Send and Retrieve Feedback#
Anyone using the Web Terminal can report a problem, or ask for something, from the terminal itself — no bug tracker account, no separate form. The Feedback control sits in the footer of the sidebar, next to the Docs link, and opens a dialog with a text box, a channel picker, and two attachment checkboxes.
Every submission is recorded on the deployment it was sent from, whichever
channel the sender picks. Facility staff read those records back with
osprey feedback list and osprey feedback export — see
Read the feedback a deployment collected below.
Choose a channel#
Pick Local when the report is for this deployment’s operators (the usual case in a control room). It keeps the report on this deployment, and it is the only channel that works with outbound access blocked.
Local and Email are always offered. Between them sits one channel per issue tracker the facility configured (a GitHub repository or a GitLab project), captioned however the facility named it. Switching between channels keeps the typed text and both checkboxes as they were.
Channel |
What it does |
Use it when |
|---|---|---|
Local |
Writes the report to this deployment and nothing else. Send is the only button. |
The report is for the people who run this deployment — the usual case in a control room, and the only one that works with outbound access blocked. |
An issue tracker (GitHub or GitLab) |
Also opens a new-issue tab on that tracker. Without session context the issue body arrives complete. With it, the whole report (text, metadata, session context) is copied to the clipboard, and the issue body is a single line saying to paste it there: the context is far too large to prefill. |
The report belongs with the people who read that tracker, and the sender has an account there. |
Also opens a mail draft to the configured address, following the same rule: complete when it fits, one paste when session context (or a very long report) is attached. |
There is a maintainer address to write to and no issue tracker in the loop. |
The tracker and Email channels never send anything by themselves. OSPREY prefills a tab or a draft in the sender’s own browser and mail client; the sender still reviews it and presses send. The server posts nothing off the host, on any channel.
Both outbound channels write the same record on the deployment as a Local send does, so the facility keeps its own copy of every report even when the conversation continues in a public issue.
A tracker the deployment has not configured is not offered at all. If the Email channel has no address behind it, its radio button still appears, but the action is refused with a short explanation rather than aiming the report somewhere nobody reads. See Documentation and feedback keys.
What gets attached#
The dialog’s (?) control opens a plain-language disclosure of exactly what
leaves the page — the same list as this section.
Always sent, with no checkbox to turn it off:
the feedback text that was typed,
a timestamp,
the sender’s username, when the deployment knows it (in a single-user or development setup this falls back to the account the server runs under).
Deployment metadata — the checkbox is on by default. It adds:
the OSPREY version,
the preset the deployment was built from, with a short prefix of its hash,
the channel-finder mode,
the application name,
the sender’s browser,
an escalate to OSPREY link, when the deployment has one configured: a prefilled upstream issue that carries the deployment’s identity and none of the report’s text, for a maintainer who forwards a framework bug.
Turning it off does not remove the text, the timestamp, or the username, which travel either way.
Session context — the checkbox is off by default, and is greyed out with “no session yet” when there is no live session to attach. It adds:
the session’s id — on the tracker and Email channels the issue title or mail subject also carries a short prefix of it, so the message can be matched to the deployment’s own record even if the paste step is missed,
the session’s tool-call and agent event log,
its chat history,
the terminal scrollback,
an artifact inventory listing titles and ids only, in two sections: the artifacts this session owns (its row in the gallery sidebar), and artifacts no session owns from the same time window.
The second section is not this session’s work: it may come from the chat panel or a CLI conversation. Only titles and ids appear — never contents — but a sender attaching context to a public issue should know it is there.
Session context has a size budget. When a session is over it, the oldest content within each part is dropped first, so a long session is attached in part, with its most recent activity kept. The outbound copy (the clipboard payload behind a GitHub issue or a mail draft) is trimmed harder than the record kept on the deployment, which holds a much larger version. Where the session had no readable transcript, the dialog says so and the outbound copy carries the text, the metadata and a one-line note instead of a silently empty context.
Copy session context#
A small Copy button on the session-context row puts the same trimmed context on the clipboard without sending anything and without writing a record. It is enabled whenever a session exists — the checkbox does not need to be ticked. Reach for it when the clipboard has been overwritten after a send, or when only the context itself is wanted — for a chat message, say, rather than a report.
Read the feedback a deployment collected#
Feedback is stored where the sender’s own data lives, not in a central inbox.
Two verbs read it back, both run from the deployment repository (or with
--repo DIRECTORY naming another one):
osprey feedback list # one line per submission, newest first
osprey feedback export --output feedback.json
osprey feedback export > feedback.json # same document, redirected
list prints one row per submission: the record id, who sent it, when (UTC),
which channel it went out on, and the start of the message. It reads submission
headers only, so it stays quick on a large store. A deployment nobody has
written to yet prints one sentence rather than an empty table.
export writes the whole record — who sent it, when, from which session and
which version of the application, how it went out, where it was truncated — and
the captured session alongside it, as one JSON array with one object per
submission. Records are written as they are read, so a large store never has to
fit in memory. With --output, the file holds the export and everything else
stays on the terminal. Without it, stdout carries the JSON document and nothing
else, and errors and the closing summary go to standard error.
On a multi-user deployment, osprey feedback reads each person’s
workspace volume through a container, so the deployment’s container runtime
has to be running. A single-user deployment is read straight off disk.
Warning
On a multi-user deployment, never give agent_data.base_dir a
~-relative path: osprey feedback cannot see what is written there.
Use a path relative to the deployment repo, or an absolute one.
The feedback still gets written, but into a volume these commands do not mount.
The multi-user reader never consults agent_data.base_dir: it mounts each
person’s workspace volume and reads a fixed path inside it. With a
~-relative value, osprey feedback list exits 0 reporting that no
feedback has been sent. Nothing detects that case, so do not read an empty
listing as proof that nobody has written in.
The single-user reader resolves the store through that key, so relocating it
relocates the store and the reader together, and it refuses a ~-relative value loudly
with that explanation.
Where the records live
The two deployment shapes are read in two different ways, and the command
picks the right one for you. What decides is whether the deployment has a
web-terminal roster: a non-empty modules.web_terminals.users list.
Multi-user deployment. Each person on the roster has their own web terminal and their own workspace volume, and their feedback stays in it. Reading the deployment’s feedback starts a short-lived, read-only container on each person’s workspace in turn and reads the store through it. The deployment’s container runtime has to be running. With it stopped, the command fails loudly instead of reporting a deployment with no feedback in it.
A workspace that cannot be read is named on screen, and the rest of the deployment is still listed or exported, so one broken volume never hides everybody else’s reports. The export’s closing summary distinguishes a workspace that contributed nothing from one that was only partly readable. If nothing at all could be read, the command stops rather than printing “no feedback”.
A roster user’s workspace volume is created when their web terminal
container first starts. A repository that has been built but never brought
up with osprey up therefore has no volumes to read. Every user then reads
as unreadable, and the command stops with that rather than reporting an
empty deployment.
Single-user deployment. There is one store, on this machine, under the deployment’s own agent-data directory, and it is read straight off disk: no container, nothing to start first. This is also the shape a development setup has.
Exit status is 0 when the read succeeded, including a deployment that has
collected no feedback. It is non-zero when the repository has no build, when
every workspace failed to read, and when an export stopped part-way. A part-way
export still parses: it holds what was read, and the summary says how much that
was.
Point the channels at your facility#
web.feedback.trackers lists the issue trackers the dialog offers — a
self-hosted GitLab, the upstream GitHub repository, both — and
web.feedback.email names the address behind the Email channel. The table,
the shipped defaults, and what a blank value means are in web: — documentation, feedback, the bars and the chat pool;
the tracker list itself is described in Issue trackers. The Local
channel is always available and needs no configuration at all.
Note
The feedback dialog is behind the terminal’s own gate: only a browser holding a session can open it, the same as for any other terminal action. It draws no line within a session, though — anyone sitting at a signed-in terminal can send a report — so read a stored report as coming from that terminal, not from a named person.
See also
- Run the Web Terminal
Running the terminal the feedback dialog lives in.
- web: — documentation, feedback, the bars and the chat pool
The full
websettings table.- Multi-User Support
Per-user web terminals and the roster that decides how feedback is stored and read.
- CLI Reference
osprey feedback listandosprey feedback exportwith every option.