## Problem to solve In ambient (@mention) sessions, Duo Developer sometimes posts its findings to the discussion itself, and the platform then delivers the agent's final answer to the same discussion a few seconds later. The result: two near-duplicate comments, same author (the flow service account), different wording. This is not sporadic. Between 2026-09-04 and 2026-09-08, 18 of 68 platform-delivered final answers (26%) from the `duo-developer-gitlab-org` service account were preceded within 5 minutes by a service-account-authored note on the same issue or MR. 7 of the 18 were hand-checked: all 7 were genuine duplicates. This is a heuristic over one service account and four days, but it is an order of magnitude above "occasionally". Production example (2026-09-02, session `6856010`, work item #2803): - Agent-posted reply (no "View session" link): https://gitlab.com/gitlab-org/modelops/applied-ml/code-suggestions/ai-assist/-/work_items/2803#note_3773948235 - Platform-delivered reply (with "View session" link, 22 seconds later): https://gitlab.com/gitlab-org/modelops/applied-ml/code-suggestions/ai-assist/-/work_items/2803#note_3773949759 ## What actually happens (verified from the session trace) Exported trace for session `6856010` (`/api/v4/ai/duo_workflows/workflows/6856010/trace.jsonl`), flow `developer` (production pins `^3.0.0`): 1. The rendered goal already contains the delivery instruction, verbatim: *"Reply to @bastirehm in your response, not by posting a comment — your response reaches them automatically, so a comment would duplicate it."* (from the mention goal template, `ee/app/models/ai/catalog/goal_templates/developer/mention.rb` in `gitlab-org/gitlab`). 2. The agent's first action is `read_file /root/.agents/skills/glab/SKILL.md` — the skill bundled with glab v1.111.0, which the executor installs globally (`ee/lib/gitlab/duo_workflow/glab_setup.rb`). Its description mandates this: *"Whenever a task touches GitLab in any way, consult this skill first."* 3. After ~25 minutes of investigation and creating MR !6721, the agent reasons: *"The MR has been created... Let me now post a reply to the issue discussion with the findings."* 4. It finds the discussion containing the mention note, then posts via `run_command` → `glab api .../discussions/<id>/notes -F body=@/tmp/issue_reply.md` — a near-verbatim replay of the skill's "Threaded replies on issues" recipe (heredoc to a temp file, `-F body=@file`). 5. It then emits its final answer with the same content. The platform's `GitlabDuoNote` messaging adapter posts it as a threaded reply 22 seconds later. Two facts shape the problem: - **The instruction exists and has not held.** The mention goal template has carried an anti-double-post instruction since 2026-06-19 and was iterated three times (2026-06-19, 2026-07-09, 2026-08-04). The observed occurrence postdates all three iterations. - **There is no note tool to block.** The ambient developer flows expose only `read_file, find_files, list_dir, grep, create_file_with_contents, edit_file, mkdir, todo_write, run_command`. The agent posted through `run_command` + `glab api`, and shell access means it always can. Platform delivery itself is exactly-once (work item 611422, !251096): the duplicate does not come from a repeated delivery job but from the agent posting on its own. ## Where this could come from Hypotheses, with the evidence for each. They are not mutually exclusive, and none is proven in isolation. 1. **Mental-model mismatch.** The shared system prompt (`common/developer/system/1.0.0.jinja`, ai-assist) frames the agent as autonomous: *"you receive a task, execute it, and persist your results to GitLab."* In that frame the final answer reads as a status report to the operator, not as the reply. In the trace, the agent posted first and emitted its final answer second — as if the post was the delivery and the final answer the wrap-up. 8 of the 18 sampled duplicates open with a variant of *"I've posted my analysis as a reply to @X... Here's a summary."* 2. **Goal framing.** The goal is narrated in third person ("@user is talking to you in a conversation on this issue") and names a place where a conversation exists — a place the agent knows how to write to. It also ships executable reply targets: the `#note_<id>` anchor and `<message id="...">` attributes. The agent used that exact note ID to locate the discussion it posted to. The verb "Reply" maps directly onto a well-trained API action, and the prohibition ("not by posting a comment") names the very action it tries to prevent. 3. **Skill affordance at decision time.** The glab skill is loaded at the start of every GitLab-touching session, by mandate. It teaches *how* to post comments and threaded replies in detail, and says nothing about *when not to*. At the moment of decision, the skill's recipe is fresher and more actionable than a 25-minute-old sentence in the goal. 4. **Instruction distance.** At the posting decision the model sat at ~180k input tokens, after hundreds of tool calls. A control session on the same issue (same flow, same goal wording, same skill, also read the skill first) did not double-post: it was a 4-minute, read-only session (~86k tokens). Short sessions comply; long sessions with many GitLab writes fail. No compaction events were observed in the failing session, so "compaction dropped the rule" is not the explanation — recency and salience are. 5. **Reinforcing ambient context.** `gitlab-org/gitlab`'s AGENTS.md asks AI-authored comments to carry `<:robot:>` tags — it presupposes the agent posts comments; 8 of the 18 duplicates carried the tags. During its work, the agent in session `6856010` also read other flows' prompts that model the pattern "finish the work, then post a summary comment." 6. **Non-uniform delivery semantics.** Same flow, same system prompt: for @mentions the platform delivers the final answer; for assign triggers the goal tells the agent to "@mention them in a comment" — i.e. to post itself. The delivery mode exists only as prose in the goal, and the agent must infer which regime it is in. 7. **Scope drift (possibly a separate issue).** The task was "investigate". The agent implemented a fix, opened an MR, and then treated "update the issue" as the next step of a coding-task completion pattern. The double-post may be one symptom of a broader completion ritual rather than an isolated misunderstanding. ## What any fix has to live with - The two posts differ in wording each time, so text-similarity dedup cannot catch them. - Suppressing or forcing an empty final answer is unsafe with the current delivery path: when the last agent message is empty, delivery falls back to an earlier non-empty message. Observed in session `6858368`, where the delivered note was a reasoning turn that carried a `todo_write` call. - The shared system prompt is a generic building block used by chat, ambient, and other flows. Per-invocation conditionals in it create prompt variants (cache churn) and can go stale when a session is continued. The goal, by contrast, is re-rendered for every mention: each mention starts a fresh session. - Rendering tests prove a prompt contains a sentence; they do not prove behavior changes. Any prompt-level fix needs behavioral validation before it counts. ## Links / references - Production occurrence: https://gitlab.com/gitlab-org/modelops/applied-ml/code-suggestions/ai-assist/-/work_items/2803#note_3773948235 - Session trace: https://gitlab.com/gitlab-org/modelops/applied-ml/code-suggestions/ai-assist/-/automate/agent-sessions/6856010 (export: `/api/v4/ai/duo_workflows/workflows/6856010/trace.jsonl`) - Control session (no duplicate): agent session `6858368` - Mention goal template: https://gitlab.com/gitlab-org/gitlab/-/blob/master/ee/app/models/ai/catalog/goal_templates/developer/mention.rb - Shared ambient system prompt: https://gitlab.com/gitlab-org/modelops/applied-ml/code-suggestions/ai-assist/-/blob/main/ai_gateway/prompts/definitions/common/developer/system/1.0.0.jinja - glab skill provisioning: https://gitlab.com/gitlab-org/gitlab/-/blob/master/ee/lib/gitlab/duo_workflow/glab_setup.rb (pins glab v1.111.0, `glab skills install --global`) - Platform dedup fix (not related): https://gitlab.com/gitlab-org/gitlab/-/work_items/611422 and https://gitlab.com/gitlab-org/gitlab/-/merge_requests/251096 - Earlier prompt-level attempts: !6766, !6767, !6769 (drafts)