The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
threadplane	Ready	Preview, Comment	Jun 20, 2026 3:23am

Claude finished @blove's task in 2m 28s —— View job

PR Review

• Read CLAUDE.md and understand repo guidelines
• Read the PR diff and changed files
• Review Python backend changes (graph.py)
• Review Angular frontend changes (app.config.ts, subagents.component.ts)
• Review e2e test changes
• Review fixture changes
• Post findings

Summary

Solid follow-up to #711. The core diagnosis is correct — the old flat @tool never created a tools:<call_id> namespace, so SubagentTracker had nothing to match and the inline cards never rendered. Compiling a real child StateGraph and renaming role → subagent_type fixes both problems cleanly. The single-call-per-subagent tradeoff to stabilize aimock replay is well-motivated and well-documented.

Correctness observations

task_description vs description in namespace matching

SubagentTracker.matchSubgraphToSubagent (in subagent-tracker.ts) first tries to match a subgraph namespace to a registered tool-call by comparing args['description'] (lines 137, 144) to the child graph's reported description. The task tool args carry task_description, not description, so this path always misses and the code falls through to the positional "first unmatched pending/running" fallback (line 152). For sequential deterministic dispatch this is fine — registration order equals dispatch order. But it means the description-based fast-path is effectively dead for this pattern. No bug, but worth documenting: if a future cap dispatches subagents with the same task_description for different slots, positional matching is the only disambiguator.

_SUBAGENT_PROMPTS.get silent fallback — inline comment posted (minor; a log line is enough).

Count assertion timeout — inline comment posted (cosmetic; safe in practice).

Other notes

• Removing ChatSubagentsComponent / ChatSubagentCardComponent from the cockpit imports is correct — the inline card is now composed inside <chat> by default, so the explicit sidebar tray was redundant.
• subagentToolNames: ['task'] in app.config.ts is the right toggle. The SubagentTracker's default already includes 'task', but being explicit here makes the cap self-documenting. ✓
• Fixture reduction from ~102 to ~26 lines (removing the nested tool-call round-trips) is expected and clean.
• _final_text defensively handles both string and list-block content formats — good coverage for streaming edge cases.

Verdict

Approve with minor nits. The two inline suggestions are low-priority (a log line and an explicit timeout). The implementation is correct and the e2e coverage is meaningfully stronger than before.

Claude finished @blove's task in 3m 18s —— View job

PR Review

• Read CLAUDE.md and understand repo guidelines
• Read PR diff and changed files
• Review Python backend (graph.py)
• Review Angular frontend (app.config.ts, subagents.component.ts)
• Review e2e tests and fixtures
• Post findings

Summary

Solid follow-up. The root cause diagnosis is correct: the old flat @tool never created a tools:<call_id> namespace, so SubagentTracker had nothing to match — hence no cards. Compiling a real child StateGraph and renaming role → subagent_type fixes both registration (isValidSubagentType(args['subagent_type']) now passes) and namespace matching.

Verification of tracker behaviour

Confirmed by reading subagent-tracker.ts:

• Registration (registerFromToolCalls, line 74): reads args['subagent_type'] — the new tool arg name. Passes for "research" / "booking" / "itinerary". ✓
• Matching (matchSubgraphToSubagent, lines 135–154): the description-based fast-paths read args['description'], which doesn't exist in the new tool args, so they always fall through to the positional fallback (first unmatched pending/running subagent). For sequential deterministic dispatch this is correct. The PR description documents this behaviour accurately.
• Card persistence: getSubagents() returns all subagents with status !== 'pending', so completed cards remain in the DOM — the toHaveCount(3) assertion is correctly placed after submitAndWaitForResponse. ✓

Open nits (from previous review, not yet addressed)

Two inline suggestions re-posted against the current commit — both low-priority:

1. _SUBAGENT_PROMPTS.get silent fallback (graph.py:124) — an unknown subagent_type silently falls back to _ITINERARY_PROMPT; a print on the else-path makes it debuggable.
2. toHaveCount(3) without explicit timeout (c-subagents.spec.ts:23) — safe in practice since the agent is idle by then, but an explicit { timeout: 15_000 } self-documents intent.

One new minor observation

ChatOpenAI(model="gpt-5-mini", streaming=True) is instantiated fresh on every _subagent_node invocation (graph.py:126). For a demo cap this is fine, but moving it to module-level would avoid recreating the client object on each of the three subagent calls. Not a correctness issue.

Verdict

Implementation is correct. The e2e coverage is meaningfully stronger. Two pre-existing nits remain (inline suggestions on the current commit above).

Claude finished @blove's task in 1m 54s —— View job

PR Review

• Read CLAUDE.md and repo guidelines
• Read the full PR diff and changed files
• Check if prior review nits were addressed
• Review for correctness, type-safety, and new issues
• Post findings

Summary

This synchronize push merges recent main commits into the branch — no substantive changes to the PR implementation. The core work (real subagent_subgraph, subagent_type arg, subagentToolNames: ['task'], fixture re-record) remains correct. The sync introduced no conflicts.

Two outstanding nits (flagged in prior two passes, still unaddressed)

1. graph.py:124 — silent fallback — _SUBAGENT_PROMPTS.get(subagent_type, _ITINERARY_PROMPT) silently produces an itinerary response for any unrecognised subagent_type. Inline suggestion re-posted: add a print on the else-path.

2. c-subagents.spec.ts:23 — toHaveCount(3) without explicit timeout — inherits the project default (5 s). Safe now since submitAndWaitForResponse already idles, but a { timeout: 15_000 } self-documents intent. Inline suggestion re-posted.

Both are one-liners. Everything else looks good.