Preserve agentId anchors in parallel-Task stitch + tool-param UI fix#115
Merged
Preserve agentId anchors in parallel-Task stitch + tool-param UI fix#115
Conversation
Parallel `Task` tool_uses emit a chain of assistant tool_use entries whose sibling tool_result anchors each reference a different subagent via agentId. Variant 2 of `_stitch_tool_results` classified the intermediate assistant subtrees as "dead-end" and let the walker mark their descendants — including agentId-bearing tool_results — as skipped, so only the outermost subagent session had an attachment point the traversal could find. Fix: add `_collect_agent_anchors` to walk an assistant subtree for UserTranscriptEntry descendants with agentId, and surface those anchors into the stitch result alongside the assistant dead-ends. The anchors still end up in `skipped` via `_collect_descendants` (since they're descendants of the assistant), but now they also live in the DAG-line's chain — which is what `build_session_tree` reads when choosing attachment points for subagent sessions. Regression tests cover the 2-teammate case, 3-teammate case (the experiments/worktrees shape), and an end-to-end splice verifying a subagent entry is actually reached through the anchor. Fixes issue #91 symptom: only 2 of 6 subagents rendering in a teammates session. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The generic tool renderer (render_params_table) wraps long parameter values in <details class='tool-param-collapsible'>. The CSS hid the entire <summary> on open, which also removed the browser's default disclosure arrow — leaving no visible way to collapse the expanded value. Wrap the preview text in <span class='tool-param-preview'> so the [open] rule can hide just the preview, not the whole summary. The summary (with its list-item marker / arrow) stays visible, and a "collapse" hint via ::after makes the click target obvious. Covers both call sites in tool_formatters.py: structured (JSON) and simple (string) long values. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info⚙️ Run configurationConfiguration used: defaults Review profile: CHILL Plan: Pro Run ID: 📒 Files selected for processing (1)
✅ Files skipped from review due to trivial changes (1)
📝 WalkthroughWalkthroughThis change preserves user entries that carry Changes
Estimated code review effort🎯 4 (Complex) | ⏱️ ~50 minutes Possibly related PRs
Suggested labels
Poem
🚥 Pre-merge checks | ✅ 3✅ Passed checks (3 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
🧹 Nitpick comments (3)
claude_code_log/html/templates/components/message_styles.css (1)
734-745: Minor: "collapse" hint is hardcoded in CSScontentand not localizable.The
summary::after { content: "collapse" }literal bypasses any i18n pipeline — unlike text emitted via the Python templates, CSScontentstrings can't be translated at render time. If localization is on the roadmap, consider sourcing this label from a CSS custom property set per-locale (e.g.,content: var(--label-collapse, "collapse")) or emitting the hint as a real element in the template instead. For a single-language dev tool this is fine as-is.🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@claude_code_log/html/templates/components/message_styles.css` around lines 734 - 745, The CSS uses a hardcoded string in .tool-param-collapsible[open] > summary::after (content: "collapse") which prevents localization; change it to use a CSS custom property (e.g., content: var(--label-collapse, "collapse")) or move the hint into the template as a real element so the text can be produced by the i18n pipeline; update the stylesheet rule for summary::after and ensure templates set --label-collapse per locale (or render an element inside summary that gets translated) so the hint becomes localizable.claude_code_log/dag.py (1)
389-400: Variant 2 anchor preservation looks correct.
extracted_anchorsare appended todead_endsand timestamp-sorted alongsideuser_dead + assistant_children, so the relative emission order is deterministic and the continuation (user_with_cont) stays last. Because the downstream_walk_session_with_forksloop (Lines 491-495) runs_collect_descendantson everystitched[:-1]entry, anchors also get recorded inskipped, which keeps the coverage check happy — no risk of the fallback timestamp-sort kicking in for these sessions.One small, non-blocking note:
_collect_agent_anchorsis invoked once peracinassistant_children. If two assistant siblings ever shared a descendant (unlikely in a proper tree, but possible after orphan promotion), the same anchor UUID could be appended twice. If you want to be defensive, aset()dedupe before sort would be cheap:Optional dedupe
extracted_anchors: list[str] = [] for ac in assistant_children: extracted_anchors.extend(_collect_agent_anchors(ac, session_uuids, nodes)) # Stitch: dead-end children first, then the continuing user child - dead_ends = user_dead + assistant_children + extracted_anchors + dead_ends = list(dict.fromkeys(user_dead + assistant_children + extracted_anchors)) dead_ends.sort(key=lambda c: nodes[c].timestamp)🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@claude_code_log/dag.py` around lines 389 - 400, The code collects anchors into extracted_anchors by calling _collect_agent_anchors for each assistant child, which can produce duplicate UUIDs if two assistant siblings share a descendant; to fix, dedupe extracted_anchors (e.g., convert to a set and back to a list) before appending to dead_ends so each anchor UUID appears only once, then sort dead_ends by nodes[c].timestamp as before to preserve deterministic emission and the final user_with_cont ordering.test/test_dag.py (1)
1166-1301: Regression coverage for the anchor-preservation fix is well targeted.The three scenarios (2-teammate, 3-teammate, end-to-end splice) directly exercise the regression path in
_stitch_tool_resultsVariant 2 and would have caught the original bug. A couple of small, optional tightenings you may want to consider:
test_inner_anchor_preserved_in_dead_end_subtreeandtest_three_parallel_agents_all_anchors_preservedonly assert membership. Pinning the expected order (e.g.["a", "tool1", "tool2", "res2", "res1", "att1", "cont"]) would lock in the timestamp-sort behaviour of the newdead_ends + extracted_anchorspath and catch future reorderings.test_anchor_attachment_routes_subagent_sessioncould additionally assert thatsub1is emitted immediately afterres2in the traversal, which is the user-visible property the UI depends on:Optional stronger assertion
- traversed_uuids = [getattr(e, "uuid", "") for e in traverse_session_tree(tree)] - assert "sub1" in traversed_uuids, ( - "subagent entry not reached via anchor — agent session not spliced" - ) + traversed_uuids = [getattr(e, "uuid", "") for e in traverse_session_tree(tree)] + assert "sub1" in traversed_uuids, ( + "subagent entry not reached via anchor — agent session not spliced" + ) + # Subagent must be spliced right after its anchor. + assert traversed_uuids.index("sub1") == traversed_uuids.index("res2") + 1
- Minor: the local
from claude_code_log.converter import _integrate_agent_entriescould move to module-level imports for consistency with the other tests.🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@test/test_dag.py` around lines 1166 - 1301, The tests currently only assert membership; tighten them by (1) replacing the broad membership checks in test_inner_anchor_preserved_in_dead_end_subtree and test_three_parallel_agents_all_anchors_preserved with exact expected-order assertions (e.g. assert the main traversal/UUID list equals the timestamp-sorted sequence like ["a","tool1","tool2","res2","res1","att1","cont"] or contains those in that order) to lock in the dead_ends+extracted_anchors ordering, (2) in test_anchor_attachment_routes_subagent_session add a stronger assertion that "sub1" is immediately after "res2" in the traversal result (check the index of "res2" and assert traversal[index+1] == "sub1"), and (3) move the local import of _integrate_agent_entries to the module-level imports for consistency with other tests.
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Nitpick comments:
In `@claude_code_log/dag.py`:
- Around line 389-400: The code collects anchors into extracted_anchors by
calling _collect_agent_anchors for each assistant child, which can produce
duplicate UUIDs if two assistant siblings share a descendant; to fix, dedupe
extracted_anchors (e.g., convert to a set and back to a list) before appending
to dead_ends so each anchor UUID appears only once, then sort dead_ends by
nodes[c].timestamp as before to preserve deterministic emission and the final
user_with_cont ordering.
In `@claude_code_log/html/templates/components/message_styles.css`:
- Around line 734-745: The CSS uses a hardcoded string in
.tool-param-collapsible[open] > summary::after (content: "collapse") which
prevents localization; change it to use a CSS custom property (e.g., content:
var(--label-collapse, "collapse")) or move the hint into the template as a real
element so the text can be produced by the i18n pipeline; update the stylesheet
rule for summary::after and ensure templates set --label-collapse per locale (or
render an element inside summary that gets translated) so the hint becomes
localizable.
In `@test/test_dag.py`:
- Around line 1166-1301: The tests currently only assert membership; tighten
them by (1) replacing the broad membership checks in
test_inner_anchor_preserved_in_dead_end_subtree and
test_three_parallel_agents_all_anchors_preserved with exact expected-order
assertions (e.g. assert the main traversal/UUID list equals the timestamp-sorted
sequence like ["a","tool1","tool2","res2","res1","att1","cont"] or contains
those in that order) to lock in the dead_ends+extracted_anchors ordering, (2) in
test_anchor_attachment_routes_subagent_session add a stronger assertion that
"sub1" is immediately after "res2" in the traversal result (check the index of
"res2" and assert traversal[index+1] == "sub1"), and (3) move the local import
of _integrate_agent_entries to the module-level imports for consistency with
other tests.
ℹ️ Review info
⚙️ Run configuration
Configuration used: defaults
Review profile: CHILL
Plan: Pro
Run ID: 334a1eb4-1635-4784-a10d-ea78b5e7acb0
📒 Files selected for processing (5)
claude_code_log/dag.pyclaude_code_log/html/templates/components/message_styles.cssclaude_code_log/html/tool_formatters.pytest/__snapshots__/test_snapshot_html.ambrtest/test_dag.py
The detail-level filter and compact-Markdown mode landed in #96 and #114 but weren't surfaced in the README. Add: - Key Features bullet calling out the pairing with --format md for feeding past conversations to an LLM - New Usage subsection "Feeding Past Conversations to an LLM" with the --detail low --format md --compact combo and a short level reference - Problem-statement entry for LLM-based analysis/experience building - Markdown Output Features bullet for --compact Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This was referenced Apr 19, 2026
This was referenced Apr 26, 2026
Merged
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Tasktool_uses (e.g. teammates orchestrating multiple subagents) produced tool_result anchors nested inside sibling assistant subtrees. The DAG walker's variant 2 stitch classified those subtrees as dead-end and dropped the agentId-bearing anchors, leaving subagent sessions with no attachment point. Only the outermost subagent rendered. Fixes the issue Support "teammates" #91 symptom where 4 of 6 subagents were missing from a teammates session.<details class='tool-param-collapsible'>had[open] > summary { display: none }, hiding the browser's disclosure marker along with the preview — expanded values had no visible collapse affordance.Changes
claude_code_log/dag.py+test/test_dag.py(commit 6b37e8d)_collect_agent_anchors(uuid, session_uuids, nodes)walks an assistant subtree to findUserTranscriptEntrydescendants withagentIdset._stitch_tool_resultsvariant 2, after dead-end checks pass, extract agent anchors fromassistant_childrensubtrees and include them indead_ends(timestamp-sorted). Anchors end up in both the chain (preserving attachment points) andskipped(via_collect_descendants(assistant, ...)) — that's fine,walked ∩ skippedis allowed.experiments/worktreesshape), end-to-end splice verifying a subagent entry is reached through the anchor.claude_code_log/html/tool_formatters.py+ CSS (commit a0ed7d7)<span class='tool-param-preview'>at bothrender_params_tablecall sites (structured JSON and simple string).display: noneon the whole<summary>withdisplay: noneon.tool-param-previewonly, plus::after { content: "collapse"; }for an explicit hint. Summary keepslist-itemdisplay so the browser's disclosure arrow stays visible.Test plan
uv run pytest -n auto -m "not (tui or browser)"— 864 pass, 7 skip (19 pre-existing pagination failures on main, unrelated)uv run pytest -n auto -m browser— 29 pass, 1 skipuv run pyright— 0 errorsuv run ty check— all checks passeduv run ruff check/ruff format --check— cleanexperiments/worktreesregenerates 488K (2 subagents) → 768K (all 6 subagents)Relates to #91.
🤖 Generated with Claude Code
Summary by CodeRabbit
New Features
Bug Fixes
Tests
Documentation