingestion calls as fire-and-forget. Always check both response.success and response.results?.rejected. A call can partially fail: some objects are rejected while others are accepted, and response.success can be true even when rejected.length > 0.
15. Keep access control IDs consistent — When using USER or GROUP principals in permissions, the id value must exactly match the externalId used in the corresponding graph.setUsers() / graph.setGroups() call. Mismatched IDs (e.g. user-42 vs accountId:user-42) silently make objects invisible to users who should have access.
16. Use orchestration for refresh scheduling — When the user wants periodic re-ingestion, use graph.scheduleOrUpdateTask + a taskRunner handler (the platform-managed approach) instead of a scheduledTrigger. Orchestration handles cadence, retry, and fan-out natively. Only fall back to scheduledTrigger if the user explicitly needs a lightweight cron-style trigger without platform task tracking.
17. taskId must be a valid UUID — Both scheduleOrUpdateTask and scheduleChildTask require task.taskId to be a UUID string. Use uuid() (v4) for root tasks and uuidv5('<scanId>:<parentTaskId>:<itemId>', APP_NAMESPACE) for child tasks. Persist root task IDs in KVS so re-running onConnectionChange reuses the same ID (making it an update, not a competing schedule).
18. Write TaskInfo to KVS before scheduling; never delete it on completion — Always write the TaskInfo record to KVS keyed by taskId before calling scheduleOrUpdateTask or scheduleChildTask. On success, set completed: true on the record — do NOT delete it. Deleting causes duplicate deliveries to report ENTITY_NOT_FOUND and overwrite the earlier success.
19. Branch on response.error for orchestration responses — scheduleOrUpdateTask and scheduleChildTask return { status: 'ACCEPTED' } on success, not { success: true }. Check if (response.error) to detect failures — do NOT check if (!response.success).

MCP Prerequisites

Agent Workflow — Complete Steps 0–7 in Order

Step 0: Prerequisites

Check Node.js (node -v, requires 22+), Forge CLI (forge --version), and login (forge whoami). Install missing tools:

npm install -g @forge/cli

Tell the user to run forge login in their terminal if not authenticated.

Step 1: Discover Developer Spaces

Note: forge developer-spaces list does NOT exist in Forge CLI 12.x. You cannot list developer spaces non-interactively.

forge create requires an interactive TTY to select a developer space. Ask the user to run it themselves:

Tell the user:
  cd <parent-directory>
  ATL_FORGE_ATTRIBUTION_SKILL_NAME=forge-connector forge create --template blank <app-name>

  When prompted, select a Developer Space and let it complete.
  Come back when done.

The --dev-space-id flag in the scaffold script is optional and can be omitted — the script has been updated to skip it when not provided.

Step 1.5: Discover Data & Map to Object Types

Do this before scaffolding. Ask the user the following questions to determine the correct Teamwork Graph object type(s). Do not assume or default to atlassian:document.

Questions to ask the user

1. What external system or tool are you connecting? e.g. Google Drive, ServiceNow, Salesforce, GitHub, Confluence, Slack, Figma, Zendesk

2. What kind of content do you want to make searchable in Rovo? Prompt with examples to help them identify it:

   • Files, pages, wiki articles, reports, PDFs → likely atlassian:document
   • Tasks, tickets, issues, bugs, stories → likely atlassian:work-item
   • Chat messages, emails, comments → likely atlassian:message or atlassian:comment
   • Projects, workspaces, boards → likely atlassian:project
   • Code repositories → likely atlassian:repository
   • Pull requests / merge requests → likely atlassian:pull-request
   • Git commits → likely atlassian:commit
   • Design files (Figma, Sketch) → likely atlassian:design
   • Video recordings → likely atlassian:video
   • Calendar events, meetings → likely atlassian:calendar-event
   • Threads, channels → likely atlassian:conversation
   • Customer accounts or organisations → likely atlassian:customer-organization
   • Team spaces or org units → likely atlassian:space

3. Is the content a single type or a mix? If mixed (e.g. a project management tool with tasks and documents), plan to ingest each as its own object type. The scaffold supports one primary type — you can add more objectTypes entries in manifest.yml later.

4. Does the admin need to supply credentials (API key, URL, OAuth token) to connect? Yes → use --has-form-config in the scaffold command. No (data comes entirely from within Atlassian) → omit the flag.

5. How often does the source data change? This determines the sync cadence. Use orchestration (scheduleOrUpdateTask + taskRunner) for all platform-managed refresh — it handles cadence, retry, and fan-out natively. Only use the legacy scheduledTrigger approach if you explicitly need a lightweight cron with no task tracking.

   Frequency	Orchestration interval
   Hourly	scheduleInterval: { value: 1, timeUnit: 'hour' }
   Daily	scheduleInterval: { value: 1, timeUnit: 'day' }
   Every N minutes (1–60)	scheduleInterval: { value: N, timeUnit: 'minute' }
   Static / one-off	No scheduled refresh needed — ingest once in onConnectionChange

   Record the chosen interval before proceeding to Step 2.

6. Who should be able to see the ingested content in Rovo Search? This determines the permissions.accessControls on each object. Ask:

   • "Is all this content publicly accessible, or does the source system restrict who can see what?"
   • "Do you want Rovo Search results to respect those source-system permissions?"

   Map the answer to the correct principal model:

   Source system access model	accessControls to use
   Publicly accessible, no restrictions	principals: [{ type: 'EVERYONE' }]
   Specific named users have access	principals: [{ type: 'user', id: '<atlassian-account-id>' }] — one entry per user
   Team or group based (e.g. Confluence space, Google Workspace group)	principals: [{ type: 'group', id: '<group-id>' }] — one entry per group
   Private / owner only	single user principal with the owner's Atlassian account ID
   Mixed (per-object ACLs from the source)	fetch ACLs per item during ingestion and map each to a user or group principal

   Do NOT default to EVERYONE unless the user explicitly confirms content is publicly accessible. Using EVERYONE on restricted content leaks data to users who shouldn't see it in Rovo Search.

   Record the chosen permission model before proceeding to Step 2. Reference it when writing the setObjects call in Step 3.

Mapping decision

Based on the answers, select the best-fit type from the Object Types table below. Only fall back to atlassian:document if the content genuinely has no better match (e.g. arbitrary file attachments). For types marked ❌ in the "Indexed in Rovo" column (atlassian:build, atlassian:deployment, atlassian:test), warn the user that those objects will not appear in Rovo Search or Rovo Chat.

Record the chosen object type(s) and permission model before proceeding to Step 2.

Step 2: Scaffold the Connector App

Run from the skill directory (the directory containing this SKILL.md). Replace <object-type> with the type determined in Step 1.5. --dev-space-id is optional:

python3 -m

Add --dev-space-id <id> only if you have the ID from a previous step.

Object type — use the type chosen in Step 1.5. Do NOT default to atlassian:document without first completing the discovery questions above.

Form config flag — add --has-form-config if the admin must provide API credentials or connection details (determined in Step 1.5 question 4). Omit it for apps that operate entirely within Atlassian (no external credentials needed).

If scaffold fails because forge create needs a TTY: The scaffold script will print a manual fallback command. Have the user run forge create interactively, then continue from Step 3 — the scaffold script only needs to write manifest.yml and src/index.js after the directory exists.

Step 3: Customize the Generated Code

After scaffolding (or after the user runs forge create interactively):

The blank template generates src/index.js (JavaScript, not TypeScript). Edit it to add your API calls. The scaffold generates working handler skeletons; fill in your business logic.

Key files to edit

setObjects — ingest data into Teamwork Graph

Use the graph named export — do NOT destructure setObjects directly:

• Max 100 objects per call — batch large datasets with a loop
• id must be unique per connectionId
• connectionId is required in every graph.setObjects() call

Document Categories (for atlassian:document.type.category)

getObjectByExternalId — look up a single object

Step 4: Deploy and Install

You MUST run the deploy script — do not only give the user manual forge deploy commands.

The deploy script lives in the forge-app-builder skill, not in this skill. Derive its directory from the path of this SKILL.md: go up two levels (skills/forge-connector/ → skills/) then into forge-app-builder/. Run all commands below from that directory.

Step 5: Connect via Atlassian Administration

After deployment, tell the user to:

1. Go to Atlassian Administration → Apps → [site] → Connected apps
2. Find the app → View app details → Connections tab
3. Click Connect under the connector
4. Fill in any configuration fields (if formConfiguration was defined)
5. Click Connect — this triggers onConnectionChange with action: CREATED and starts data ingestion

Step 6: Monitor with forge tunnel

Use forge tunnel during development to stream live logs directly to your terminal as the connector functions execute. This is the fastest way to catch errors in onConnectionChangeHandler, validateConnectionHandler, and setObjects calls without waiting for forge logs.

Tell the user to run this in their own terminal (it requires an interactive session):

With the tunnel active, any invocation of the connector functions (e.g. clicking "Connect" in Atlassian Admin, or triggering a scheduled re-ingestion) will stream output immediately. Look for:

• [connector] Fetched N items — confirms fetchExternalData() ran
• [connector] Batch 1: N accepted, 0 rejected — confirms setObjects succeeded
• Any uncaught errors or thrown exceptions from validateConnectionHandler

If the tunnel is not running, use forge logs instead to inspect past invocations:

Tunnel vs logs — when to use which:

Note: forge tunnel must be run by the user in an interactive terminal — do not attempt to run it via the agent.

Step 7: End-to-End Verification (optional)

Before running any checks, ask the user:

"Would you like to run end-to-end verification checks before deploying to production? This confirms the connection, ingestion, Rovo Search visibility, and permission boundaries are all working correctly."

If the user says no or wants to skip, move on — do not run or describe the checks. If the user says yes, work through every check below in order.

Check 1 — Connection established

In Atlassian Administration → Apps → Connected apps, the connector should show status Connected. If it shows an error or pending state, go back to Step 6 and inspect forge tunnel or forge logs output.

Check 2 — validateConnection passed (if configured)

If the app has a validateConnection function, confirm the admin saw a success message when clicking Connect. If not, check logs for the return value — it must be { success: true }, not a thrown error.

Check 3 — onConnectionChange fired and ingestion ran

In forge logs or the tunnel output, confirm:

Look for all three signals:

• Handler was invoked: log line from onConnectionChangeHandler with action: CREATED
• Data was fetched: e.g. [connector] Fetched N items
• setObjects succeeded: e.g. [connector] Batch 1: N accepted, 0 rejected

If setObjects returned { success: false }, the error detail is in result.error — surface it to the user and fix before continuing.

Check 4 — Objects visible in Rovo Search

1. Open Rovo Search on the Jira site (allow up to 5 minutes for indexing after ingestion)
2. Search for a word that appears in at least one ingested object's displayName or content text
3. Filter by the connector's nickname (set by admin at connection time)

At least one result from the connector should appear. If nothing shows up after 5 minutes:

• Confirm setObjects logged N accepted with N > 0 (Check 3)
• Confirm the object's permissions match the logged-in user (an EVERYONE principal or a user/group principal that includes the test user)
• Re-check that write:object:jira scope is present in manifest.yml and the app was redeployed after any scope change

Check 5 — Permission boundary (skip only if EVERYONE was used)

If the connector uses user or group principals:

1. Log in as a user who should have access → confirm the object appears in Rovo Search
2. Log in as a user who should not have access → confirm the object does not appear

If a restricted object is visible to an unauthorised user, re-check the accessControls principals in setObjects and redeploy.

Check 6 — Rovo Chat references connector data

Ask Rovo Chat a question whose answer exists only in the ingested content, e.g.:

"What is the status of [title of an ingested item]?"

Rovo Chat should cite the connector as a source. If it cannot find the content, Checks 3 and 4 likely have an unresolved issue.

Check 7 — Scheduled re-ingestion fires (if configured)

If a scheduledTrigger was added:

1. Temporarily set interval: fiveMinutes in manifest.yml, redeploy, and wait one cycle
2. Confirm forge logs shows a fresh ingestion run from refreshIngestionHandler
3. Restore the original interval and redeploy before going to production

Production readiness gate

If the user chose to run verification, only proceed to a production deploy (forge deploy -e production) when all applicable checks above pass:

Manifest Reference

Key rules:

• Scopes are read:object:jira, write:object:jira, delete:object:jira — NOT read:graph:teamwork / write:graph:teamwork (those fail forge lint)
• function: is declared under modules:, not at the top level
• Egress uses address: not a bare string (run forge lint --fix to auto-correct)
• formConfiguration uses form: [{ type: header, properties: [...] }] — NOT fields: or beforeYouBegin:

Minimal connector (no admin config, no OAuth)

Use when the app operates entirely within Atlassian — no external credentials needed.

Connector with orchestration (platform-managed refresh)

Add an orchestration.taskRunner block to any connector variant to enable platform-managed periodic re-ingestion. Wire it alongside your existing onConnectionChange handler.

timeoutSeconds is only valid on a function referenced by a consumer module (like orchestration.taskRunner). Setting it on an unreferenced function fails manifest validation.

Connector with admin form config (API key / URL)

Use when the admin must provide credentials to connect to an external system.

Handler Signatures

Critical: Forge passes the request directly as the first argument — it is NOT wrapped under event.payload. Config form values are at request.configProperties, not event.payload.config. Getting this wrong causes TypeError: Cannot destructure property of undefined.

onConnectionChange

validateConnection

refreshIngestion (scheduled trigger)

Object Types

Objects in bold are indexed in Rovo Search and Rovo Chat.

Rovo Search / Rovo Chat Surfacing

Once ingested:

• Objects appear in Rovo Search under a subfilter named after the connector's nickname (set by admin at connection time)
• Rovo Chat can reference and cite connector objects in responses when queried about topics related to the ingested content
• Data is not available immediately — allow a few minutes for indexing after onConnectionChange fires

To verify ingestion is working:

1. Open Rovo Search on the Jira site
2. Search for text that appears in an ingested object's name or properties
3. Filter by the connector nickname to narrow results

Batching Pattern for Large Datasets

Orchestrated Re-Ingestion (recommended)

Use orchestration when the user selects a refresh schedule. The platform invokes taskRunner on the configured cadence per connection — no external scheduler needed. See Orchestration concepts for the full reference.

Install the uuid package for task ID generation:

Three IDs you receive in every taskRunner invocation

onConnectionChange — schedule the root task

Call scheduleOrUpdateTask on both CREATED and UPDATED. Reusing the same taskId makes the call an in-place update of the existing schedule.

taskRunner — run ingestion with idempotency

Task types (types.FORGE_TASK_TYPES)

taskType is a closed enum — use the SDK constants, never raw strings.

Fan-out with child tasks (large datasets)

For large datasets, fan out from the root taskRunner into per-item children. Use deterministic UUIDs (uuidv5) for child tasks so duplicate deliveries resolve to the same TaskInfo row.

Scheduled Re-Ingestion (legacy — prefer orchestration)

Prefer orchestration (the section above) when the user wants platform-managed periodic sync. Use scheduledTrigger only for simple cron-style triggers where platform task tracking is not needed.

To keep data fresh with a basic scheduled trigger:

Scripts

The scaffold script is in this skill's directory. The deploy script is in the forge-app-builder skill directory — always cd there (or derive the path from this SKILL.md's location) before running it.

Development Best Practices

Follow these practices to avoid common issues. See the official connector requirements for the full guide.

Do not ignore ingestion errors

Always inspect the response from setObjects, setUsers, or setGroups. A call can fail completely or partially reject individual items — and response.success may be true even when some objects were rejected.

Keep access control IDs consistent

When using USER or GROUP principals in object permissions, always ingest the corresponding users or groups first via graph.setUsers() / graph.setGroups(). The id in accessControls must be the exact same string as the externalId passed to those calls. Mismatched IDs silently make objects invisible to users who should have access.

Declare capabilities

Add a capabilities block to the graph:connector module. Without it, admins see "Capabilities not declared by developer" on the configuration screen, which reduces confidence. Set values that accurately reflect your connector's behaviour:

Data deletion responsibilities

When an admin disconnects, Atlassian automatically removes all ingested data. Your onConnectionChange handler with action: DELETED only needs to clean up app-side resources (stop sync jobs, remove webhooks, delete stored credentials).

Troubleshooting

Naming and Logo Guidelines

• Use the official service name as the connector name (e.g. Google Drive, not Drive Connector by Acme)
• Use the official service logo for icons — do not modify or combine with your own branding
• These guidelines apply only to the graph:connector module; your Forge app itself may use your own branding