fix: resolve subagent type showing _unknown while running by JayantDevkar · Pull Request #57 · JayantDevkar/claude-code-karma

JayantDevkar · 2026-04-13T13:15:22Z

What's broken

When you open a subagent's detail page while it's still running, the type badge shows _unknown instead of the actual type (e.g. oh-my-claudecode:critic). Once the agent finishes, the type magically appears. Beyond that, the agent status system has several gaps compared to session status.

Issues addressed

Issue	Gap	Description
#56	Type detection	Agent type shows `_unknown` while running
#58	SSR pre-fetch	Agent page missing server-side live session data
#59	Smart polling	Agent polling doesn't stop after agent completes
#60	Status granularity	No active/idle distinction for running agents
#61	API endpoint	No dedicated agent live-status endpoint

Root cause (#56)

We figure out which agent belongs to which Task tool call by matching the agentId field inside the tool_result message. Problem is, that tool_result only gets written to the parent session JSONL after the agent finishes. So while it's running, we have no way to link agent → Task → type, and it falls through to the default _unknown classification.

Changes

1. Prompt-based type matching (#56)

Added Phase 2.5 to get_all_subagent_types() — for any agent that's still unmatched after tool_result extraction:

Read its first JSONL line to grab the initial prompt (first 100 chars)
Normalize it (lowercase, collapse whitespace)
Compare against Task tool_use input prompts from the parent session

Also added a guard in _determine_subagent_type() to convert internal markers (_unknown, _warmup) to null in the API response.

2. SSR live session pre-fetch (#58)

Added liveSession fetch to agents/[agent_id]/+page.server.ts so the status badge appears immediately on page load instead of waiting for client-side polling.

3. Smart polling termination (#59)

Added logic in ConversationView.svelte to stop refreshing agent data once the subagent has completed, even if the parent session is still running.

4. Active/idle status for agents (#60)

Frontend heuristic: derive active vs idle from the agent's last message timestamp. No hook changes needed.

5. Dedicated agent live-status endpoint (#61)

Added /agents/{encoded}/{uuid}/agents/{agentId}/live-status that reads the parent session's state file and returns just the subagent's status entry.

Testing

All 1474 existing tests pass
Verified prompt matching against real session data
Lint clean (ruff check)

Files changed

File	What changed
`api/services/subagent_types.py`	Phase 2.5 prompt matching + 3 helper functions
`api/routers/subagent_sessions.py`	Filter `_` markers from API response + live-status endpoint
`frontend/src/routes/projects/[project_slug]/[session_slug]/agents/[agent_id]/+page.server.ts`	SSR live session fetch
`frontend/src/routes/projects/[project_slug]/[session_slug]/agents/[agent_id]/+page.svelte`	Pass liveSession prop
`frontend/src/lib/components/conversation/ConversationView.svelte`	Smart polling termination
`frontend/src/lib/components/conversation/ConversationHeader.svelte`	Active/idle heuristic for agents

Add prompt-based matching (Phase 2.5) to get_all_subagent_types() as a fallback for running agents where no tool_result exists yet. Compares each unmatched agent's initial prompt against Task tool_use input prompts from the parent session JSONL. Also filter internal classification markers (_unknown, _warmup) from the API response in _determine_subagent_type() so the frontend shows "Other" instead of raw internal strings. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

JayantDevkar

Walkthrough of the changes for review. The core idea: we already have the agent's prompt and the Task input prompt available before the agent finishes — we just weren't matching them.

JayantDevkar · 2026-04-13T13:19:10Z

api/services/subagent_types.py

+    # Phase 2.5: Match unmatched agents by prompt content
+    # Fallback for running agents where no tool_result exists yet.
+    # Compares each unmatched agent's initial prompt against Task input prompts.
+    unmatched_agent_files = [f for f in agent_files if f.stem.removeprefix("agent-") not in result]


Why this exists: Before this change, an agent only got its type after Phase 1 matched its agentId from a tool_result. But tool_result is only written when the agent finishes. This new phase runs before the prefix/content classification (Phase 3-4), so running agents get typed correctly instead of falling through to _unknown.

We only build the prompt lookup table when there are actually unmatched agents — no extra JSONL scanning for sessions where everything already resolved.

JayantDevkar · 2026-04-13T13:19:10Z

api/services/subagent_types.py

+                agent_id = agent_file.stem.removeprefix("agent-")
+                agent_prompt = _get_agent_initial_prompt(agent_file)
+                if agent_prompt:
+                    key = _normalize_key(agent_prompt)


How the matching works: The Task input.prompt field contains the exact text that becomes the agent's first UserMessage. We normalize both sides (lowercase + collapse whitespace) and compare the first 100 chars. This is the same approach collect_subagent_info() in collectors.py:469-471 already uses for the session subagents tab — we're just bringing it to the shared extraction function.

Edge case: if two Task calls have identical first-100-char prompts but different types, the last one wins. In practice this doesn't happen since prompts are unique per spawn.

JayantDevkar · 2026-04-13T13:19:10Z

api/services/subagent_types.py

+    return " ".join(text.lower().strip().split())
+
+
+def _extract_task_prompt_types(path: Path) -> dict[str, str]:


Three new helpers — kept minimal and self-contained:

_normalize_key() — same logic as utils.normalize_key(), duplicated here to keep subagent_types.py dependency-free (it only imports json/logging/re/Path)

_extract_task_prompt_types() — single-pass JSONL scan, same pattern as _extract_types_from_raw_jsonl() above

_get_agent_initial_prompt() — reads only the first line of the agent JSONL, same pattern as _classify_by_first_message()

JayantDevkar · 2026-04-13T13:19:10Z

api/routers/subagent_sessions.py

-    return types.get(agent_id)
+    result = types.get(agent_id)
+    # Convert internal classification markers to None for the API response
+    if result and result.startswith("_"):


Defensive filter: get_all_subagent_types() uses internal markers like _unknown, _warmup, _teammate, _compact for classification. These should never leak into the API response. Previously _unknown was returned verbatim and the frontend displayed it as a badge label. Now any _-prefixed result becomes null, which the frontend renders as "Other" via getSubagentTypeDisplayName().

…dle, smart polling (#58, #59, #60, #61) Gap 1 (#58): Add server-side live session fetch to agent detail page so status badge renders immediately on page load instead of waiting for client-side polling. Gap 4 (#61): Add GET /agents/{encoded}/{uuid}/agents/{agentId}/live-status endpoint that extracts subagent status from parent session's state file. Gap 3 (#60): Derive active/idle from entity end_time for running agents using 15s idle threshold. Reuses session statusConfig colors. Gap 5 (#59): Stop refreshing agent JSONL data after subagent completes while continuing live-status polling at 5s rate for badge accuracy. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Previously, the live-sessions fetch was awaited sequentially before the Promise.all, adding an unnecessary round-trip to every agent detail page load. All 6 fetches are independent, so they should fire simultaneously. Also fixes ruff formatting in subagent_sessions.py and adds review.md with remaining polish suggestions for the PR.

Remove review.md — PR review artifact, not project documentation. Fix /live-status endpoint docstring to accurately document the three response cases (404, 200 with subagent=null, 200 with subagent data). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

JayantDevkar requested a review from the-non-expert April 13, 2026 13:15

JayantDevkar commented Apr 13, 2026

View reviewed changes

JayantDevkar and others added 3 commits April 13, 2026 19:48

perf: fetch live session status in parallel with agent data on SSR load

aea211c

fix: lint fixes

5970abd

the-non-expert approved these changes Apr 14, 2026

View reviewed changes

the-non-expert and others added 2 commits April 14, 2026 10:17

JayantDevkar merged commit 4067d87 into main Apr 14, 2026
20 checks passed

JayantDevkar deleted the fix/56-subagent-type-unknown-while-running branch April 14, 2026 15:41

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

fix: resolve subagent type showing _unknown while running#57

fix: resolve subagent type showing _unknown while running#57
JayantDevkar merged 6 commits intomainfrom
fix/56-subagent-type-unknown-while-running

JayantDevkar commented Apr 13, 2026 •

edited

Loading

Uh oh!

JayantDevkar left a comment

Uh oh!

JayantDevkar Apr 13, 2026

Uh oh!

JayantDevkar Apr 13, 2026

Uh oh!

JayantDevkar Apr 13, 2026

Uh oh!

JayantDevkar Apr 13, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

		return " ".join(text.lower().strip().split())


		def _extract_task_prompt_types(path: Path) -> dict[str, str]:

Conversation

JayantDevkar commented Apr 13, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What's broken

Issues addressed

Root cause (#56)

Changes

1. Prompt-based type matching (#56)

2. SSR live session pre-fetch (#58)

3. Smart polling termination (#59)

4. Active/idle status for agents (#60)

5. Dedicated agent live-status endpoint (#61)

Testing

Files changed

Uh oh!

JayantDevkar left a comment

Choose a reason for hiding this comment

Uh oh!

JayantDevkar Apr 13, 2026

Choose a reason for hiding this comment

Uh oh!

JayantDevkar Apr 13, 2026

Choose a reason for hiding this comment

Uh oh!

JayantDevkar Apr 13, 2026

Choose a reason for hiding this comment

Uh oh!

JayantDevkar Apr 13, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

JayantDevkar commented Apr 13, 2026 •

edited

Loading