feat: session token breakdown in footer and /status

[encyc]

Summary

Re-submit of #1666, rebased onto current main (v0.8.45) after the CodeWhale rebrand.

Add accumulated session token tracking with input / cache-hit / output breakdown:

• SessionState: new total_input_tokens, total_cache_hit_tokens, total_cache_miss_tokens, total_output_tokens fields
• Turn outcome handler: accumulate per-turn token breakdown from usage struct
• StatusItem::Tokens: new footer chip, enabled by default
• Footer chip: compact format — 12K in · 8.1K cch · 2.5K out
• /status command: expanded with session input / cache / output rows
• /clear and /load: reset accumulated breakdown (per-runtime-session, not persisted)

Verified

• cargo fmt --check ✓
• cargo check ✓
• cargo clippy ✓ (no new warnings)
• cargo test --workspace --all-features — 3329 passed; 2 pre-existing failures unrelated to this change

Greptile Summary

This PR adds accumulated session token tracking (input / cache-hit / cache-miss / output) to the TUI footer chip and /status command. It is a rebase of #1666 onto the CodeWhale-branded main.

• SessionState gains four u32 accumulator fields and a reset_token_breakdown helper that is called consistently from /clear, /load, and apply_loaded_session, keeping all token-reset paths in sync.
• TurnOutcome handler accumulates per-turn input and output tokens unconditionally; cache telemetry is gated on prompt_cache_hit_tokens being Some, which avoids double-counting when the provider doesn't report cache data.
• Footer chip (StatusItem::Tokens, enabled by default) renders a compact N in · N cch · N out label; the chip is hidden until the first turn completes. The /status command gets three new rows showing the same breakdown in raw counts.

Confidence Score: 5/5

Safe to merge; all accumulation, reset, and display paths are correctly wired and symmetric.

The accumulation logic is straightforward: four counters incremented per turn, reset in all three load/clear paths, displayed in the footer chip and /status. The cache-telemetry guard (if let Some(hit_tokens)) correctly prevents double-counting when the provider omits cache fields. Config enum and UI mappings are complete with no missing arms. The only gaps are missing test assertions for the new /status rows.

The status_report_includes_runtime_fields test in crates/tui/src/commands/status.rs would benefit from asserting the new Session input/cache/output rows.

Important Files Changed

Filename	Overview
crates/tui/src/tui/app.rs	Adds four u32 accumulator fields and a reset_token_breakdown helper to SessionState; defaults are zero and the Default impl is updated correctly.
crates/tui/src/tui/ui.rs	Accumulates per-turn input/output/cache-hit/cache-miss tokens in the TurnOutcome handler; cache telemetry is gated on prompt_cache_hit_tokens being Some, avoiding the double-count bug. Also calls reset_token_breakdown inside apply_loaded_session.
crates/tui/src/tui/footer_ui.rs	Adds footer_session_tokens_spans and wires it into the Tokens chip slot; chip is hidden when both input and output are zero, consistent with other optional chips.
crates/tui/src/commands/status.rs	Adds three new rows to /status output (Session input/cache/output); existing test does not assert the new rows, so a regression there would go undetected.
crates/tui/src/commands/core.rs	Calls reset_token_breakdown in the /clear handler alongside the existing token-counter resets; correctly symmetric with the load and session-select paths.
crates/tui/src/commands/session.rs	Calls reset_token_breakdown in the /load handler; comment correctly explains that the breakdown is per-runtime-session, not persisted.
crates/tui/src/config.rs	Adds StatusItem::Tokens to the enum and all five required match arms (key, label, description, defaults, all); no arms are missing.
crates/tui/src/config_ui.rs	Mirrors StatusItem::Tokens into StatusItemValue with both From impls updated symmetrically.

Sequence Diagram

sequenceDiagram
    participant Engine
    participant ui.rs as run_event_loop
    participant SessionState
    participant FooterUI as footer_ui.rs
    participant StatusCmd as /status

    Engine->>ui.rs: "TurnOutcome { usage }"
    ui.rs->>SessionState: "total_input_tokens += usage.input_tokens"
    ui.rs->>SessionState: "total_output_tokens += usage.output_tokens"
    alt prompt_cache_hit_tokens is Some(hit)
        ui.rs->>SessionState: "total_cache_hit_tokens += hit"
        ui.rs->>SessionState: "total_cache_miss_tokens += miss (or input-hit)"
    end

    Note over ui.rs,SessionState: /clear or /load or apply_loaded_session
    ui.rs->>SessionState: reset_token_breakdown()

    SessionState-->>FooterUI: total_input/cache_hit/output_tokens
    FooterUI-->>FooterUI: render N in · N cch · N out chip

    SessionState-->>StatusCmd: total_input/cache_hit/cache_miss/output_tokens
    StatusCmd-->>StatusCmd: render Session input/cache/output rows

Loading

_{Reviews (2): Last reviewed commit: "fix: address review feedback — current_s..." | Re-trigger Greptile}

[gemini-code-assist]

Code Review

This pull request introduces tracking and displaying of accumulated session token usage (input, cache hit, cache miss, and output tokens) across the TUI application. Feedback focuses on refactoring the repeated token-resetting logic into a helper method on SessionState to keep the code DRY. Additionally, improvements are suggested to handle models/providers that do not support cache telemetry, ensuring the UI (footer and status view) does not display misleading zero-values or redundant cache metrics when telemetry is unavailable.

[gemini-code-assist]

To avoid duplicating the resetting of the accumulated token breakdown fields across multiple commands and UI event loops, consider adding a helper method on SessionState.

For example:

impl SessionState {
    /// Reset the accumulated token breakdown fields to zero.
    pub fn reset_token_breakdown(&mut self) {
        self.total_input_tokens = 0;
        self.total_cache_hit_tokens = 0;
        self.total_cache_miss_tokens = 0;
        self.total_output_tokens = 0;
    }
}

[gemini-code-assist]

Use the newly introduced reset_token_breakdown helper method on SessionState to reduce boilerplate and keep the code DRY.

Suggested change

	app.session.total_input_tokens = 0;
	app.session.total_cache_hit_tokens = 0;
	app.session.total_cache_miss_tokens = 0;
	app.session.total_output_tokens = 0;
	app.session.reset_token_breakdown();

[gemini-code-assist]

Use the reset_token_breakdown helper method on SessionState here as well to simplify the session loading logic.

Suggested change

	app.session.total_input_tokens = 0;
	app.session.total_cache_hit_tokens = 0;
	app.session.total_cache_miss_tokens = 0;
	app.session.total_output_tokens = 0;
	app.session.reset_token_breakdown();

[gemini-code-assist]

Use the reset_token_breakdown helper method on SessionState here to keep the session loading logic consistent and clean.

    app.session.reset_token_breakdown();

[gemini-code-assist]

If the active model or provider does not support cache telemetry, prompt_cache_hit_tokens will be None. In this case, defaulting to 0 and subtracting it from input_tokens falsely attributes all input tokens as cache misses. This leads to misleading metrics (e.g., showing 100% cache misses for unsupported models).

Only accumulate cache hit and miss tokens if prompt_cache_hit_tokens is actually reported (Some).

                        if let Some(hit_tokens) = usage.prompt_cache_hit_tokens {
                            app.session.total_cache_hit_tokens = app
                                .session
                                .total_cache_hit_tokens
                                .saturating_add(hit_tokens);
                            let cache_miss = usage.prompt_cache_miss_tokens.unwrap_or_else(|| {
                                usage.input_tokens.saturating_sub(hit_tokens)
                            });
                            app.session.total_cache_miss_tokens = app
                                .session
                                .total_cache_miss_tokens
                                .saturating_add(cache_miss);
                        }

[gemini-code-assist]

If the active model does not support cache telemetry, total_cache_hit_tokens and total_cache_miss_tokens will both be 0. Displaying 0 cch in the footer is redundant and clutters the UI for unsupported models. Consider omitting the cch chip entirely when no cache telemetry has been recorded.

    if session.total_input_tokens == 0 && session.total_output_tokens == 0 {
        return Vec::new();
    }
    let in_str = format_token_count_compact(u64::from(session.total_input_tokens));
    let out_str = format_token_count_compact(u64::from(session.total_output_tokens));
    let text = if session.total_cache_hit_tokens == 0 && session.total_cache_miss_tokens == 0 {
        format!("{in_str} in · {out_str} out")
    } else {
        let cache_str = format_token_count_compact(u64::from(session.total_cache_hit_tokens));
        format!("{in_str} in · {cache_str} cch · {out_str} out")
    };
    vec![Span::styled(
        text,
        Style::default().fg(palette::TEXT_MUTED),
    )]

[gemini-code-assist]

For models/providers that do not support cache telemetry, showing 0 hit / 0 miss is less clear than explicitly stating not reported (similar to how cache_summary handles it). Consider showing not reported when both hit and miss counts are zero.

Suggested change

	push_row(
	&mut out,
	"Session cache:",
	&format!(
	"{} hit / {} miss",
	app.session.total_cache_hit_tokens, app.session.total_cache_miss_tokens
	),
	);
	let session_cache = if app.session.total_cache_hit_tokens == 0 && app.session.total_cache_miss_tokens == 0 {
	"not reported".to_string()
	} else {
	format!(
	"{} hit / {} miss",
	app.session.total_cache_hit_tokens, app.session.total_cache_miss_tokens
	)
	};
	push_row(
	&mut out,
	"Session cache:",
	&session_cache,
	);

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 6cde1c74f3

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Restore current session ID when applying a loaded session

apply_loaded_session no longer assigns app.current_session_id, so after resuming/loading through this path the app keeps None or a stale previous ID. That breaks session scoping: startup queue restore compares against current_session_id, and later saves/syncs use this field to decide which session to update, so loading session B after working in session A can cause updates to be written to A (or a new ID) instead of B.

Useful? React with 👍 / 👎.