[BUG] [v0.0.7] SubagentExecutor summary turn skips total_tokens update — session.tokens_used under-reports when summary is requested (executor.rs:493-497)

[echoforge2200]

Project

cortex

Description

When SubagentExecutor::run_subagent() requests a mandatory summary turn (lines 444-501 in executor.rs), it merges the summary turn's token counts into turn_ctx.tokens at lines 493-497:

// Update token counts
turn_ctx.tokens.input_tokens += summary_turn_ctx.tokens.input_tokens;
turn_ctx.tokens.output_tokens += summary_turn_ctx.tokens.output_tokens;
turn_ctx.tokens.cached_tokens += summary_turn_ctx.tokens.cached_tokens;
turn_ctx.tokens.reasoning_tokens += summary_turn_ctx.tokens.reasoning_tokens;

However, turn_ctx.tokens.total_tokens is not updated. The TokenUsage struct (defined in agent/mod.rs:234-247) stores total_tokens as a separate field — it is NOT computed from input_tokens + output_tokens. The proper way to update it is via TurnContext::add_tokens() (orchestrator.rs:85-88) which does self.tokens.total_tokens += input + output, but the manual addition at lines 493-497 skips this field entirely.

Later at line 533, session.record_turn(...) uses turn_ctx.tokens.total_tokens which now excludes the summary turn's tokens:

session.record_turn(
    turn_ctx.tool_results.len() as u32,
    turn_ctx.tokens.total_tokens as u64,  // Missing summary tokens
);

Error Message

No error — the token count is silently wrong.

Debug Logs

No response

System Information

Cortex version: v0.0.7
OS: Linux
Affected file: src/cortex-engine/src/tools/handlers/subagent/executor.rs (lines 493-497, 533)
Related: src/cortex-engine/src/agent/mod.rs (lines 234-247 — TokenUsage struct)
Related: src/cortex-engine/src/agent/orchestrator.rs (lines 85-88 — add_tokens method)

Steps to Reproduce

1. Spawn a subagent that completes its task but does NOT include summary markers (e.g., "## Summary for Orchestrator") in its response.
2. The executor detects missing summary at line 447: !has_summary_output(&result.response).
3. A summary turn is executed (lines 454-469).
4. If the summary succeeds, token counts are merged at lines 493-497 — but total_tokens is skipped.
5. session.record_turn() at line 533 records turn_ctx.tokens.total_tokens which excludes the summary turn's tokens.
6. SubagentSession.tokens_used and SubagentResult.token_usage under-report actual token consumption.

Expected Behavior

turn_ctx.tokens.total_tokens should be updated alongside the other token fields when merging summary turn tokens. The missing line should be:

turn_ctx.tokens.total_tokens += summary_turn_ctx.tokens.total_tokens;

Or equivalently, use turn_ctx.add_tokens(summary_turn_ctx.tokens.input_tokens, summary_turn_ctx.tokens.output_tokens) which correctly updates all three fields.

Actual Behavior

total_tokens is not updated with summary turn tokens, causing session.tokens_used to under-report. This affects billing/quota tracking and user-facing token usage reports in SubagentResult.token_usage and ProgressEvent::Completed.total_tokens.

Additional Context

The TokenUsage struct stores total_tokens independently:

pub struct TokenUsage {
    pub input_tokens: u32,
    pub output_tokens: u32,
    pub total_tokens: u32,    // NOT computed — must be updated manually
    pub cached_tokens: u32,
    pub reasoning_tokens: u32,
}

The add_tokens method correctly maintains the invariant total_tokens = input + output, but the manual field-by-field addition at lines 493-497 breaks it by omitting total_tokens.