Circleback action items to tasks: testing webhook signatures, owners, and replay keys
Moving meeting action items into a task tracker introduces two decisions immediately: what to do with an unassigned task, and how to handle the same payload arriving again. I built a local adapter from Circleback’s documented webhook schema and tested signatures, missing owners, completed items, and invalid inputs. The adapter keeps uncertain owners and deadlines unresolved and produces a stable key from the meeting and action item IDs.
The experiment was run on September 30, 2026 with synthetic data. It uses no real meeting recordings or customer information. I have not tested delivery from Circleback, writes to an external task tracker, or transcription accuracy. The results below describe the code that processes a payload after receipt.
On October 2, I also tested SQLite storage. Two processes repeatedly saving the same items left three tasks. The added experiment below includes the old-replay case that can reopen a completed task.
For the wider workflow—Slack delivery, huddle capture, and a nightly notes sync—see six Circleback settings for Slack and AI. This article focuses on the next step: exporting action items into a task tracker.
Which work belongs in the task tracker
The target workflow is moving agreed follow-ups into the system where a team tracks ownership and completion. If meeting notes already serve that purpose, another integration may add maintenance without helping anyone finish work. Exporting makes sense when the destination is where people actually update tasks.
I chose a small output contract: title, description, owner email, completion status, and a link back to the meeting. The adapter does not guess an account from a name or infer a deadline from prose. Those guesses could assign work to the wrong person or record a deadline nobody agreed to.
Circleback automations combine conditions and actions. Its documentation describes filters such as tags and participants; leaving conditions empty applies the automation to every meeting. For an initial connection, use a dedicated test scope. Official automation guide
Mapping the documented fields
The adapter uses actionItems fields id, title, description, assignee, and status. An assignee can be null, and status is PENDING or DONE. The action item schema I consulted has no deadline field. Official webhook schema
The right column describes decisions made for this experiment, rather than product features.
| Incoming data | Adapter output |
|---|---|
| Meeting ID and action item ID | A stable key containing both |
| Title and description | Original text, with outer whitespace removed from the title |
| Assignee email | Email when present; otherwise null and an owner-review flag |
| Status | Preserve PENDING or DONE |
| No deadline | dueDate: null |
| Meeting ID | A source meeting URL |
An item with a name but no email still needs an ownership review. Completed items remain in the output, allowing the destination to update an existing task instead of silently retaining its old pending state.
Giving AI the context to continue work also means passing on what remains unresolved. Here, null records an outstanding confirmation rather than hiding a missing value. Keep the source meeting link so the next person can check the basis for an assignment.
Verify the original body before parsing
A signed JSON body and a reserialized object need not contain the same bytes. Spaces and line breaks can change even when the parsed object is identical. Verification therefore runs before JSON parsing, using Circleback’s documented x-signature and HMAC-SHA256 scheme. Signature verification documentation
I used Web Crypto for the local implementation. It rejects malformed signatures, empty secrets, and incorrect secrets. No production secret appears in the sample.
import { verifyCirclebackSignature, buildActionRows } from './circleback-webhook.mjs'
// request comes from the webhook receiver implemented by the operator.
const rawBody = await request.text()
const valid = await verifyCirclebackSignature(
rawBody,
request.headers.get('x-signature'),
signingSecret,
)
if (!valid) throw new Error('Reject request before parsing')
const rows = buildActionRows(JSON.parse(rawBody))This is an integration sketch, not an HTTP server. In the local test, the original body with whitespace passed verification. Parsing and reserializing that body failed verification against the same signature. Check the receiving framework’s body-parser configuration when connecting a real endpoint.
Isolate signature failures with two files
To investigate a signature mismatch first, save these two files in the same directory. Run them with Bun or Node.js 22+. The example uses synthetic data and a test secret, so it needs no product account or production signing secret.
bun signature-demo.mjs
# With Node.js
node signature-demo.mjs{
"synthetic": true,
"sameParsedJson": true,
"differentBytes": true,
"original": true,
"rewrittenJson": false,
"tamperedBody": false,
"wrongSecret": false,
"malformedSignature": false,
"replay": true,
"requestBodyUsed": true,
"secondReadRejected": true,
"verifiedParsedMeetingId": "demo-署名"
}In this synthetic example, parsing produces the same JSON value, but rewriting changes the bytes. The original body passes; rewritten or tampered bodies, a wrong secret and a malformed signature fail. Replaying the identical body passes. Signature verification alone does not prevent processing the same input twice.
The demo performs no HTTP reception, external requests or storage. It does not test live Circleback delivery or the product's replay defenses. Destination-side duplicate handling is a separate concern, addressed by the keys and upsert in the next section.
Do not call request.json() after the signature check
With a Fetch Request that has a body, calling await request.json() after await request.text() reads the same body a second time. The demo constructs a local Request. In both Bun and Node, text() sets bodyUsed to true, and the subsequent json() rejects with TypeError. Fetch body consumption specification
After the signature passes, parse the string already read with JSON.parse(rawBody). The receiver snippet above follows this order. The demo's verifiedParsedMeetingId is demo-署名, obtained from that verified string. This does not test live HTTP reception or body parsers such as Express middleware.
A valid signature does not validate JSON or task fields
verifyCirclebackSignature returning true does not guarantee that JSON.parse or buildActionRows will succeed. The signature checks the body against the secret; the next steps check whether this sample can handle the task data. Circleback documents nullable assignee and a status of PENDING or DONE. Webhook fields and signature documentation
Save these two files in one directory and run them with Bun or Node.js 22+.
- Test validation after signatures: payload-validation-demo.mjs
- Signature verifier and task adapter: circleback-webhook.mjs
bun payload-validation-demo.mjs
# With Node.js
node payload-validation-demo.mjsOn 2026-10-17, Bun and Node produced the same results below. Each of the first six inputs was signed using its own body. For the last input, the empty-items body was signed, then only the body was replaced with {.
| Synthetic input | Signature result | Processing result |
|---|---|---|
actionItems: [] |
true | accepted, 0 rows |
| One item with a null assignee | true | accepted, 1 row, null ownerEmail, true needsOwnerReview |
Invalid JSON: { |
true | SyntaxError at the json stage |
JSON null |
true | TypeError at the shape stage |
status: 'UNKNOWN' |
true | TypeError at the shape stage |
| Two items with the same action ID | true | TypeError at the shape stage |
Body changed to { after signing |
false | Stops at signature; JSON parsing is skipped |
The SyntaxError for parsing { also matches the ECMAScript JSON.parse specification. Valid JSON such as null is still not a meeting object. Rejecting duplicate IDs is a constraint of this adapter, not evidence that Circleback sends duplicate IDs.
The demo checks the signature, JSON syntax, and task shape in sequence, reporting the failed stage and error name. It returns no task rows for invalid inputs and performs no HTTP reception, external requests, or storage. This is a local synthetic-input test; live Circleback delivery and HTTP response or retry policies remain untested. Decide how the receiver records failures and responds when connecting it.
Replays and task updates
A meeting can contain several tasks, so the meeting ID alone cannot identify each row. I used JSON.stringify([meetingId, actionId]). Unlike joining IDs with a separator, the array representation distinguishes combinations even when an ID contains that separator.
Transforming the same input twice produced identical keys. Changing a title and status while keeping its IDs also retained the key. A destination-side upsert can then update the same task.
Stable keys do not prevent duplicate writes on their own. Two concurrent deliveries could both create a task unless the destination enforces uniqueness or stores the external task ID. Use an atomic upsert backed by a unique constraint, or an equivalent destination-specific mechanism. The September 30 test covered key stability. On October 2, I extended it to the SQLite storage example below.
Two SQLite writers leave only three tasks
I tested replayed action items in a local database using Bun 1.3.11's built-in SQLite driver. This storage example belongs after signature verification. It runs neither a webhook receiver nor an external service.
The table defines the meeting/action key as task_key TEXT NOT NULL PRIMARY KEY. An incoming key uses SQLite's upsert, ON CONFLICT(task_key) DO UPDATE, to update that row's title, owner, status, and other fields. There is no separate “check whether the row exists, then insert” step.
The three items from a meeting are written in one transaction through db.transaction(...).immediate(rows). A SQL failure rolls back that meeting's batch. A test that failed on the second item left zero rows, including no partial first item. Values use bound parameters; a title containing quotes and SQL-like text was retained as text.
Save these four files in the same directory and run Bun. The storage demo is Bun-only. The earlier signature and transformation demo still supports Node.js.
- Existing signature verifier and adapter: circleback-webhook.mjs
- SQLite store: task-store.mjs
- Three synthetic items: task-store-fixture.json
- Two-writer runner: task-store-demo.mjs
bun task-store-demo.mjsThe demo creates a SQLite file in a temporary directory. Both writer processes finish setup before they receive a start signal. Each saves the same three items 20 times: 40 batches and 120 row upserts. The demo removes its own temporary directory when it ends.
{
"writers": 2,
"batches": 40,
"rowUpserts": 120,
"storedTasks": 3,
"pending": 2,
"pendingOwnerReview": 1,
"afterUpdateStatus": "DONE",
"afterOldReplayStatus": "PENDING"
}Only three tasks remained after the concurrent writer phase: two pending, with one needing an ownership review. A missing owner email and deadline remained null. A separate test with the same action IDs in a different meeting left six rows, keeping tasks from different meetings distinct.
An old replay can reopen DONE even without duplicate rows
The last two fields show the limit. Updating item 101 to DONE produced afterUpdateStatus: "DONE". Replaying its original PENDING payload then changed that same row back to PENDING. The primary key and upsert prevent duplicate rows; they cannot determine which input is newer.
This sample accepts the last input written. Before connecting it to a real workflow, decide whether meeting data may overwrite a status someone updated in the destination. Fetching the source's current state or comparing a trustworthy update version would require further implementation. The selected adapter fields do not establish event order, and this example does not implement that policy.
Seven tests cover replays, updates, separate meetings, batch rollback, malformed input, the old-replay result, and two-writer storage. These counts describe synthetic data in local SQLite. They do not measure real Circleback delivery, an external task tracker, or storage performance.
Keep destination-owned completion separate from meeting replays
On October 6, I tested a policy that lets the task tracker own its status. This separate sample focuses only on status storage. It stores the incoming meeting value as sourceStatus and a destination override as localStatus. With no override, the displayed status follows the meeting value. With an override, it follows the destination.
COALESCE(local_status, source_status) AS statusThe webhook upsert updates the meeting value, title, and meeting URL, leaving the local override untouched. After marking item 101 DONE in the destination and saving its original PENDING payload 40 times, the source remained PENDING while the local and effective values remained DONE. Three tasks remained.
Save these four files in the same directory and run Bun. This new demo uses an in-memory SQLite database, with no network connection or writes to an existing database.
- Signature verifier and adapter: circleback-webhook.mjs
- Three synthetic items: task-store-fixture.json
- Separate status store: task-status-store.mjs
- Replay and clear demo: task-status-demo.mjs
bun task-status-demo.mjs{
"replays": 40,
"storedTasks": 3,
"sourceStatus": "PENDING",
"localStatus": "DONE",
"effectiveStatus": "DONE",
"afterClearStatus": "PENDING"
}The final field shows an explicit clear. Calling setLocalStatus(key, null) displays the last received source value, PENDING. Before clearing an override, fetch the current meeting state or have someone check the value that will take over. This sample still cannot determine which received payload is newer.
DONE is not permanently locked. Setting a local PENDING displays PENDING even if the source says DONE. This is a workflow choice: a replay must not change a status explicitly chosen in the destination. It does not fit a workflow that must always follow the meeting's status.
Eight tests cover replays, explicit reopening, clearing, invalid statuses and unknown keys, empty or malformed meetings, and batch rollback. Two database connections wrote sequentially; closing and reopening the database preserved DONE. This new sample has not been tested with simultaneous writes from separate processes.
An old replay still overwrites the title. This status-only sample does not store owners or deadlines. External task APIs, permissions, and removal of deleted meeting items need their own design. It uses a separate table from the earlier demo and performs no migration. These results come from synthetic data and local SQLite, not real meeting delivery.
Results from three synthetic action items
The fixture contains one assigned pending item, one unassigned pending item, and one completed item. Its email address, owner@example.com, is a test value. Running the companion demo produces these counts alongside two example referral URLs:
{
"rows": 3,
"pending": 2,
"pendingOwnerReview": 1,
"stableKeys": true
}One of the two pending items requires an ownership review. The completed item remains in the output but is excluded from the pending count. These results show that the adapter preserves the supplied state. They say nothing about how accurately Circleback extracts work or owners from an actual conversation.
I also tested a missing action array, a string ID, an empty title, an unknown status, a malformed assignee, and duplicate action IDs within one payload. Each invalid input stops transformation. An empty action array returns zero rows normally; a meeting with no action items should not create tasks.
Reproduce the experiment
Save these three files into the same directory and run the demo with Bun or Node.js 22+. No dependencies or product account connections are required.
bun demo.mjs
# With Node.js
node demo.mjsThe demo starts no listener and writes to no external service. Change an owner email to null or change an item’s status to inspect how those inputs affect the local output.
Deciding whether to connect real meetings
For a real connection, scope the automation to approved meetings and select only the outcomes the destination needs. This adapter does not use recordings or full transcripts. Review the destination, exported fields, and participant communication before enabling delivery. The documented UI steps are in the webhook setup guide.
A synthetic-data test cannot establish deployment readiness. Use a small set of approved meetings to check missed follow-ups, owner matching, completion updates, and delivery failures. Start with human review before notifications are sent, so an incorrect assignment is not immediately circulated.
This integration is worth evaluating when a team already maintains follow-ups in a task tracker and manual copying is a recurring burden. For a few meetings that are already manageable by hand, maintaining a webhook receiver and storage may be harder to justify. Decide that before building the connection.
The companion article on Dub referral tracking with bilingual UTMs separates language-specific link activity from commission attribution.
Go deeper on this topic
SaaS integrations
Give AI the context to continue the work: source conversations, open tasks, and unresolved details, from adoption decisions to integration code.
Article 22 of 23 in this series.
Next reads
Track Dub referral links in Japanese and English without breaking referral IDs
Adding UTMs can re-encode existing referral parameters even when their decoded values stay the same. Reproduce spaces, tildes, and duplicate keys locally, with limits for signed URLs and commission attribution.
Claude Code 2.1.325: Sandboxed Bash Prompts and Skill Approval
Review sandbox autoAllow, scoped ask and deny rules, and scheduled Skill approval. Read-only audit with 729 synthetic cases; CLI comparison unverified.
Claude Code Permission Wildcards: What * Allows and How to Audit
Compare Claude Code 2.1.322 and 2.1.323 in isolated tests: broad Bash wildcards, spaces, variable warnings, and a read-only settings audit.
Use the topic hub as a map, then continue to the next article in the series. New here? Start at the home hub →
FAQ
Does a valid signature guarantee valid JSON and task fields?
In the synthetic demo, invalid JSON signed with its matching body passes verification, then JSON.parse throws SyntaxError. The adapter rejects JSON null, an unknown status, and duplicate IDs with TypeError. Empty actionItems and null assignees are accepted. This tests the local adapter; live Circleback delivery and HTTP responses remain untested.
Can I call request.json() after verifying the signature?
For a Fetch Request with a body, reading the body with text() and then json() rejects with TypeError. The published synthetic demo checks this in Bun and Node. Read once, verify the signature, then pass that string to JSON.parse. Live Circleback reception and Express body parsers were not tested.
Does a passing webhook signature also block replays?
This local verifier accepts an identical body and signature again. Signature checks alone do not prevent repeated processing; destination-side keys and upserts are separate. The two published files run a synthetic demo, not a test of live Circleback delivery or the product’s replay defenses.
Did you benchmark real meeting transcription?
No. This is a local integration experiment using synthetic payloads and official documentation. Japanese transcription accuracy, real meeting summaries, and paid-plan benefits were not measured.
Does the sample prevent duplicate tasks?
The SQLite example uses the meeting/action key as a primary key with an upsert. Two writers saving the same three items in 40 batches left three tasks. An old PENDING replay can still reopen DONE, so event ordering needs a separate policy. A separate example preserves a local DONE through 40 old replays; clearing its override returns to the last received source. External task trackers have not been tested.