Cloud Agents: mini probes PASS, eng-shaped implement prompts die at kickoff (~3 msg, 0 tools) — 2026-09-28

[dguardi]

Where does the bug appear (feature/product)?

Cloud Agent (GitHub, Slack, Web, Linear)

Describe the Bug

Where: Cloud Agents & Automations (cursor.com/agents, scheduled/event)

New Cloud Agents on a private GitHub repo often plan briefly then die with ~3 messages and 0 tool calls. Tiny one-tool probes (echo / ls / wc / Read README) PASS. Eng-shaped implement / draft-PR prompts (~800–1.4k chars), including a one-fold feature ticket, consistently FAIL the same way (planning line, then error, tools=0). A mid-size non-implement control (Read + wc) FAILED earlier the same evening and PASSED on retest ~1h later, while implement tickets kept dying — including after a Cursor client update. ASCII-only prompts and Cloud Agents MCP limited to X + Gmail do not fix it. status.cursor.com showed Cloud Agents Operational during this.

Steps to Reproduce

1. Launch Cloud Agent on a connected private repo with: PROBE ONLY. One Shell: echo kickoff-ok → PASS (tools run).
2. Launch with a short eng-shaped implement prompt (~800–900 chars; draft one PR for a small config/UI change; no secrets) → FAIL (~3 msg, 0 tools).
3. Optionally retest a mid-size PROBE-ONLY Read+wc control — may PASS even while step 2 still FAILs.

Expected Behavior

After env ready, the first tool runs (as in the mini probes), including for normal-sized implement prompts.

Actual Behavior:
Planning/thinking text appears, then the run stops. Tool count stays 0. Empty-transcript errors also seen on some launches.

Operating System

Other

Version Information

OS: Other (Cloud Agent backend / web agents UI)

Cloud Agent backend issue (launched via cursor.com / agent API). Not IDE-version-specific. Cursor client update did not clear the implement-ticket failures.

For AI issues: which model did you use?

Default Cloud Agent model (not overridden for these probes).

For AI issues: add Request ID with privacy disabled

PASS bc- (lookup):
bc-55c25ed3-d655-5bbf-aee9-9f8cdde084c7
bc-f0da4263-c0ac-509b-96a4-2fdd193f84cf
bc-9d4274d2-786c-5876-8e44-149a2001ffdc
bc-2926dfa6-2111-529a-9e4a-913005f085ff
bc-7c75a483-78b6-5f3e-8239-e1f1bc76d9e6
bc-d24d3ec3-82ed-56a6-8634-d5fe5886ad7c
bc-7b0ae522-f2bd-538e-a068-144f44436ff2

FAIL bc- (lookup):
bc-3747951b-9028-59f3-82fd-9cd9ac942fb5
bc-195762e3-6651-5f74-932d-5f5443469c3c
bc-b8af91fd-022c-569e-a896-ac2467b8c4b0
bc-b342661d-ec2a-53b4-90e9-2e0c8cc02389
bc-1de2fc9e-b0c5-5ea6-8fb1-b4fa99cb423c
bc-df3ec4cd-6f77-545d-ba68-e2f385fe79e9

Additional Information

• Cloud Agents MCP on this account: only X and Gmail (no custom HTTP MCP).
• Unicode/special-char hygiene ruled out (ASCII-only mid/implement prompts still failed; mini Read PASS).
• Related public threads (different / partial match): Cloud Agents stuck on "Planning next moves" — env running, zero model output (2026-09-03) ; Cloud agents no longer working at all
• Do not need repo name or org name for lookup — bc- IDs should be enough.

Does this stop you from using Cursor

Sometimes - I can sometimes use Cursor

[mohitjain (Mohit)]

Hey @dguardi,

The failing runs are being started from a short (7-character) commit ID, while the ones that pass start from the default branch. The cloud agent start step can’t resolve a short commit ID right now, so it can’t find the starting point and stops right after the planning line (the ~3 messages, 0 tools you’re seeing). Prompt size and content aren’t the cause.

To unblock, either start the run from a branch (like your default branch), or pass the full 40-character commit hash instead of the short form (git rev-parse <short-sha> prints it). Since these are launched programmatically on your side, that’s where the short ID is coming from, so pointing it at a branch or the full hash should get them running.

This isn’t anything you misconfigured; it’s on our side and something we’re tracking. Mind trying one run started from your default branch to confirm it goes?

[dguardi]

Thanks @mohitjain — that matches what we saw.

Confirm probe just now from default branch (starting_ref=main, not a short SHA): PASS — tools ran, KICKOFF_OK.

bc-31f1fc5b-260f-5704-8da4-7486cb73e0d3

We’ll launch from a branch or full 40-char hash on our side going forward. Glad you’re tracking the short-ID resolve on yours.

[Colin]

Hey @dguardi,

Quick follow-up on this one: starting a cloud agent from a short commit ID now works. The short ID resolves to the full commit and the agent starts normally, so you no longer need to point runs at a branch or the full 40-character hash.

This was a change on our side, so there’s nothing to update. Let me know if you still see a run stop right after the planning line when it starts from a short commit ID.