LavaFlow
Turn LavaFlow on from the CLI so Volcano diagnoses your project's runtime errors and opens pull requests that fix them, then follow what it did.
What it is
LavaFlow uses AI to diagnose this project's errors from their logs and
deployed source, and opens fix pull requests in your connected repository.
It covers every function, frontend and durable function in the project.
Turning it on is the whole setup: there is nothing to install and no key to
bring. You review and merge its pull requests like any other, unless you
choose autonomy merge: then LavaFlow can merge them, and a fix it holds or
stops is yours to review and merge on GitHub.
Where LavaFlow isn't available, every volcano cloud lavaflow command says
why:
- When LavaFlow isn't open to the project or its owner yet, every command says "LavaFlow isn't available for this project yet" and exits with code 4. Volcano checks this before the owner's plan.
- When LavaFlow is open to the project but the owner's plan doesn't include it, every command says "LavaFlow needs a plan that includes it, and the project owner's plan doesn't" and exits with code 5.
- On a Volcano API that has no LavaFlow, every command says "LavaFlow isn't available on this Volcano API" and exits with code 4.
A project that already has LavaFlow records, such as saved settings or past
runs, keeps them readable while LavaFlow isn't available for it:
status, settings, stats, resources, list and show still work, and
disable still turns LavaFlow off. enable says "LavaFlow can't be turned on
for this project: it isn't open to the owner's account, or the owner's plan
doesn't include it" before it asks anything, and exits with code 4. It says
the same when the owner's plan couldn't be read.
settings set exits with the plan's code 5 or code 4, as above.
How it relates
- Belongs to a project. It reads the runtime errors of the project's functions, frontends and durable functions.
- Opens pull requests in the repository the Git connection names. Without one, LavaFlow diagnoses issues and opens no pull requests. See git.md.
- Fixes code it can read at the commit that built it. See Source commit.
- Cloud only. LavaFlow runs on Volcano's cloud service, which local
development doesn't run, so
volcano lavaflowrefuses and points tovolcano cloud lavaflow.
CLI operations
| Operation | Command |
|---|---|
| Show whether LavaFlow is on | volcano cloud lavaflow status [--json] |
| Turn LavaFlow on | volcano cloud lavaflow enable [--yes] [--json] |
| Turn LavaFlow off | volcano cloud lavaflow disable [--json] |
| Show the settings in full | volcano cloud lavaflow settings [--json] |
| Change the settings | volcano cloud lavaflow settings set [--autonomy a] [--human-review-path p]... [--clear-human-review-paths] [--allow-ungated-workflows true|false] [--yes] [--json] |
| Show issues and fix pull requests per day | volcano cloud lavaflow stats [--days n] [--json] |
| Show what LavaFlow did for each resource | volcano cloud lavaflow resources [--days n] [--json] |
| List runs | volcano cloud lavaflow list [--status s] [--resource id] [--limit n] [--cursor c] [--json] |
| Show a run | volcano cloud lavaflow show <run-id> [--json] |
| Start a run on an issue | volcano cloud lavaflow runs start <issue-id> [--json] |
Every command takes --json for the API's own fields and values.
Text from a run, such as its diagnosis, evidence and title, prints as plain
text. A control character in it, such as a color code or a carriage return
copied from a log, prints as a visible escape such as \x1b, and with
--json it stays escaped inside the JSON string.
LavaFlow belongs to the project's owner. For anyone else, every command says
"project not found", as for a project that doesn't exist, and exits with
code 4. A project access token can run the read commands for its own
project, but only the owner's own session can turn LavaFlow on or off,
change its settings and start runs: enable, disable, settings set and
runs start refuse a project access token before they ask or send anything,
with code 5. Sign in with
volcano login instead, and if VOLCANO_TOKEN holds a project access token,
unset it first.
When the owner's plan can't be read, a command says "LavaFlow couldn't read
the project owner's plan; try again later" and exits with code 5. A change
that enable, disable or settings set sends can also be refused for a
read-only account; the API doesn't tell that apart from a plan it couldn't
read, so the command names both, adds what Volcano said, and exits with
code 5. runs start does the same for a read-only or inactive account and a
plan it couldn't read (see Start a run).
Turn it on
volcano git connect
volcano cloud lavaflow enableLavaFlow uses AI to diagnose this project's errors from their logs and deployed source, and opens fix pull requests in your connected repository.
LavaFlow sends this project's logs and source code to an AI service that Volcano operates.
Turn LavaFlow on? Type 'y' or 'yes' to confirm: y
✓ LavaFlow is on
LavaFlow: on
Autonomy: open fix pull requests
Human review paths: none
Ungated workflows: not allowedenable says what LavaFlow does and where it sends the project's data, then
asks before it turns LavaFlow on. --yes skips the question, not those
sentences, which go to stderr so --json output stays JSON. Without --yes, enable needs a
terminal to answer: in a script or CI job it exits non-zero and changes
nothing.
disable turns it off. No new run starts; past runs and their pull requests
stay readable.
Check the status
volcano cloud lavaflow statusLavaFlow: on
Autonomy: open fix pull requests
Merging: available
Repository: acme/shop
Ungated workflows: not allowed
Updated: 2026-10-30 12:00While LavaFlow is off, the status says what it does and how to turn it on.
When you use a project access token, which can read but not change the
settings, the status says so instead.
When it is set to open pull requests and no repository is connected, the
status says to run volcano git connect and then, as the owner,
volcano cloud lavaflow enable: LavaFlow starts no run for a repository other
than the one it was turned on for, so after you connect or change the
repository, turn it on again. When LavaFlow isn't available for
a project that has LavaFlow records, the status notes that it can't be turned
on or changed, and that the owner can still turn it off.
Merging says whether LavaFlow can merge its fixes for the project; settings
shows it too. While it reads not available for this project,
settings set --autonomy merge is refused. When merging isn't available while
LavaFlow is on and available at autonomy merge, merging is paused, and it
reads:
Merging: paused; waiting fixes resume when it's back, new fixes open as pull requests for reviewA fix already waiting merges once merging is back, and a new fix opens its pull request for you to review. The line is left out when the project's capabilities could not be read or don't list merging.
With --json, repository_state is connected, not_connected, or
unknown when the connection could not be read. available is true when
LavaFlow can be turned on and changed for the project and false when it
can't; it is left out when the project's capabilities could not be read.
merge_available is true when LavaFlow can merge its fixes for the project
and false when it can't; it is left out when the Merging line is.
Choose how far it goes
volcano cloud lavaflow settings set --autonomy diagnose_only--autonomy takes one of three values:
open_pr: LavaFlow diagnoses each issue and opens a fix pull request for you to review and merge.diagnose_only: LavaFlow diagnoses each issue and leaves any fix to a person.merge: LavaFlow can also merge its fix pull requests. See Merge fixes.
A setting you don't pass keeps its value.
LavaFlow writes and reviews fixes by your repository's AGENTS.md and
CLAUDE.md files, so tell it about your project there: how to run the tests,
conventions, and areas to treat with care. To make a person approve every fix
that changes a path, use --human-review-path; see Merge fixes.
Merge fixes
At autonomy merge, LavaFlow can merge its fix pull requests. A fix built on
a commit the deploy reported, at either
verification level, is always held for a person.
Only a fix on a commit that a push to the repository deployed is free of that
hold. A fix it holds or stops is yours to review and merge on GitHub.
volcano cloud lavaflow settings set --autonomy mergeTurning on merge lets LavaFlow's fixes, including those from runs it starts on its own, run your CI with the secrets your pull request workflows get.
--autonomy merge prints that and asks first. --yes skips the question but
still prints it, on stderr, so --json output stays JSON. Without a terminal
and without --yes, it changes nothing and exits with a non-zero code.
Only the project's owner can turn merging on, signed in with volcano login
while that sign-in lasts, or with a platform token that has an expiry. Under
merge, the same goes for enable and for changing the human review paths.
The owner's GitHub account must also
be able to push to the project's repository. Volcano refuses these changes
with person_required for a platform token that never expires, a sign-in that
ended, an owner whose GitHub account can't push to the repository, or a
project with no repository. The command says why in Volcano's words and exits with code 5.
Checking that the owner can push
says why the owner can't. While VOLCANO_TOKEN is set, its token is used
instead of a volcano login sign-in. A project access token is refused before
anything is sent, also with code 5. Merging applies to the
repository the project is connected to when you turn it on: once the project
is connected to another, autonomy reads open_pr until you turn merge on
again. Merging is only for a project where
it's available, which status shows; elsewhere --autonomy merge says
"LavaFlow merging isn't available for this project yet" and exits with
code 1. When status shows merging isn't available while LavaFlow is, it
says so before it warns or asks. A project already at merge can still save
--autonomy merge again.
--human-review-path names a path whose changes always hold a fix for a
person, such as db/migrations/**. Repeat it for each path, up to 50, each 1
to 200 characters. It replaces the saved paths; --clear-human-review-paths
removes them all. Volcano's own held paths, such as deploy configuration and
authentication code, always hold a fix too. settings prints the paths. Like
--autonomy merge, both flags need merging to be available for the project;
elsewhere they say "LavaFlow merging isn't available for this project yet"
and exit with code 1.
volcano cloud lavaflow settings set --human-review-path 'infra/**' --human-review-path 'db/migrations/**'show prints Merge: in its Fix: block with where the merge is:
waiting for CI, queued to merge, merged, held or stopped. A merged
fix says merged by LavaFlow, or merged by someone else when a person
merged it while LavaFlow waited. While merging is paused, a
waiting fix says paused while merging is off instead. held and
stopped are final: LavaFlow doesn't touch the pull request again, and it's
yours to review and merge on GitHub. A held fix lists each reason in Volcano's
words, with the paths it names; a stopped fix gives its reason, such as
checks failed. While the pull request is open, the Fix: block ends with a
line saying it's yours, and in list the run shows Held: yours to merge or
Stopped: yours to merge as its status. Once someone merges or closes the
pull request, show keeps the Merge: line and its reasons, and list
shows the run's own status.
Fix:
Pull request: #42 https://github.com/acme/shop/pull/42
State: open
Opened: 2026-10-30 09:03
Merge: held
- It changes a path in human_review_paths. (db/migrations/0042_orders.sql)
The pull request is now yours to review and merge on GitHub.Allow ungated workflows
By default, LavaFlow doesn't open a fix's pull request where a workflow in
the repository could run it, or put its text into a command, with the
repository's secrets before a person approves it: the fix stays on its run
instead. Only the project's owner can turn that on, signed in as a person:
with volcano login, or another platform token that expires. It applies to the
repository the project is connected to now, while their GitHub account can push
to it:
volcano cloud lavaflow settings set --allow-ungated-workflows trueWorkflows in the repository may then run LavaFlow's pull requests, or put their text into commands, with the repository's secrets before a person reviews them. LavaFlow writes those changes from the project's error logs, which its users can influence. Turn it on only if those workflows keep secrets away from untrusted code; gating them on a label a person adds is safer.
Connecting the project to another repository, or disconnecting it, turns the setting off, and connecting the first repository again doesn't turn it back on.
settings set prints that warning and asks before it turns the setting on.
--yes skips the question but still prints the warning, on stderr, so
--json output stays JSON. Without a terminal and without --yes, it changes
nothing and exits with a non-zero code. When Volcano refuses it, for example
for a platform token that never expires, or while the owner's GitHub account
can't push to the repository, the command says why in Volcano's words and
exits with code 5. Any of the owner's credentials turns it off, with
--allow-ungated-workflows false, which asks nothing. While LavaFlow isn't
available for the project, Volcano turns the setting off only together with
LavaFlow, so the command turns LavaFlow off too and says so on stderr.
status and settings show it as Ungated workflows:, allowed for the connected repository or not allowed.
When a run's pull request opened past ungated workflows, show lists the
first ungated workflow job on each branch or commit LavaFlow checked: its file,
the branch or commit whose workflows run it, the job and the event that runs
it. Other jobs on the same branch or commit may run the pull request too.
Ungated workflows: the pull request opened past these, the first one on each branch or commit checked
- .github/workflows/preview.yml on main: job deploy-preview, runs on pull_request_targetA workflow file that couldn't be read shows as "couldn't be read".
See what it did
stats counts the issues LavaFlow took on and the fix pull requests it opened,
per UTC day, today included. Merges aren't counted. --days takes 1 to 90 and
defaults to 30.
volcano cloud lavaflow stats --days 7resources lists every function, frontend and durable function: when
LavaFlow last analyzed it, and the issues and fix pull requests in the window.
A resource whose issues suddenly jumped is flagged issue rate jumped, with
the reason: at least 3 issues in the last 24 hours, and at least 3 times its
daily average over the 29 days before.
volcano cloud lavaflow resourcesType Name Last analyzed Issues Fix PRs
----------------------------------------------------------------------------------------------------
function checkout 2026-10-30 09:00 6 3 issue rate jumped
durable function nightly-report never 0 0Start a run
LavaFlow starts runs on its own only for issues that hold an exception your
code threw or logged. Start one yourself for any other
issue of the project, or for one recorded before LavaFlow was turned on. Name
the issue by its ID, such as the Issue that show prints for an earlier
run, or an ID from the issues API. The run follows
the project's settings and starts queued.
volcano cloud lavaflow runs start 5b2d8e1c-3f4a-4c6b-9d7e-8a1f2b3c4d5e✓ LavaFlow run started
Run: 0d6e9b1a-2f4c-4e8d-9a7b-3c5d1e2f4a6b
Status: queued
Follow it with volcano cloud lavaflow show 0d6e9b1a-2f4c-4e8d-9a7b-3c5d1e2f4a6b--json prints the run as the API returned it.
Only the project's owner can start a run, from a live sign-in, while LavaFlow
is on: a sign-in from volcano login, or a platform token that expires. When
the run doesn't start, the command says what to do:
| When | The command says | Exit code |
|---|---|---|
| LavaFlow is off | Turn it on with volcano cloud lavaflow enable | 1 |
| The project no longer links the repository LavaFlow was turned on for | Connect it again with volcano git connect, or turn LavaFlow on again | 1 |
| The project links another repository | Turn LavaFlow on again to fix issues in it | 1 |
| The issue has a run in progress | The run's ID and the show command for it, or list when it isn't among the newest 100 runs | 1 |
| The issue started 5 runs, or the project 200, in the last 24 hours | When to try again | 1 |
| The Volcano API serves LavaFlow but is older than starting runs | This Volcano API can't start LavaFlow runs yet | 1 |
| The project has no such issue | The project has no issue with that ID | 4 |
| The token never expires, or its sign-in ended | Run volcano login, or use a platform token that expires | 5 |
| The owner's account is read-only or inactive, or their plan couldn't be read | All three, and what Volcano said | 5 |
Follow the runs
A run works on one issue. list shows the newest first, with the resource,
the run's title, and the fix pull request. A run's title is the title of the
fix it proposed, or its diagnosis summary; until LavaFlow has either, the
title shows -. Narrow it with
--status (queued, diagnosing, fixing, reviewing, pr_open,
completed, failed or cancelled) or --resource with an ID from
resources --json. When there are more, the
list ends with the command for the next page. With --json, pass
next_cursor to --cursor, with the same --limit, --status and
--resource.
volcano cloud lavaflow list --status failedA running run's status says which stage it is in: "Running: finding the root
cause" (diagnosing), "Running: writing the fix" (fixing) or "Running:
reviewing the fix" (reviewing). An open pull request shows as "PR open". A
status the CLI doesn't know prints as the API gives it.
show prints where the run's issue came from, the commit a run read the code
at, its review, a Fix: block, and its diagnosis and root cause with the
evidence behind them. A run that failed says why.
The Issue source: line says what kind of issue the run works on:
issue.source | Issue source: |
|---|---|
runtime | runtime error |
build | failed build |
code_quality | Code Quality finding |
report | reported problem |
A run on a failed build works on the build's first error. LavaFlow can't run
your build, so its pull request says so: your repository's CI, or the deploy
after you merge, is the first build of the fix. Build issues are in preview
and are not available in every environment yet. A source your CLI doesn't
know prints as the API gives it. A server that doesn't send the source, and a
run that works on no issue, print no Issue source: line.
A run with a fix has a Checks: line. LavaFlow doesn't run your repository's
tests before it opens the pull request, so the line says they didn't run. Your
repository's CI may still run on the pull request after it opens.
The line reads
LavaFlow didn't run the repository's tests before opening this pull request.
for a run that opened one, and LavaFlow didn't run the repository's tests.
for a run that didn't.
When the review ends without passing the fix, a Review stopped: line says
why, in the words of the API's review.stopped:
review.stopped | Review stopped: |
|---|---|
rounds | The last review round still found problems. |
no_revision | LavaFlow gave no revised fix. |
budget_exhausted | The run used its token budget before LavaFlow could revise the fix, not the daily spend limit. |
context_limit | The fix's conversation grew too long for LavaFlow before it could revise it. |
timed_out | The run's time for review ran out. |
reviewers_failed | No reviewer of the last round finished. |
too_few_reviewers | Fewer reviewers could run than a round needs to pass the fix. |
unreadable | A retried run couldn't read the review saved with its fix, so a person must review it. |
A reason your CLI doesn't know shows as Other with its code. When no
reviewer could run, the review reads Review: error; no reviewer could run,
followed by Review stopped: with the reason.
The Source: line gives the commit's short SHA and where it came from:
| Shown | Meaning |
|---|---|
deployed from a push | A push to the project's connected repository deployed it. |
reported by the deploy; deployed files match | The deploy reported it, and Volcano matched the deployed files against it. |
reported by the deploy; found on the production branch | The deploy reported it, and Volcano confirmed only that the production branch holds it: the repository level function and durable function deploys get. Volcano doesn't compare the deployed files with it. A fix on such a commit opens only while the production branch still holds it. |
production branch; may not be deployed | A reported problem's run read the production branch at a commit it pinned. |
A run recorded before Volcano kept the origin, or with an origin your CLI
doesn't know, shows the commit alone. A run that read no code, such as one
that diagnosed from logs alone, has no Source: line.
The Fix: block always says what happened to the run's fix:
- With a pull request: its link, its state (
open,draft,mergedorclosed), when it opened, and where its merge is (see Merge fixes). A draft says why it's one, such as LavaFlow's review not passing the fix. A pull request opened before LavaFlow recorded the time has no open time. - With none: why LavaFlow opened none, and what to do next, such as reading
the diagnosis or reviewing another LavaFlow pull request that is still
open. A run that finished without a recorded reason says
No fix pull request was opened. LavaFlow didn't record why. - While the run is queued or working:
No fix yet: LavaFlow is still working on this run.
volcano cloud lavaflow show 9c1e4f2a-7b3d-4e8f-a1c2-5d6e7f8a9b0cRun: 9c1e4f2a-7b3d-4e8f-a1c2-5d6e7f8a9b0c
Status: PR open
Resource: function checkout
Title: Load the cart before checkout reads it
Issue: 5b2d8e1c-3f4a-4c6b-9d7e-8a1f2b3c4d5e
Issue source: runtime error
Source: commit 4f1c2b3 (deployed from a push)
Started: 2026-10-30 09:00
Review: passed after 2 round(s)
Checks: LavaFlow didn't run the repository's tests before opening this pull request.
Fix:
Pull request: #42 https://github.com/acme/shop/pull/42
State: open
Opened: 2026-10-30 09:03
Diagnosis
Checkout reads the cart before it loads.show --json and list --json print the run as Volcano's API returned it,
including fields newer than your CLI.