[Bug]: Subagent thread bar shows Medium for a Claude subagent that runs at high effort #15896

Activity

1. isinghmitesh commented on Oct 5, 2026

   

   isinghmitesh

   on Oct 5, 2026

   More actions

   Also seeing an effort-display mismatch on macOS with T3 Code Nightly 0.0.46-nightly.20261005.2667 on 2026-10-05, in the opposite direction to this report.

   A Claude Opus 5.5 parent launched three native Explore agents whose definitions pin model: claude-sonnet-5-5 and effort: medium. T3 displayed all three as Sonnet 5.5 at high effort. Their underlying Claude Code JSONL transcripts recorded effort: "medium" and perTurnEffort: "medium" on every Sonnet assistant record: 29, 40, and 5 records respectively, 74 total.

   The child model selections in orchestration_v2_projection_threads lacked effort options. Two were { "instanceId": "claudeAgent", "model": "claude-sonnet-5-5" }; the third became <synthetic> after a usage-limit error. Its preceding Sonnet records still consistently showed medium.

   We have verified the transcript values and missing stored options, but have not traced which UI formatting path produced the high label. This supports the broader mismatch between displayed effort and recorded runtime effort, without establishing that the exact fallback in this issue is responsible for our variant.

   Added the fuller reproduction to #15214 as well. Showing the actual recorded effort, or unknown when unavailable, would make the subagent label reliable.

   Investigated and posted with Codex on the user's behalf, using local transcript metadata and read-only T3 state inspection.
2. LunarRed commented on Oct 7, 2026

   

   LunarRed

   on Oct 7, 2026

   More actions

   Same on T3 Code nightly 2026-10-07. A Claude Opus 5.5 subagent started with the Agent tool and effort: "max" showed Claude Opus 5.5 Medium in the subagent bar. Its transcript (tasks/<agentId>.output) recorded "effort":"max" and "perTurnEffort":"max" on all 143 requests. Here the effort came from the Agent tool's effort parameter, not from agent frontmatter, so the bar ignores both sources.
3. NicolasStuff commented on Oct 8, 2026

   

   NicolasStuff

   on Oct 8, 2026

   More actions

   Additional confirmation on macOS with T3 Code Nightly 0.0.46-nightly.20261005.2702, investigated on 2026-10-08.

   The parent was explicitly asked to launch three Opus 5.5 subagents at xhigh. These were native Claude Agent children, not T3 delegate_task children.

   For all three children:

   • The native launch metadata records model: "opus" and effort: "xhigh".
   • Their own Claude Code JSONL transcripts record both effort: "xhigh" and perTurnEffort: "xhigh" on every assistant record that includes those fields: 100, 146, and 174 records, respectively, at inspection time. None of those records says medium.
   • T3's read-only orchestration_v2_projection_threads state contains the following child selection, with no effort options:

   {"instanceId":"claudeAgent","model":"claude-opus-5-5"}

   This confirms the display mismatch for an explicit per-invocation effort request, without depending on agent frontmatter or assumed inheritance. The user's initial impression was that the parent ignored the requested effort; the runtime records show that it did not.

   Latest release/source check: the latest published nightly at inspection time is v0.0.46-nightly.20261008.2833, commit a6ec88f7a716fc421bd22c2484881c44110f9375 (also the checked-out main HEAD). Its source still has the relevant paths:

   • rememberClaudeSubagentLaunch retains model information but does not read the Agent input's effort.
   • syncSubagentThreadModel replaces the selection with { instanceId, model } when the native model changes.
   • formatModelSelectionEffort still resolves an absent effort through the catalog default.

   #15710 is still open and addresses the false default label. This comment adds independent runtime evidence rather than duplicating that PR or this issue. Learning and displaying the actual native effort remains the related work discussed in #15214.

   Verification boundary: the installed nightly and local transcript metadata were inspected directly; the October 8 nightly was checked in source, not installed or run. No full transcripts, task content, project identifiers, or credentials are attached.

   Investigated and posted with GPT-6-Astra in the Codex harness inside T3 Code, on the user's behalf.