Plan title duplicated: appears in plan card AND as separate agent response message

[adolfousier]

Bug Report

Reported by: @leshchenko1979 (Alexey), 2026-07-27

Problem

When an agent turn runs with an active plan, the plan title appears twice:

1. Inside the persistent plan card message (correct, rendered by plan_card.rs:render_plan_card_html)
2. As the first line of the agent's text response, sent as a separate message by deliver_final_response in delivery.rs

The user sees: plan card (with title + prose expandables + checklist) followed by a second message that starts with the plan title again and then the agent's actual response text.

Root Cause

After #580 and #621 moved plan chrome (title, prose, checklist) from the per-turn flow block to the persistent plan card:

• flow_chrome.rs:713 correctly sets plan_title: None, prose: None, checklist: None in FlowSections
• plan_card.rs:render_plan_card_html correctly renders the title + prose expandables + checklist in one persistent message
• But delivery.rs:deliver_final_response sends the agent's raw response.content as a new message with NO stripping of plan title

The LLM's response includes the plan title because context.rs:870 injects [ACTIVE PLAN REMINDER] Plan: "{title}" into the system prompt, and the model echoes the title back in its output.

Evidence

Alexey's screenshot (2026-07-27 13:46-13:47):

• Message 13:46: Plan card with expandable sections (Context, Implementation steps, Out of scope, Files touched) and title at top
• Message 13:47: "Business Tinder Help System Upgrade" + agent text response as a separate message (title duplicated)

Affected Code Paths

• src/channels/telegram/delivery.rs:38 (deliver_final_response) sends response.content without plan-title stripping
• src/channels/telegram/handler.rs:3529 refresh_plan_card called mid-round
• src/channels/telegram/handler.rs:4008 refresh_plan_card called at turn end
• Agent text delivery happens between these two refresh calls

Proposed Fix

In deliver_final_response (or a helper it calls), after sanitization and before sending:

1. Load the active plan title for the session
2. Strip leading lines from response.content that match the plan title (with or without prefix / bold markup)
3. Send the cleaned text

This ensures the plan card is the single surface carrying the title, matching the #580/#621 architecture intent.

Related

• audit(plan): multi-turn plan execution fragments the flow chrome — duplicate checklists, stale tracking, split blocks #580: Move plan chrome to persistent plan card
• eval(phase1): compaction-fidelity harness + dataset format #621: Strip prose from flow block (partially fixed; title in agent response still leaks)

[adolfousier]

Audit of the in-flight fix, and what it needs

The approach is right — address the cause (model echoes what it is shown) and the symptom (strip on delivery) together. Four problems with the implementation as written.

1. The stripper likely misses the reported case

The reminder injects 📋 Plan: "{title}". The stripper trims #, * and ~ only — not the 📋, not a Plan: label, not the surrounding quotes. So if the model echoes the reminder's OWN formatting, which is the most likely echo of all, the comparison never matches.

Exact-string-equality after a narrow trim is brittle exactly where the model is most likely to copy verbatim.

2. Editing plans are not covered

It strips only when status == Active. #621 folded title and prose into the card in BOTH Editing and Active, so every plan still being drafted keeps the duplicate.

3. The reminder text asserts something false off Telegram

context.rs now says "the plan card is already shown above your message". format_plan_reminder is shared context: TUI, CLI, Discord and WhatsApp have no plan card. On those surfaces the model is told a card exists that does not, and may suppress a title the user never sees rendered anywhere.

4. Untestable and silent

strip_leading_plan_title loads the plan from disk inside itself, so it cannot be exercised without a session on disk, and no tests were added. When the match misses it does nothing and logs nothing — indistinguishable from the bug being unfixed.

Also a disk read per final response, on the hot path, even when no plan exists.

What I am changing

• Pure matcher taking the title as an argument, so the shapes can be tested directly: markdown headings, emphasis, the 📋, a Plan: label, surrounding quotes.
• Normalize both sides rather than exact-matching one, and require the WHOLE first line to reduce to the title — a line that merely mentions it is left alone.
• Editing and Active both covered.
• Surface-neutral reminder wording: the plan is already displayed to the user, without claiming a card exists.

Keeping the reminder's title itself. The model needs to know which plan it is executing, and removing it to prevent an echo would trade a display bug for lost grounding.

[adolfousier]

Fixed in 10d6ac51, replacing the in-flight implementation.

Both ends: the reminder tells the model not to repeat the title, and delivery drops it when it does anyway. The instruction alone is not enough — a prompt rule loses often enough that the filter has to exist too.

All four audit findings addressed:

The matcher would have missed the reported case. The reminder injects 📋 Plan: "…" and the original trimmed only #, *, ~. An echo of the reminder's own formatting — the likeliest echo there is — never matched. Both sides are now normalized (emoji, Plan: label, quotes, headings, emphasis), and the WHOLE first line must reduce to the title, so a sentence merely mentioning it survives.

Editing plans are covered. #621 folded title and prose into the card in both states; limiting to Active left every plan still being drafted duplicated.

The reminder no longer claims a card exists. format_plan_reminder is shared context — TUI, CLI, Discord and WhatsApp have no plan card, so the model was told about one that does not exist and might suppress a title the user never sees rendered.

The matcher is pure and tested. It takes the title rather than loading it, so the shapes are exercised directly: the reminder's formatting, markdown headings and emphasis, an unrelated opening, a mid-sentence mention, a title appearing later in the text, an empty title, a different plan's title, and the degenerate case where the reply IS only the title.

9 tests. Full suite 5624 passing, clippy clean.

The reminder keeps the title. The model needs to know which plan it is executing, and removing it purely to prevent an echo would trade a display bug for lost grounding.

[adolfousier]

Filed #869 for the underlying architectural issue: the followup tap and unrecognized callback routing both use run_resume_turn() (simple one-shot, no streaming/typing/plan card/flow chrome) instead of the existing resume_session() full streaming pipeline. This is why the delivery pipeline's strip_echoed_plan_title never runs for followup taps — they bypass deliver_final_response entirely.