Skip to content

docs(fleets): fleet chat streaming, inline final reply and runner update - #206

Open
indykish wants to merge 4 commits into
mainfrom
chore/m207-fleet-chat-streaming-changelog
Open

indykish wants to merge 4 commits into
mainfrom
chore/m207-fleet-chat-streaming-changelog

Conversation

@indykish

@indykish indykish commented Sep 25, 2026 •

Copy link
Copy Markdown
Contributor

Documents the fleet chat release in agentsfleet/agentsfleet#715 (0.51.0). Merge after #715, because the API reference reads public/openapi.json from agentsfleet main.

Session notes 1

The Sep 25 changelog entry now covers the whole release: queued steers, streamed replies at a 200 millisecond cadence, lighter reasoning text, final_reply on event_complete, the bundle download retry, and the corrected bundle-failure message. Operators are told to update each agentsfleet-runner to 0.51.0, because an older runner still completes runs but shows each reply only when the run finishes.

 page                   change
 ─────────────────────  ─────────────────────────────────────────────
 changelog.mdx          Sep 25 entry: +4 bullets, runner update note,
                        API tag
 fleets/running.mdx     copyable diagnostic; bundle retry: 3 attempts
                        in 5 s, then a runner-refusal message
 fleets/troubleshooting Copy diagnostic carries safe stream counters
 runners.mdx            limit: live replies need runner 0.51.0

Values trace to agentsfleet source: 200 ms (src/runner/daemon/forwarders.zig:25), 3 attempts in 5 s (src/runner/bundle_extract.zig:48-54), 64 KiB (rustd/crates/afd_fleet/src/lease/bracket.rs:42), refusal sentence (ui/packages/app/components/domain/fleetFailureCopy.tsx:15).

verified dates are unchanged: only the edited sections were rechecked, not the whole pages (DOC-31).

Review

Review not run as a separate skill pass. The docs checker caught DOC-02 (26-word sentence) and DOC-03 (four-sentence paragraph) on fleets/running.mdx:117; both FIXED before commit 7622146.

Make

make lint            exit 0   Documentation check passed; 3 links checked
gitleaks --staged    no leaks found

🤖 Generated with Claude Code

RetriggerConfidence Score: 5/5

The PR appears safe to merge, with one non-blocking wording issue still outstanding.

Findings

  1. P2 Retry count is ambiguous ▶
Fix with agent prompt
### Issue 1
changelog.mdx:undefined-61
“Retries ... up to 3 times” suggests three retries after the first download, while the running guide says there are three attempts total. That difference makes it harder for operators to know when a failed download should stop retrying. Use “up to 3 attempts” in both places.

```suggestion
  - **A brief network drop no longer fails a run before it starts** - the runner makes up to 3 attempts to download the fleet bundle within 5 seconds.
```

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.
Summary

The PR documents fleet chat streaming, resend behavior, diagnostics, runner compatibility, and related API errors.

  • Adds release notes for streamed replies and idempotent message sends.
  • Updates fleet guidance for failed runs, unsent messages, and runner upgrades.

Reviews (3) · Last reviewed commit: "docs(fleets): UZ-AGT-016 covers another ..."

indykish and others added 2 commits September 25, 2026 07:32
…r update

The Sep 25 changelog entry gains `final_reply` on `event_complete`, the
200 millisecond reply cadence, the lighter reasoning text, the bundle
download retry, and the corrected bundle-failure message. It also tells
operators to update each `agentsfleet-runner` to 0.51.0, because an older
runner shows replies only at completion. The running page states the retry
count and budget, and the runners page lists the update as a limit.

Pairs with agentsfleet/agentsfleet#715.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Comment thread changelog.mdx

- **A stopped reply has a plain message** — chat no longer displays a runner error name as if it explained what happened. The cause remains available in the event diagnostic.
- **Streaming tool calls reach the fleet** — a model that asks to use a tool while streaming can complete the tool call and continue its answer.
- **A brief network drop no longer fails a run before it starts** — the runner retries the fleet bundle download up to 3 times within 5 seconds.

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P2 Retry count is ambiguous
“Retries ... up to 3 times” suggests three retries after the first download, while the running guide says there are three attempts total. That difference makes it harder for operators to know when a failed download should stop retrying. Use “up to 3 attempts” in both places.

Suggested change
- **A brief network drop no longer fails a run before it starts** — the runner retries the fleet bundle download up to 3 times within 5 seconds.
- **A brief network drop no longer fails a run before it starts** — the runner makes up to 3 attempts to download the fleet bundle within 5 seconds.
Prompt To Fix With AI
This is a comment left during a code review.
Path: changelog.mdx
Line: 33

Comment:
**Retry count is ambiguous**
“Retries ... up to 3 times” suggests three retries after the first download, while the running guide says there are three attempts total. That difference makes it harder for operators to know when a failed download should stop retrying. Use “up to 3 attempts” in both places.

```suggestion
  - **A brief network drop no longer fails a run before it starts** — the runner makes up to 3 attempts to download the fleet bundle within 5 seconds.
```

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

indykish and others added 2 commits September 28, 2026 08:34
…ration_id

Chat keeps every unsent message with its own Resend and Dismiss, across a
reload and in other tabs, and tells a refused send from one that got no
answer. The messages endpoint answers a repeated operation_id with the
first message's event, even after a pause, and refuses the same id with
a different message with 409 UZ-AGT-016. The overview's chat section
also drops the paw indicator and describes the Thought clock and tool
rows that replaced it.

Pairs with agentsfleet PR #717.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…stop

Brings PR #206 in line with the final agentsfleet#717: a reused
operation_id from another sender is refused (a workspace's API keys are
one sender), a repeat is answered after a stop, pause, spent budget or
full backlog, and an id sent before the release is not recognized after
it. Chat keeps Resend beside Sign in, keeps each person's unsent messages
to them on a shared browser, and names the 8,192-byte limit.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant