UPDATE_CHICKPEA_CLOUDFLARE.md

Rama main · réplica de lectura del repositorio original.

# Update Chickpea on Cloudflare

Update the user's existing Chickpea installation to the latest stable application
release. Handle the terminal and browser work, preserve their settings and data,
and verify the deployed version and Admin access. Ask for help only when you need
missing account information, sign-in, or a decision about their custom changes.

Use the ordinary guarded `npm run deploy` command. The `supportedOrigins` list
belongs to the optional receipt-based `npm run upgrade` tool; an empty list does
not prevent a normal update. Rollback support is not an update prerequisite.
Do not edit release metadata to satisfy that tool or use it for this workflow.

An update does not include sending Slack messages or running Agents, connections,
or schedules as tests. Leave those checks to the user unless they request them.

## 1. Find the existing installation

Start in the user's Chickpea folder. Inspect its Git status, deployment config,
and any private installation notes. Match the configured Cloudflare account,
Worker, public URL, and `AUTH_DB` database ID to the live installation. Reuse its
Wrangler login or named profile. Preserve local edits and credentials.

Read the live deployment settings before changing source. Keep the existing
Worker and database names and IDs, resource bindings, non-secret variables,
routes, triggers, and other service settings. A local config can be older than
the live deployment. Secret values stay in Cloudflare; do not print or copy them.

## 2. Update the source and dependencies

Choose the latest stable application release from the
[official releases](https://github.com/pejmanjohn/chickpea/releases).
Ignore `cli-v*` releases, drafts, and prereleases. Fetch its exact tag and read
its release notes. Update the checkout to that release while preserving the
user's installation settings and unrelated local changes. Resolve any conflicts
without discarding their changes; use a separate checkout if needed.

This current guide governs the update workflow, including when the release's
older update instructions require the receipt tool or a supported-origin entry.
Use the selected release's Node pin, dependency and deploy commands, configuration
format, and migration steps. Select Node from `.nvmrc` using an available Node
manager, then install the release's lockfile dependencies.
For releases with the reviewed `allowScripts` policy, use:

```sh
npm ci --strict-allow-scripts
```

v0.1.34 and later need Node 24.22.1 or a later Node 24 on the machine that runs the
update. Its dependency install refuses older Node, including 24.20.0,
the previous pin. Cloudflare Workers Builds, which the Deploy to Cloudflare
button also uses, reads `.nvmrc`. If the installation sets a `NODE_VERSION`
build variable, change it to 24.22.1 or remove it; a build on older Node stops
at dependency install. A Workers Builds build of v0.1.34 was not tested before
release, so check that its build log shows Node 24.22.1.

Keep the new release's code and migrations. Restore the installation's live
deployment settings in the authored `wrangler.jsonc`, including `account_id`,
Worker `name`, the `AUTH_DB` database name and ID, non-secret variables, and
service settings. Preserve resource identities while keeping the new release's
migrations, compatibility flags, and asset configuration. Do not copy old
application-version metadata over the new build identity or edit generated files.

Read any required migration steps. A missing `supportedOrigins` entry or a
changed source-file digest alone is not an incompatibility. Stop for a concrete
problem such as an unresolved target mismatch or a migration that would lose
data, and explain that problem rather than telling the user to wait indefinitely.

Some releases add Durable Object migrations; v0.1.27 adds two, on every
installation: `v10` for coding workers and `v11` for per-thread Slack turn
runners. v0.1.28 through v0.1.43 add none; an update from v0.1.26 or
earlier straight to any of them still applies `v10` and `v11`. Once the deploy applies them,
Cloudflare refuses `wrangler rollback` and dashboard rollbacks to any version
from before that release. Recovery from a bad update then means deploying a newer or fixed
release forward. Tell the user this before deploying such a release.

v0.1.27 also makes per-thread runners the default turn executor; nothing to
configure. If new Slack turns misbehave after the update, redeploy with
`npm run deploy -- --var SLACK_TAG_TURN_EXECUTOR:alarm` to return new turns to
the previous executor without a rollback, and remove the variable later to
return to the default. See the
[upgrading runbook](docs/runbooks/upgrading.md).

Releases that let people stop a run from Slack need nothing to configure.
Rolling back from one is safe, except that messages a stop is still holding
(usually for a few seconds) then run on the older release. A customer-owned
Slack app created before that support shows Slack's **Stop** button only after
it subscribes to the `agent_session_stopped` event; without it, people stop a
run by replying `stop` in its thread. Tell the user this optional step is in
[Slack's Stop button](SETUP_AGENT.md#slacks-stop-button), and change their
Slack app only if they ask. Installations on the shared Chickpea app need
nothing. See [Slack steering](docs/runbooks/slack-steering.md#rollback).

Releases that let Agents answer in group DMs need nothing to configure. A
customer-owned Slack app created before that support hears only `@Chickpea`
in group DMs, and reads no earlier group DM messages, until it subscribes to
the `message.mpim` event and has the `mpim:history` permission. Tell the user
this optional step is in [Group DMs](SETUP_AGENT.md#group-dms), and change
their Slack app only if they ask.

Releases that request permissions for upcoming features need nothing to
configure. A customer-owned Slack app created before them keeps working
without them, and no feature in this release uses them. Tell the user this
optional step is in
[Permissions for upcoming features](SETUP_AGENT.md#permissions-for-upcoming-features),
and change their Slack app only if they ask.

v0.1.27, v0.1.28 and v0.1.29 cannot read two kinds of record that v0.1.30
writes, and a single one stalls them after a rollback. On a shared-gateway
installation, a button click or form submission that is still queued stops the
older release's gateway queue for up to 7 days. The other is a turn on an
OpenRouter model newer than the built-in model list. If a reply, routine run or
coding task on such a model is still running at the rollback, no thread starts
a new turn. Before rolling back from v0.1.30, make sure none is running. If
Slack goes quiet after a rollback, redeploy v0.1.30 with the same deploy
command. It processes the stuck record. Roll back again only after that work
finishes.
[Recover a stalled rollback](docs/runbooks/operations.md#recover-a-stalled-rollback)
says how to recognise the stall. Later releases set such a record aside for
operator recovery instead of stalling.

v0.1.30 and earlier cannot read three kinds of turn that v0.1.31 and later write. One
that has not finished stalls every thread after a rollback. The three kinds are
a turn that starts a conversation (a new thread, the first turn in an older
thread after the update, or a routine run), a question one Agent asks another,
and a turn whose saved thread context includes a post from an app. Most finish
within seconds, but a routine run or coding task can take up to an hour. Before
rolling back from v0.1.31 or later to v0.1.30 or earlier, make sure no reply,
routine run, coding task or Agent-to-Agent question is running. If Slack goes
quiet after a rollback, redeploy the release you rolled back from with the same
deploy command, let that work finish, and roll back again.
[Recover a stalled rollback](docs/runbooks/operations.md#recover-a-stalled-rollback)
lists the log lines that identify this stall.

After a rollback from v0.1.31 or later to v0.1.30 or earlier, Agents in
existing threads no longer remember the turns they took since v0.1.31, and the older release reads apps' posts in a
thread as people's posts. A person whose last turn in a thread carried an app's
post gets failed replies from that Agent there until you update again. v0.1.31
sets unreadable records from a later release aside instead of stalling.

Rolling back from v0.1.32 to v0.1.31 needs no extra care. v0.1.31 answers a
teammate's handed-back answer that is still queued as an ordinary report, so the
Agent always replies, even where v0.1.32 would have posted nothing. A turn
that had already decided to post nothing but had not finished when you rolled
back may post the literal text `NO_REPLY` once.

Rolling back from v0.1.34 to v0.1.33 keeps Admin and Slack replies working.
The Agent-level connection secrets that v0.1.34 deleted (see step 3) do not
come back, and v0.1.33 cannot read secrets entered on v0.1.34. Each v0.1.34
start deletes any Agent-level secret that v0.1.33 saved, and connections saved
in Admin on v0.1.33 need their secret entered again after the next update.
Replies do not use these connections on either release. While on v0.1.33, an
Agent created with the name of an Agent deleted during the rollback picks up
the Agent-level secrets that Agent saved there, as v0.1.33 always did. A reply
still unsent at the rollback for an Agent that
uses a team connection needing an admin to reconnect it is set aside without
an answer, and a thread whose first v0.1.34 turn ran in that state gets
failed replies on v0.1.33 until you update again. Prefer deploying a fixed
release forward to rolling back across v0.1.34.

## 3. Deploy to the same Worker

v0.1.43 has Chickpea ask before an Agent the person has not used takes a
message at the top of their direct message with Chickpea. Chickpea still hands
it straight to an Agent the person created, or one that answered them in that
direct message in the last 7 days. For any other single Agent that fits, it
replies that the Agent can take it, with one button for that Agent and one for
Chickpea, and hands nothing over until the person picks. A message deleted
while Chickpea hands it on is neither handed over nor put on a card, and
Chickpea posts nothing for it. An edited one is not handed over either:
Chickpea says it changed, and the edit's own turn decides again. A message
that is a single word of ASCII characters, such as a pasted key, never goes on
a card, so only an Agent that may take a message without asking can take it;
otherwise Chickpea answers itself and suggests mentioning the Agent. Website
logins now belong to the team: only Owners and Admins add, grant, raise or
remove them, and a member sees a line asking an Owner or Admin to add one. A
member's personal login saved before the update is shown and managed as a team
login; its stored record is not rewritten. Everyone who may edit that Agent
now sees the login's site, label, sign-in name and last use, and a member can
no longer remove it, or delete an Agent that alone holds it, until an Owner or
Admin removes it. The storage contract raises `settings.browser.*` to 2, so
`release.json`'s `configuration` digest changes; there is no migration.
Updating needs no other step. There is no rollback path: recovery means
deploying a fixed release forward.

v0.1.42 makes Chickpea the front desk of its direct messages. A new message
at the top of a person's direct message with Chickpea that mentions no Agent
goes to the Agent that fits: Chickpea replies that the Agent is taking it, or,
when several fit, offers one button per Agent, and it still answers requests
about the workspace itself. When the person who asked an Agent for a
connection in a Slack thread finishes its setup, that Agent now continues
there and tests the connection; otherwise the thread gets one plain line. An
Agent can take a key that the person who approved its connection sends within
30 minutes, in the same conversation. A managed sign-in started in Admin now
ends on a Chickpea page that tries to close itself, and an approval preview
for a long text edit lists its key changes instead of the whole text. Neither
the storage contract nor any migration changes, so every `release.json`
digest stays the same. Updating needs no other step. There is no rollback
path: recovery means deploying a fixed release forward.

v0.1.41 lets an Agent read up to 10 files attached to one Slack message, up
from 4; a message with more still reads none of its files. Editing a message
while an Agent is answering it, or before its turn starts, now steers that
turn: the Agent reads the new text, at its next step if it had already read
the message, and an edit that comes too late for the reply in progress gets
its own reply afterward. An edit of a message with files, a command, an
approval, or a mention of more than one Agent steers nothing, and neither does
an edit that turns a plain message into a command. In Admin, a row under
**Usage → Recent activity** links to the Slack message that started it, or for
a routine, to the message the routine posted. On the shared Chickpea Slack app
the link opens in whichever Slack workspace your browser has open.
**Settings → MCP** now loads when you open it from another Settings page. On
its first start the usage store adds two nullable columns for the Slack
message and thread; the storage contract is unchanged, so `release.json`'s
`configuration` digest stays the same. Updating needs no other step. There is
no rollback path: recovery means deploying a fixed release forward.

v0.1.40 lets a person install skills by attaching `SKILL.md` files or `.zip`
and `.skill` archives to a Slack message, and lets a coding agent upload skill
files over MCP (`upload_skills`) or with `chickpea skills upload`. Agents list
the files inside an attached zip, read files inside it, and read an earlier
file in the thread with `read_slack_file`. On the shared Chickpea Slack app, a
file attached to a thread before this update has no handle. A message sent
while an Agent is working now reaches it at its next step, including during a
coding job, and the Agent can answer a question in the thread before its final
reply. A message that waits for the Agent's next turn keeps 👀 until that turn
reads it, then shows 👍 until the reply posts. On its first start the
configuration store adds a nullable column for thread file references, and the
storage contract records it, so `release.json`'s `configuration` digest
changes. Updating needs no other step. There is no rollback path: recovery means
deploying a fixed release forward.

v0.1.39 lets an Agent's author upload `SKILL.md` files and `.zip` or `.skill`
archives on the Agent's Skills tab in Admin. It lets the Claude app (claude.ai,
Claude Desktop, Cowork) and other MCP clients that do not send
`application_type` sign in to Chickpea's MCP server, and it adds an **Add to
Claude** link to **Settings → Coding agents** and the `/connect` page. Agents
read Slack text attachments of up to 64,000 characters each and 128,000 per
message, up from 32,000 and 48,000; on the shared Chickpea Slack app, the
gateway still refuses a text file over about 25 KB. On its first start
v0.1.39 removes the requester's Slack message that releases before v0.1.38
kept in setup records, because that message can hold a pasted key. Updating
needs no other step. There is no rollback path: recovery means deploying a
fixed release forward.

v0.1.38 lets an Agent search the web and read public pages the person links,
lets a coding agent send a management request of up to 1 MiB (so an Agent
with many large skills installs in one call), and lets a person set up a
custom API connection from chat with a pasted key. Web search needs an
`EXA_API_KEY` Worker secret; without it, Agents still read linked pages.
`release.json`'s `configuration` digest changes because the configuration
store now declares its stored shapes in `src/config/storage-contract.json`;
the store itself has no schema change. Updating needs no other step. There is
no rollback path: recovery means deploying a fixed release forward.

v0.1.37 lets a coding agent connected to Chickpea's MCP server do what the
signed-in person can do in Admin. It asks the person first only where Admin
asks "Are you sure?", such as deleting an Agent or routine, suspending or
removing a member, or disconnecting a connection. In Slack, a change an Agent proposes
applies only after the person approves it. Disabling an Agent now keeps its
Channels, and enabling it again brings it back to them. Updating needs no
other step. A rollback to v0.1.36 was not tested.

v0.1.36 lets people use Chickpea in Slack group DMs. There, `@Chickpea`
and any Agent that everyone in the workspace can use answer in a thread.
Updating changes nothing in Channels or DMs. The shared Chickpea app supports
group DMs. An installation on it hears Agent handles and thread replies in
group DMs, and lets Agents read a group DM's earlier messages, once an Owner or
Admin uses **Reconnect with Slack** in **Settings → Slack**. Until then, group
DMs answer only `@Chickpea`. A rollback to v0.1.35 was not tested.

v0.1.35 adds an effort level to each chat-model choice. Updating changes no
Agent's behavior: every model keeps running at its standard level until
someone picks another in Admin. One exception is an OpenRouter model that
Chickpea's built-in model tables do not list and that does not offer Medium:
with no level chosen, it now runs at the nearest level it offers. A rollback
to v0.1.34 was not tested.

v0.1.34 is a security update for the eMudhra certificate authority compromise.
Its Node requirement is in step 2. Before you deploy it:

- v0.1.34 deletes the secrets saved directly on an Agent's own MCP servers and
  API connections (Agent-level connections): bearer tokens, header values,
  pasted API credentials, and OAuth sign-ins. Team and personal connection
  accounts keep theirs. Admin's Connections tab lists only those accounts, and
  Agents use only those accounts in Slack, on v0.1.33 as on v0.1.34. Replies do
  not change, and an ordinary update needs no action. Someone who uses an
  Agent-level connection through the Admin API or a Chickpea setup link enters
  its secret again: ask Chickpea in Slack for a setup link, and finish one link
  before opening the next, or call
  `PUT /admin/api/agents/<id>/mcp/secrets/<connectionId>` or
  `PUT /admin/api/agents/<id>/api-connections/secrets/<connectionId>` with
  `expectedRevision` and `expectedCredentialBinding` set to the Agent's
  `revision` and the connection's `credentialBinding` from
  `GET /admin/api/agents/<id>`. An `MCP_AGENT_…` or `CONNECTOR_AGENT_…` secret
  supplied as a Worker variable
  applies only when its matching `…_BINDING` variable holds that connection's
  current credential binding. Admin does not show the binding; the Agent's
  record in the signed-in Admin API (`GET /admin/api/agents/<id>`) gives
  editors each connection's `credentialBinding` and variable names.
- An installation with the coding sandbox must update with
  `npm run deploy:sandbox`, as described below, so its container moves to the
  patched image.

v0.1.33 adds an optional ChatGPT plan connection for Cloudflare. Updating keeps
the current chat provider and API key. If you connect a ChatGPT plan and later
roll back to v0.1.32, switch chat back to an API key and disconnect the plan
first: that older release cannot use or manage the plan connection. An OpenAI
API key is still required for image generation.

For the core Cloudflare deployment, run the release's guarded command from the
updated checkout:

```sh
npm run deploy
```

An installation with the optional coding sandbox must keep it. If Admin
**Settings → Coding sandbox** shows **Installed but off** or **On**, or the
deploy reports that the live Worker has the coding sandbox, deploy with
`npm run deploy:sandbox` and the same flags and environment variables described
below. For example, use
`CHICKPEA_DEPLOY_TARGET=production npm run deploy:sandbox -- --profile <existing-profile>`.
This needs Docker running and the `containers:write` permission. Since
v0.1.25 the sandbox also keeps workspace checkpoints in an R2 bucket that the
deploy creates itself, which needs R2 enabled on the Cloudflare account: open
**R2 Object Storage** in the Cloudflare dashboard and enable it (the free tier
is enough; do not create a bucket). If R2 is not enabled, the command still
updates the sandbox, with checkpoints off, and ends with a line saying how to
turn them on. See
[Deploy from the command line](docs/runbooks/coding-sandbox-deployment.md#deploy-from-the-command-line).
A plain `npm run deploy` refuses to replace a live sandbox, and the guided
`npm run upgrade` tool supports only core installations.

If the installation uses a named Wrangler login, use
`npm run deploy -- --profile <existing-profile>` instead. On a host with a
Chickpea QA registry, an ordinary customer target also needs
`CHICKPEA_DEPLOY_TARGET=production`; this selects the configured Worker and does
not create a new environment.

Let the command build, apply its reviewed migrations, and finish its deployment
readiness check. Preserve its target and database guards. Do not create another
Worker or database, rotate permanent auth or encryption secrets, or reinstall
Slack. The command may print a setup link; an existing installation continues
using its usual Admin sign-in and does not need onboarding again.

## 4. Verify and finish

Confirm Cloudflare serves the selected version. Open signed-in
**Settings -> About & updates** and check the application version and source
commit. Do not send a Slack test or invoke an Agent, connection, or schedule.

If deployment or verification fails, inspect the actual serving state and fix
the reported error. Preserve the existing resources and data; do not reset or
reinstall the application to make a check pass.

Update any existing private installation notes with the release and checkout
path. Tell the user which version is installed, whether deployment and Admin
checks passed, and where the local source lives. They can test Slack themselves
when convenient.

## Optional: connect a ChatGPT plan

An ordinary update preserves the existing provider and account connection.
Only when the user asks to connect ChatGPT, follow this installation's
`/connect-chatgpt.md` guide from their local coding agent. Admin **Settings →
Models & providers → Connect ChatGPT** also provides a copyable setup request
for an existing installation. Handle the local setup, let the user sign in,
confirm the account in Admin, and help them choose an available model. Keep
any OpenAI API key for image generation. This connection requires no deployment.