Upstream openai/codex#29218: CLI resume drops markdown table tail from historical assistant message

[ctkm-aelf]

Mirrored upstream candidate selected to unblock the managed issue loop.

Upstream: openai#29218
Upstream URL: openai#29218
Rubric query: manual expanded repo-local intake after automated intake returned duplicates
Score: 8
Labels observed upstream: bug, TUI, CLI, session

Problem

When a Codex CLI session is resumed, a historical assistant message containing a plain fenced code block followed by a markdown table and task-list is rendered only through the fenced code block. The table and task-list tail are still present in persisted session data and in app-server --stdio thread/resume, so the likely failure boundary is the TUI resume/replay/rendering path rather than persistence.

Intake framing

• action_mode: identify
• canonical_key: CLI resume drops markdown table tail from historical assistant message openai/codex#29218
• owned_by_repo: yes, expected TUI/session replay path under codex-rs
• expected boundary: codex-rs/tui resume/replay/rendered assistant-message path, to be confirmed by design consensus
• duplicate_search_summary: searched local ChronoAIProject/codex issues for upstream number, normalized title, markdown table tail, and historical assistant message; no local duplicate found before creation
• next_action: plan

Acceptance direction

Design consensus should verify current resume rendering with a finalized assistant message containing a plain fenced code block, a markdown table, and task-list lines. If the bug reproduces or is source-explainable, produce a concrete implementation plan naming affected files and focused TUI/session tests. If current checkout is already fixed, close as triage already-fixed with evidence.

⟦AI:AUTO-LOOP⟧

[ctkm-aelf]

🤖 Minimal solver proposes TUI replay regression first

TL;DR

• What this is: Minimal-change design solver output for issue Upstream openai/codex#29218: CLI resume drops markdown table tail from historical assistant message #56.
• Current state or conclusion: The smallest boundary is the TUI resume/replay rendering path, with persistence and app-server protocol left unchanged.
• What the maintainer should do OR what the controller does next: Next consensus should prefer a focused codex-tui replay regression, then fix only replay finalization if the test reproduces the dropped tail.

---

Minimal framing

The issue source says the markdown tail still exists in persisted session data and in app-server thread/resume, so the proposed boundary is presentation only: the terminal UI replay path for historical assistant messages. AgentMarkdownCell is the source-backed transcript cell that re-renders assistant markdown for display; the plan is to test that path with a finalized assistant message containing a plain fenced code block, a markdown table, and task-list lines.

Option	Trade-off
Add focused TUI replay regression first	Smallest change; proves whether current checkout is already fixed or identifies the exact display boundary.
Change app-server or persistence	Rejected for this issue because the snapshot says those layers already preserve the full message.
Change markdown rendering globally up front	Rejected unless the replay test fails there; broader blast radius for unrelated markdown/table behavior.

Concrete next step: add one test in codex-rs/tui/src/chatwidget/tests/history_replay.rs, run just fmt and just test -p codex-tui, and only touch chatwidget/streaming.rs, markdown.rs, or markdown_render.rs if that focused test fails.

📎 Full raw codex output (archival)

---

solver: minimal
issue: 56
cluster: issue-56
verdict: propose

Recommended framing

最小可行边界是在 TUI 的历史 assistant 消息 replay/rendering 层补一个精确回归用例，并只在该用例失败时修正 AgentMessage replay 到 finalized markdown cell 的路径；issue 已说明 persisted session data 与 app-server thread/resume 仍保留尾部内容，所以不应改 persistence、protocol 或 rollout schema。当前 checkout 的 ThreadItem::AgentMessage replay 会进入 on_agent_message_item_completed，该函数对完整文本执行 parse_assistant_markdown 并最终用 agent markdown renderer 渲染，因此首要缺口是没有测试锁住“plain fenced code block 后接 markdown table 与 task-list”的恢复显示行为。

Concrete plan

• Files: codex-rs/tui/src/chatwidget/tests/history_replay.rs — add one replay regression that feeds a finalized ThreadItem::AgentMessage through ReplayKind::ResumeInitialMessages with a plain fenced code block, then a markdown table, then task-list lines; drain rendered history and assert the code block, table cells, and task-list tail are all present.
• Files: codex-rs/tui/src/chatwidget/snapshots/... — add or update the corresponding insta snapshot if the test renders user-visible transcript output rather than plain extracted text assertions.
• Files: codex-rs/tui/src/chatwidget/streaming.rs — only if the new replay test fails, adjust on_agent_message_item_completed so replayed finalized assistant messages are inserted as a source-backed AgentMarkdownCell (or otherwise flush any active stream/table holdback before turn completion) instead of allowing a stale streamed prefix to mask the finalized full message.
• Files: codex-rs/tui/src/markdown.rs or codex-rs/tui/src/markdown_render.rs — only if the failure is reproduced in direct markdown rendering, add the smallest parser/render fix for a normal fenced code block followed by a table and task list; avoid changing markdown-fence unwrapping for non-markdown fences unless the failing test proves that is the boundary.
• LOC delta: ~+45 / -0 for regression-only if current code passes; ~+70 / -10 if a small replay finalization fix is needed.
• Tests to add/modify: one focused TUI history replay test in codex-rs/tui/src/chatwidget/tests/history_replay.rs; run just fmt, then just test -p codex-tui. If snapshots are produced, review *.snap.new and accept with cargo insta accept -p codex-tui only for the intentional TUI diff.
• Governance/ruleset change (if any): none. This stays within existing TUI snapshot/regression-test rules and does not touch CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR, config schema, protocol APIs, or host .refactor-loop/ production authority.
• Migration path: no migration needed; persisted sessions already contain the full assistant message according to the issue source snapshot.

Risks

• A text-only assertion may miss a visual table layout regression; prefer a small snapshot if the existing test harness can snapshot the replay cell cleanly.
• If the failure only appears with an actual resumed rollout/app-server initial page, a direct replay_thread_item test may be too narrow; then extend the test to replay_thread_turns with a completed turn rather than changing app-server code.
• Rendering all finalized replay messages through a new path could disturb separators or /copy state; if a code fix is required, keep it limited to replay finalization and assert last_agent_markdown_text still sees the full visible markdown.

Escalation triggers (if any)

• none

Reasoning trace (internal — short, for meta-judge)

• Reference frame: regression testing at the suspected ownership boundary before changing data contracts; preserve source-of-truth data and repair only the presentation layer if reproducible.
• Why this is the minimum: the issue states storage and app-server resume still include the tail, and current code routes replayed ThreadItem::AgentMessage through on_agent_message_item_completed, parse_assistant_markdown, and the agent markdown renderer. The missing artifact is a focused replay test for exactly the reported markdown shape.
• What I considered but rejected: changing rollout persistence, app-server thread/resume, or markdown table detection globally; those are wider than the cited failure boundary and risk unrelated protocol/rendering changes.
• What I cannot decide alone: whether the current checkout already passes the exact upstream reproduction without adding the regression test, because the issue snapshot does not include the original transcript JSONL and no local audit artifact was provided.

⟦AI:AUTO-LOOP⟧
SOLVER_DONE:minimal:propose:add focused TUI resume replay regression, then fix only replay finalization if it reproduces the dropped markdown tail

⟦AI:AUTO-LOOP⟧

[ctkm-aelf]

🤖 Structural solver result: current checkout appears already fixed

TL;DR

• What this is: structural solver output for issue Upstream openai/codex#29218: CLI resume drops markdown table tail from historical assistant message #56, focused on the TUI resume/replay rendering boundary.
• Current state or conclusion: source inspection shows replayed assistant messages already flow through the canonical source-backed AgentMarkdownCell renderer, so no new abstraction is justified.
• What the maintainer should do OR what the controller does next: controller can triage this as already-fixed/false-positive unless it wants a small regression-test-only PR.

---

Structural finding

The current checkout already has the clean boundary this bug needs. AgentMarkdownCell is the finalized assistant-message history cell that stores raw markdown source and re-renders it at display time; this matters because resume replay should not preserve stale wrapped lines or stop after a code block.

The replay path maps a historical ThreadItem::AgentMessage through the same completed-message path used for non-streamed final assistant messages: replay_thread_item -> handle_thread_item -> on_agent_message_item_completed -> finalize_completed_assistant_message. StreamController::finalize then returns the complete accumulated raw source before reset, and the consolidated cell renders that source through the shared agent markdown renderer.

Option	Trade-off
Triage as already fixed	Matches current source evidence; no code churn.
Add one regression-test-only PR	Costs about +35 LOC; gives durable proof that a plain fenced code block followed by a table and task-list tail survives resume replay.
Add a new replay renderer abstraction	Not recommended; it would duplicate the existing source-backed finalized-message owner and create a second rendering boundary.

Recommended next step: if the controller continues to implementation, keep it to codex-rs/tui/src/chatwidget/tests/history_replay.rs and assert the replayed assistant message contains the code block, table body/header content, and both task-list tail lines after ReplayKind::ResumeInitialMessages replay.

📎 Full raw codex output (archival)

---
solver: structural
issue: 56
cluster: issue-56
verdict: abstain
---

## PROJECT_RULES clause violated (quoted verbatim)
> Snapshot tests
>
> This repo uses snapshot tests (via `insta`), especially in `codex-rs/tui`, to validate rendered output.
>
> **Requirement:** any change that affects user-visible UI (including adding new UI) must include
> corresponding `insta` snapshot coverage (add a new snapshot test if one doesn't exist yet, or
> update the existing snapshot). Review and accept snapshot updates as part of the PR so UI impact
> is easy to review and future diffs stay visual.

## Recommended framing
当前 checkout 看起来已经满足这个问题要求的结构边界：resume replay 不应该拥有独立的 markdown 截断逻辑，而应该把历史 `ThreadItem::AgentMessage` 投回同一个 completed assistant-message 管线，让 `AgentMarkdownCell` 保存完整 raw source 并按当前宽度重新渲染。源码检查显示 replay 已经经由 `replay_thread_item` -> `handle_thread_item` -> `on_agent_message_item_completed` -> `finalize_completed_assistant_message` -> `StreamController::finalize`，而 `StreamController::finalize` 在 reset 前 `take` 完整 `raw_source`，最终由 source-backed `AgentMarkdownCell` 渲染。因此结构性结论是当前 checkout 已经 source-explainable fixed；只需要补一个窄的 resume replay 回归快照/断言，防止以后重新引入“plain fence 后 tail 丢失”。Reference-frame note: single-source-of-truth replay architecture; persisted transcript text should feed one canonical finalized-message renderer, not a parallel resume renderer.

## Concrete plan
- New abstractions (if any): none. Existing abstraction is `AgentMarkdownCell`, the finalized assistant message history cell that stores raw markdown source and re-renders it through `render_markdown_agent_with_links_and_cwd`; existing replay adapter is `ChatWidget::replay_thread_item`.
- Files:
  - `codex-rs/tui/src/chatwidget/replay.rs`: no implementation change expected; current mapping of `ThreadItem::AgentMessage` to `on_agent_message_item_completed(..., from_replay)` is the correct boundary.
  - `codex-rs/tui/src/chatwidget/streaming.rs`: no implementation change expected; current `finalize_completed_assistant_message` feeds a whole non-streamed completed message into the stream controller when no live stream exists.
  - `codex-rs/tui/src/streaming/controller.rs`: no implementation change expected; current `StreamController::finalize` returns the complete accumulated raw source before reset.
  - `codex-rs/tui/src/history_cell/messages.rs`: no implementation change expected; current `AgentMarkdownCell::display_hyperlink_lines` re-renders from stored source at the requested width.
  - `codex-rs/tui/src/chatwidget/tests/history_replay.rs`: add one regression test for `ReplayKind::ResumeInitialMessages` with assistant markdown containing a plain fenced code block, followed by a markdown table and task-list lines. Assert rendered history contains code text, table body/header content, and both task-list tail lines.
  - `codex-rs/tui/src/chatwidget/snapshots/...`: add/update one `insta` snapshot only if the regression test uses snapshot output rather than direct content assertions.
- LOC delta: ~+35 / -0 if content assertions only; ~+35 / -0 plus one snapshot if using `insta`.
- Tests to add:
  - `codex-rs/tui/src/chatwidget/tests/history_replay.rs::replayed_agent_message_preserves_tail_after_plain_fenced_code_block`: constructs a replayed assistant final answer with ` `` `\n...\n```\n\n| Item | Status |\n| --- | --- |\n| table-tail | present |\n\n- [x] task tail kept\n- [ ] task tail pending`, drains `InsertHistoryCell` output, flattens spans, and asserts all tail markers are present. This is behavior coverage at the resume replay boundary, not a renderer-only unit test.
- Schema/protocol changes (if any): none; `thread/resume` and `thread/read` payloads are not changed.
- Governance/ruleset changes (if any): none.
- Runtime cost: 0 new latency hops, 0 new steady-state allocations. The regression test exercises existing allocations only: one replayed message string, one `StreamController` raw-source buffer, and one source-backed `AgentMarkdownCell` string already present today.
- If the structural answer is issue decomposition, use the controller-private `IssueDecompositionPlan` boundary and child body artifacts; forbid solver/judge/worker GitHub lifecycle, public issue factories, and any `wakeup_plan.py` decompose action/projection.
- Do not author helper-owned decomposition tracking comments or child fingerprint lines; the checked-in #403 helper is the only writer for the parent tracking grammar and per-child idempotency fingerprints.

## Risks
- The main tradeoff is closing the work as already-fixed based on source inspection rather than a live CLI resume repro. This is acceptable only if the regression test is added or an equivalent local vt100 replay proof is captured by the implementer/controller.
- A snapshot could be brittle because table layout depends on width. Keep the regression at a fixed test width and assert semantic tail content if exact table borders are not the behavior under test.
- Avoid over-engineering: do not add a new resume markdown renderer, parser abstraction, or replay-specific cell. The existing source-backed `AgentMarkdownCell` is already the structural owner.

## Escalation triggers (if any)
- none

## Reasoning trace (internal — for meta-judge)
- Why this structure beats a documented exception: A documented exception would permit replay to render differently from live/finalized assistant messages. The existing structure eliminates that class by routing replayed assistant messages through the canonical completed-message path and preserving raw source for reflow.
- Why I picked this abstraction over alternatives: `AgentMarkdownCell` already has the exact invariant needed: finalized assistant messages store raw markdown source and render with the common agent renderer. Adding a new helper would have only one real caller and would violate the repo preference against one-off helpers.
- What I cannot decide alone: Whether the managed loop wants a code PR containing only the regression test, or whether source evidence is enough for triage `already-fixed`. I can decide that no implementation abstraction change is justified on this checkout.

⟦AI:AUTO-LOOP⟧
SOLVER_DONE:structural:false-positive:current checkout routes replayed assistant messages through source-backed AgentMarkdownCell; add one focused replay regression test if a PR is still desired

⟦AI:AUTO-LOOP⟧

[ctkm-aelf]

🤖 Delete solver abstains from deletion for resume markdown-tail bug

TL;DR

• What this is: Delete-framing design solver output for issue Upstream openai/codex#29218: CLI resume drops markdown table tail from historical assistant message #56.
• Current state or conclusion: No deletion or collapse is justified because CLI resume transcript replay is an active user-facing path.
• What the maintainer should do OR what the controller does next: Continue with minimal/structural planning for a focused TUI replay regression test and fix, or triage already-fixed if that test passes unchanged.

---

Delete-solver finding

The suspected boundary is the TUI replay/rendering path, not a removable subsystem. ChatWidget::replay_thread_turns is the resume replay entry point that rehydrates persisted turns into the interactive transcript, and on_agent_message_item_completed is the completed assistant-message path reused for replayed ThreadItem::AgentMessage items. Removing either path would break codex resume transcript display instead of fixing the markdown tail loss.

Option	Outcome	Trade-off
Delete replay/rendering path	Rejected	Breaks resumed CLI transcript display.
Collapse into another path now	Rejected	No simpler equivalent path was found that preserves replay semantics and live-side-effect avoidance.
Keep path and add focused coverage	Recommended next step	Tests the reported code-block/table/task-list tail shape directly.

The next concrete check should replay a historical assistant message containing a plain fenced code block followed by a markdown table and task-list lines, then assert the table/task-list tail appears in the TUI history cell/display output. If that passes on the current checkout, the controller can close as already-fixed with evidence; if it fails, the fix should stay scoped to replay/rendering rather than persistence.

---

📎 Full raw codex output (archival)

---
solver: delete
issue: 56
cluster: issue-56
verdict: abstain
---

## Classification
e

## Recommended action
删除/折叠不成立：这个问题指向 CLI/TUI resume 的历史 assistant message 回放与渲染路径，属于真实用户可见能力，不是死代码、孤儿功能或可由现有更简单路径直接替代的抽象。当前证据显示持久化和 app-server `thread/resume` 仍保留 markdown tail，风险边界更可能在 `codex-rs/tui/src/chatwidget/replay.rs` 将 `ThreadItem::AgentMessage` 回放到 `on_agent_message_item_completed`，以及 `codex-rs/tui/src/chatwidget/streaming.rs` 的非流式 finalized assistant message 渲染/flush 路径。建议让 minimal/structural solver 规划一个小修复和聚焦测试：用包含 plain fenced code block、markdown table、task-list tail 的历史 assistant message 覆盖 `ResumeInitialMessages` 回放，断言 table/task-list tail 出现在 history cell/display output 中；若当前 checkout 已经通过该测试，再按 already-fixed triage 处理。

## Concrete plan (if propose)
- Files to delete: none
- Caller migrations: none
- Tests to delete: none
- LOC delta: 0
- Philosophy/CLAUDE.md/SPEC/Tier changes (if any): none

## Reverse-evidence (why this is safe to delete)
- Not safe to delete: `codex-rs/tui/src/app/thread_routing.rs` calls `ChatWidget::replay_thread_turns` for `ReplayKind::ResumeInitialMessages`, so removing replay would break resumed CLI transcript display.
- Not safe to delete: `codex-rs/tui/src/chatwidget/replay.rs` maps persisted `ThreadItem::AgentMessage` into the same `on_agent_message_item_completed` path used by completed assistant messages, preserving status/copy-history behavior and avoiding live-only side effects during replay.
- Not safe to delete: `codex-rs/tui/src/chatwidget/streaming.rs` uses `finalize_completed_assistant_message` to render non-streamed completed messages when no stream controller exists, which is exactly the shape used by replayed historical assistant messages.
- Not safe to delete: `codex-rs/tui/src/history_cell/messages.rs` documents `AgentMarkdownCell` as the finalized source-backed cell needed for width-aware re-rendering, including markdown tables, on terminal resize.
- Not safe to delete: `parse_assistant_markdown` is shared by live assistant completion, replay/transcript rendering, and resume picker preview; its tests show non-directive markdown is preserved while only app git-action directives are stripped/reformatted.
- Public API break check: no deletion proposed. The affected code is internal TUI crate surface, but deleting it would break the external user workflow `codex resume` transcript rendering.
- External repo dependency check: no external repositories were touched or required for this abstain decision.
- Tests covering the path are load-bearing: `codex-rs/tui/src/chatwidget/tests/history_replay.rs` already verifies resumed initial messages render assistant/user history; it lacks the specific code-block/table/task-list regression case and should be extended, not deleted.

## Risks
- If the upstream bug is already fixed elsewhere in the current checkout, a new regression test may pass and the issue should be closed as already-fixed rather than fixed by code.
- If the markdown tail is dropped before `ThreadItem::AgentMessage` reaches TUI replay, the minimal fix may belong in app-server protocol thread history reconstruction rather than TUI rendering; the issue snapshot says app-server `thread/resume` retains the tail, but this solver did not reproduce end-to-end.
- If the actual failure is terminal scrollback insertion rather than history-cell rendering, the right fix may be in initial history replay buffering/insertion rather than `on_agent_message_item_completed` itself.

## Escalation triggers (if any)
- none

## Reasoning trace (internal — for meta-judge)
- Reference-frame note: deletion-first / YAGNI is applicable, but reliability-preserving refactoring says not to delete an active state-rehydration path that owns a user-visible recovery workflow unless a replacement with equivalent semantics exists.
- Why I claim this can be deleted: I do not claim deletion is safe. Classification is e: genuinely needed/right-sized enough for this issue, with no current deletion/collapse dependency.
- What I checked to verify safety: searched TUI/app-server/core resume, history, assistant message, and markdown paths; read `chatwidget/replay.rs`, `thread_transcript.rs`, `chatwidget/streaming.rs`, `history_cell/messages.rs`, `git_action_directives.rs`, `resume_picker.rs`; checked call sites for `replay_thread_turns`, `on_agent_message_item_completed`, `AgentMarkdownCell`, `AgentMessageCell`, and `parse_assistant_markdown`; checked git log for `chatwidget/replay.rs`, showing it was extracted as a behavior-preserving replay module in `8fe0ecb045`.
- What I cannot decide alone: whether the current checkout reproduces the exact upstream #29218 rendering loss, because this delete role must not write code and should not force a deletion-oriented fix. Minimal/structural consensus should run or add the focused replay test.

⟦AI:AUTO-LOOP⟧
SOLVER_DONE:delete:abstain:resume assistant-message replay/rendering is a live CLI transcript capability; no deletion or collapse is justified, add focused TUI regression coverage instead

⟦AI:AUTO-LOOP⟧

[ctkm-aelf]

🤖 Meta-judge triages issue #56 as already fixed

TL;DR

• What this is: Design-consensus meta-judge decision for issue Upstream openai/codex#29218: CLI resume drops markdown table tail from historical assistant message #56.
• Current state or conclusion: The current checkout already routes resumed historical assistant markdown through the source-backed AgentMarkdownCell, so this is triage-only already-fixed rather than an implementation dispatch.
• What the maintainer should do OR what the controller does next: The controller should close or label this as triage only, with no implement worker dispatched unless a new reproducer is provided.

---

Consensus decision

The three solver outputs do not authorize implementation. Minimal proposed a regression-first plan, structural marked the current checkout as source-explainable already fixed, and delete abstained because resume replay is a live CLI transcript feature that should not be removed.

AgentMarkdownCell is the finalized assistant-message history cell that stores raw markdown source and re-renders it at the current terminal width. That matters here because a resumed historical ThreadItem::AgentMessage should flow through the same completed-message renderer as a live completed assistant message, not through a separate resume-only renderer that could stop at the first fenced block.

Option	Decision	Trade-off
Triage as already-fixed	Chosen	Matches source inspection and avoids dispatching implementation from a mixed propose/false-positive/abstain solver set.
Add a regression-test-only PR	Not authorized by this judge	Useful future hardening, but not supported by implementation-bearing consensus in this round.
Change replay, persistence, or app-server protocol	Rejected	The issue snapshot says persisted data and app-server resume already retain the markdown tail.

Evidence checked in the current checkout:

• codex-rs/tui/src/chatwidget/replay.rs maps historical ThreadItem::AgentMessage into on_agent_message_item_completed during ReplayKind::ResumeInitialMessages, so resume replay does not use a separate truncating markdown path.
• codex-rs/tui/src/chatwidget/streaming.rs concatenates all text content, parses assistant markdown once, and finalizes the whole visible markdown when non-empty.
• codex-rs/tui/src/streaming/controller.rs takes the complete raw_source before reset and returns it for consolidation.
• codex-rs/tui/src/history_cell/messages.rs defines AgentMarkdownCell as source-backed and re-renders from stored markdown through the shared agent markdown renderer.

New evidence required before reopening: provide a current-session or persisted-rollout reproducer showing that the table/task-list tail is present in the replayed ThreadItem::AgentMessage text but absent after AgentMarkdownCell rendering on this checkout.

---

📎 Full raw codex output (archival)

---
issue: 56
cluster: issue-56
convergence_round: 1
solver_verdicts:
  minimal: propose
  structural: false-positive
  delete: abstain
decision: triage
action_mode: triage
canonical_key: openai/codex#29218
next_action: triage
---

## Decision
Triage-only consensus: 当前 checkout 对这个 issue 的核心边界已经是 source-explainable fixed。minimal 提议先补 TUI resume replay 回归并仅在复现时修 replay finalization；structural 明确给出 false-positive/already-fixed 证据，指出 replayed `ThreadItem::AgentMessage` 已进入 canonical completed assistant-message path 并由 source-backed `AgentMarkdownCell` 保存完整 raw markdown；delete abstain 也确认 resume replay/rendering 是活跃能力、不能删除，并同意下一步只应是覆盖验证。因为三者没有形成 implementation-bearing 共识，且 source 检查支持 structural 的 already-fixed 判断，本轮不授权实现分发，只允许 triage 关闭或标记。

Reference-frame note: single-source-of-truth replay rendering is the applicable frame; persisted transcript text should flow through one finalized assistant markdown renderer instead of a parallel resume-only renderer.

## If triage-only consensus
- Resolution: already-fixed
- Canonical key: openai/codex#29218
- Existing issue/PR/branch/commit evidence: local checkout `357e6fb7a0` on branch `consensus-rnd/issues`; source evidence checked in `codex-rs/tui/src/chatwidget/replay.rs`, `codex-rs/tui/src/chatwidget/streaming.rs`, `codex-rs/tui/src/streaming/controller.rs`, and `codex-rs/tui/src/history_cell/messages.rs`.
- New evidence required before reopening: provide a current-session or persisted-rollout reproducer showing the table/task-list tail is present in the replayed `ThreadItem::AgentMessage` text but absent after `AgentMarkdownCell` rendering on this checkout.
- Controller action: no implementation dispatch; close or label as triage only.

## Evidence Summary
- `codex-rs/tui/src/chatwidget/replay.rs` maps historical `ThreadItem::AgentMessage` through `on_agent_message_item_completed` during `ReplayKind::ResumeInitialMessages`, so resume replay does not use a separate truncating markdown path.
- `codex-rs/tui/src/chatwidget/streaming.rs` concatenates all text content, parses assistant markdown once, and calls `finalize_completed_assistant_message` with the full visible markdown when non-empty.
- `codex-rs/tui/src/streaming/controller.rs` finalizes by taking the complete `raw_source` before reset and returning it for consolidation.
- `codex-rs/tui/src/history_cell/messages.rs` defines `AgentMarkdownCell` as the finalized source-backed assistant message cell and re-renders from stored `markdown_source` via `render_markdown_agent_with_links_and_cwd` at display width.
- Existing coverage in `codex-rs/tui/src/chatwidget/tests/history_replay.rs` verifies resumed initial assistant/user messages render, but no exact plain-fence/table/task-list regression was found by bounded search. That gap is acceptable as future hardening, not implementation authorization from this mixed solver set.

## Round audit trail (links to local artifacts)
- solver-minimal: .refactor-loop/logs/phase9-issue56-r1-minimal.log
- solver-structural: .refactor-loop/logs/phase9-issue56-r1-structural.log
- solver-delete: .refactor-loop/logs/phase9-issue56-r1-delete.log
- work-unit-producer: manual-issue (prompt-only provenance)
- work-unit-source-ref: gh-issue-56
- action_mode: triage
- canonical_key: openai/codex#29218
- duplicate_search_summary: issue snapshot says local ChronoAIProject/codex issues were searched for upstream number, normalized title, `markdown table tail`, and `historical assistant message`; no local duplicate found before issue creation. This judge did not perform additional duplicate searches outside the three solver artifacts and router-injected snapshot.
- branch: consensus-rnd/issues
- PR: none
- commit: 357e6fb7a0
- unique_diff_status: not-applicable
- tests_verification: not run; this was a design-consensus arbitration with bounded source inspection, and no code changes were authorized.
- resolution: already-fixed
- next_action: triage

⟦AI:AUTO-LOOP⟧
META_JUDGE_DONE:triage:already-fixed:current checkout replays historical assistant markdown through source-backed AgentMarkdownCell, so no implementation dispatch is authorized

⟦AI:AUTO-LOOP⟧