feat(hermes): add Hermes Agent usage adapter - #1026

sequenceDiagram
  participant User as CLI
  participant Run as adapter::hermes::run
  participant Loader as load_entries
  participant DB as Hermes state.db
  participant Pricing as PricingMap
  participant Summary as summarize_entries
  participant Reporter as report_from_rows
  User->>Run: ccusage hermes <report> [--since/--until/...]
  Run->>Loader: load_entries(shared, pricing)
  Loader->>DB: open state.db / query sessions
  DB-->>Loader: session rows (tokens, messages, cost)
  Loader->>Pricing: lookup model/provider candidates
  Loader-->>Run: Vec<LoadedEntry>
  Run->>Summary: summarize_entries(entries, kind)
  Summary-->>Reporter: Vec<UsageSummary>
  Reporter-->>User: JSON or formatted table

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Open in StackBlitz

npx https://pkg.pr.new/ryoppippi/ccusage@1026

npx https://pkg.pr.new/ryoppippi/ccusage/@ccusage/ccusage-darwin-arm64@1026

npx https://pkg.pr.new/ryoppippi/ccusage/@ccusage/ccusage-darwin-x64@1026

npx https://pkg.pr.new/ryoppippi/ccusage/@ccusage/ccusage-linux-arm64@1026

npx https://pkg.pr.new/ryoppippi/ccusage/@ccusage/ccusage-linux-x64@1026

npx https://pkg.pr.new/ryoppippi/ccusage/@ccusage/ccusage-win32-arm64@1026

npx https://pkg.pr.new/ryoppippi/ccusage/@ccusage/ccusage-win32-x64@1026

commit: 338d94c

ccusage performance comparison

This compares the PR build against the base branch build on the same CI runner.

Committed fixture performance

Committed small fixtures for stable PR-to-PR feedback and explicit Claude/Codex command coverage.

Fixtures: Claude apps/ccusage/test/fixtures/claude (0.00 MiB, 2 files), Codex apps/ccusage/test/fixtures/codex (0.00 MiB, 1 files)
Runtime: package ccusage bin from apps/ccusage/package.json through bun -b, --offline --json, measured by hyperfine with 2 warmups and 7 runs.

Large real-world-shaped fixture performance

Generated fixtures shaped from aggregate local log statistics: thousands of JSONL files, many small sessions, and a long tail of larger sessions. No real prompts, paths, or outputs are stored in the fixtures.

Fixtures: Claude /home/runner/work/_temp/ccusage-large-fixture (1.01 GiB, 2,597 files), Codex /home/runner/work/_temp/ccusage-large-codex-fixture (1.01 GiB, 2,597 files)
Runtime: package ccusage bin from apps/ccusage/package.json through bun -b, --offline --json, measured by hyperfine with 0 warmups and 1 runs.

Package size

Lower medians and smaller packed package sizes are better. CI runner noise still applies; use same-run ratios as directional PR feedback, not release guarantees.

@cubic-dev-ai review it

@cubic-dev-ai review it

@ryoppippi I have started the AI code review. It will take a few minutes to complete.

@coderabbitai review it

#!/bin/bash
cat apps/ccusage/src/adapter/hermes/parser.ts
echo "---FILE_BOUNDARY---"
cat apps/ccusage/src/adapter/hermes/pricing.ts
echo "---FILE_BOUNDARY---"
cat apps/ccusage/src/adapter/hermes/paths.ts
echo "---FILE_BOUNDARY---"
cat apps/ccusage/src/adapter/hermes/index.ts
echo "---FILE_BOUNDARY---"
cat apps/ccusage/src/adapter/sqlite.ts
echo "---FILE_BOUNDARY---"
cat apps/ccusage/src/adapter/opencode/paths.ts
echo "---FILE_BOUNDARY---"
cat apps/ccusage/src/adapter/opencode/loader.ts
echo "---FILE_BOUNDARY---"
cat apps/ccusage/src/adapter/index.ts
echo "---FILE_BOUNDARY---"
cat apps/ccusage/src/adapter/types.ts
echo "---FILE_BOUNDARY---"
cat apps/ccusage/src/commands/index.ts
echo "---FILE_BOUNDARY---"
cat apps/ccusage/src/config-loader-tokens.ts
echo "---FILE_BOUNDARY---"
cat apps/ccusage/test/cli-output.test.ts

@ryoppippi Great PR! The Hermes adapter follows the existing patterns well and the SQLite helper extraction is a nice improvement. Here are my findings:

Review: feat(hermes): add Hermes Agent usage adapter

🔴 Issues

1. createHermesStateDb duplicated in three places

The same schema-creation helper appears independently in:

• apps/ccusage/src/adapter/hermes/parser.ts (in the import.meta.vitest block)
• apps/ccusage/src/adapter/hermes/index.ts (in the import.meta.vitest block)
• apps/ccusage/test/cli-output.test.ts

If the sessions table schema ever gains a new column (e.g., cost_currency), all three copies need updating. Consider extracting this into a single shared test fixture helper (e.g., src/adapter/hermes/_test-helpers.ts or a test/helpers/hermes.ts).

2. PROVIDER_ALIASES is narrow — unknown providers fall back to 'hermes'

apps/ccusage/src/adapter/hermes/parser.ts

const PROVIDER_ALIASES = new Map<string, string>([
  ['anthropic', 'anthropic'],
  ['claude', 'anthropic'],
  ['openai', 'openai'],
  ...
]);

Common providers like mistral, cohere, deepseek, bedrock, together, and fireworks are absent. When a session has one of these as billing_provider, normalizeProvider falls through to inferProviderFromModel, and if the model name doesn't match any known pattern, provider is set to 'hermes'.

The consequence is in createModelCandidates:

if (entry.provider !== 'hermes') {
  candidates.push(`${entry.provider}/${entry.model}`);
}

When provider is 'hermes', the provider/model candidate is never tried, so LiteLLM lookup falls back to 0 silently.

Consider either expanding the aliases map or always adding the raw billing_provider/model candidate when billing_provider is non-null:

function createModelCandidates(entry: HermesUsageEntry, rawProvider?: string): string[] {
  const candidates = [entry.model];
  if (entry.provider !== 'hermes') {
    candidates.push(`${entry.provider}/${entry.model}`);
  }
  // Also try the raw billing_provider value if it wasn't recognized
  if (rawProvider != null && rawProvider !== entry.provider) {
    candidates.push(`${rawProvider}/${entry.model}`);
  }
  return Array.from(new Set(candidates));
}

3. Silent $0 fallback in calculateHermesCost with no warning

apps/ccusage/src/adapter/hermes/pricing.ts

for (const candidate of createModelCandidates(entry)) {
  const result = await fetcher.calculateCostFromTokens(tokens, candidate);
  if (Result.isSuccess(result) && Number.isFinite(result.value) && result.value >= 0) {
    return result.value;
  }
}
return 0; // <-- silent zero

When LiteLLM has no pricing info for any candidate, costs are silently reported as $0.00. This is especially confusing because the user will see a session with tokens but zero cost. Adding a logger.warn (similar to how loadReadonlySqliteRows logs failures) would make this diagnosable without the docs:

logger.warn(`No pricing found for model '${entry.model}' (provider: '${entry.provider}'); cost reported as $0`);
return 0;

🟡 Warnings

4. o\d regex won't match future multi-digit OpenAI o-series models

apps/ccusage/src/adapter/hermes/parser.ts

/^(?:gpt|chatgpt)(?:[-/]|$)|^o\d(?:[-/]|$)/u

o1, o3, o4 all match, but o10-mini (hypothetical) would not because \d matches exactly one digit. Consider \d+:

/^(?:gpt|chatgpt)(?:[-/]|$)|^o\d+(?:[-/]|$)/u

5. Hermes integration test is conditionally skipped

apps/ccusage/test/cli-output.test.ts

it.skipIf(getSqliteDatabaseFactory(() => {}) == null)(
  'renders Hermes Agent direct JSON from state.db',

This is the only end-to-end test for the Hermes adapter. In CI environments without native SQLite bindings, this test is a no-op. Since the Hermes adapter is SQLite-only (unlike OpenCode which has a JSON fallback), consider documenting the CI requirement or adding a non-SQLite-gated test that exercises the parser directly with fixture data.

6. Negative cost values are silently clamped to zero

apps/ccusage/src/adapter/hermes/parser.ts

const costUSD =
  toNonNegativeNumber(row.actual_cost_usd) ?? toNonNegativeNumber(row.estimated_cost_usd);

toNonNegativeNumber uses Math.max(number, 0), so a negative value (e.g., a credit/refund in actual_cost_usd) is silently treated as 0. If this is intentional, a code comment explaining the decision would help future readers.

🟢 Nits

• loadReadonlySqliteRows return typing: The result.value ?? [] fallback is safe but slightly odd since readRows is typed as (db) => T[] — it should never return null/undefined. A comment explaining this covers the edge case from withSqliteDatabase's return type would clarify intent.

• Missing hermes weekly is clearly intentional and properly guarded by getUnsupportedAgentReportError. 👍

• detectHermes vs detectHermesStateDb: The two-function split (one pure path check, one with SQLite support gate) is clean and consistent with the OpenCode pattern. 👍

• Deduplication by sessionId first-seen in loadHermesUsageEntries is consistent with OpenCode's approach. 👍

@coderabbitai review it

@cubic-dev-ai review it

@cubic-dev-ai review it

@ryoppippi I have started the AI code review. It will take a few minutes to complete.

@ryoppippi Sure, I'll kick off a fresh review of PR #1026 now!

@coderabbitai review it

@cubic-dev-ai review it

@cubic-dev-ai review it

@ryoppippi I have started the AI code review. It will take a few minutes to complete.

@ryoppippi Sure! Triggering a fresh review of PR #1026 now.

@coderabbitai review

Updated this PR to merge current main and port Hermes Agent support to the Rust CLI implementation.

This now wires Hermes into direct reports and all-agent reports, reads state.db from HERMES_HOME/default ~/.hermes, preserves recorded costs/messageCount, and includes reasoning tokens in total-token accounting and cost calculation.

Local validation passed:

• nix develop --command cargo test --manifest-path rust/Cargo.toml --workspace hermes
• nix develop --command pnpm run format
• nix develop --command cargo test --manifest-path rust/Cargo.toml --workspace
• nix develop --command pnpm typecheck
• nix develop --command pnpm run test

#!/bin/bash
# Find all Rust files related to Hermes
fd -e rs hermes rust/
echo "---FILE_BOUNDARY---"
# Look at Rust adapter structure
tree rust/ --gitignore 2>/dev/null | head -80
echo "---FILE_BOUNDARY---"
# Find Rust source files related to Hermes
rg -rn --type=rust "hermes" rust/ -l

#!/bin/bash
cat rust/crates/ccusage/src/adapter/hermes.rs
echo "---FILE_BOUNDARY---"
cat rust/crates/ccusage/src/adapter/mod.rs
echo "---FILE_BOUNDARY---"
cat rust/crates/ccusage/src/adapter/all.rs
echo "---FILE_BOUNDARY---"
cat rust/crates/ccusage/src/cli.rs
echo "---FILE_BOUNDARY---"
cat rust/crates/ccusage/src/config.rs
echo "---FILE_BOUNDARY---"
cat rust/crates/ccusage/src/config_schema.rs
echo "---FILE_BOUNDARY---"
cat rust/crates/ccusage/src/main.rs
echo "---FILE_BOUNDARY---"
# Check if there are any test files for Hermes in Rust
rg -rn --type=rust "hermes" rust/ -C 3

@ryoppippi Thanks for the update — the Rust core logic in hermes.rs looks solid (SQLite loading, cost calculation, reasoning token accounting, deduplication, test coverage). However, the rg scan reveals a critical rename artifact across several files where an intermediate identifier "n" was used instead of "hermes". These would break the Rust CLI silently or noisily at runtime.

Review: Rust CLI port — Hermes Agent

🔴 Critical: Stale "n" identifier scattered across 7 files

It looks like "hermes" was partially renamed from an intermediate name "n" and the rename didn't propagate consistently. Here's the full audit:

rust/crates/ccusage/src/cli.rs

The is_command mismatch alone means ccusage hermes session would fail with "Unknown command 'hermes'" at runtime, while ccusage n session would accidentally work.

rust/crates/ccusage/src/config.rs

// Current (broken):
fn is_agent_command(command: &str) -> bool {
    matches!(command, "claude" | "codex" | "opencode" | "amp" | "n" | "pi")
}

// Should be:
fn is_agent_command(command: &str) -> bool {
    matches!(command, "claude" | "codex" | "opencode" | "amp" | "hermes" | "pi")
}

rust/crates/ccusage/src/config_schema.rs

// Current (broken) — config key "hermes" block silently ignored:
pub(crate) n: Option<HermesConfig>,

// Should be:
pub(crate) hermes: Option<HermesConfig>,

The schema test asserting "ccusage-config" properties also checks for "n" instead of "hermes", and the config example uses "n" as the key.

rust/crates/ccusage/src/main.rs

// Current (broken):
Some(Command::Hermes(args)) => adapter::n::run(args),

// Should be:
Some(Command::Hermes(args)) => adapter::hermes::run(args),

rust/crates/ccusage/src/adapter/all.rs

// Current (broken) — wrong module alias + wrong agent label string:
use crate::{
    adapter::{amp, codex, n, opencode, pi},
    ...
};
// ...
append_agent_rows(&mut rows, &mut detected_agents, "n", load_n_rows(...));
// ...
fn load_n_rows(...) -> Result<AgentRows> {
    let mut entries = n::load_entries(shared, pricing)?;
    let summaries = n::summarize_entries(&entries, kind)?;
    Ok(AgentRows { rows: summary_rows("n", summaries), ... })
}

This means the all-sources report would label Hermes data as "n" (no matching arm in agent_label), showing raw "n" in the terminal table.

Also note agent_label has no "hermes" arm — it would fall through to _ => agent, displaying the raw string. Add:

"hermes" => "Hermes Agent",

rust/crates/ccusage/src/adapter/hermes.rs (internal inconsistencies)

// Default home directory:
vec![home.join(".n")]          // should be: home.join(".hermes")

// Fallback provider:
} else { "n" }                 // should be: "hermes"

// cost function name:
fn calculate_n_cost(...)       // should be: calculate_hermes_cost

// message ID prefix:
id: Some(format!("n:{}", ...)) // should be: format!("hermes:{}", ...)

// project arc:
project: Arc::from("n")        // should be: Arc::from("hermes")

// model_candidates guard:
if entry.provider != "n" {...} // should be: entry.provider != "hermes"

// path discovery:
fn n_state_db_paths()          // should be: hermes_state_db_paths()

// test temp dir:
format!("ccusage-n-{name}-...")  // cosmetic, should be: "ccusage-hermes-..."

// test name:
fn loads_billable_n_sessions_from_state_db()  // cosmetic

// test assertion:
id: Some("n:session-1".to_string())  // should be: "hermes:session-1"
project: Arc::from("n")              // should be: Arc::from("hermes")

The .n default path is the most impactful here: with HERMES_HOME unset and ~/.hermes being the correct location, the binary would look in ~/.n and find nothing.

🟡 Previously raised issues (TypeScript side — still applicable)

The earlier findings from the prior reviews remain open:

1. createHermesStateDb duplicated in 3 places — schema drift risk
2. PROVIDER_ALIASES too narrow — mistral, cohere, deepseek, etc. fall back to "hermes" and miss the provider/model pricing candidate
3. Silent $0 fallback in calculateHermesCost when no pricing is found
4. o\d regex should be o\d+ for future multi-digit o-series model names

✅ What looks good in the Rust port

• SQLite loading with open_with_flags(...with_read_only()) — correct isolation
• read_u64/read_f64/read_non_negative_f64 helpers are well-structured
• Session deduplication via HashSet is consistent with other adapters
• Reasoning tokens wired into extra_total_tokens / cost candidates correctly
• messageCount propagated into LoadedEntry
• Two real unit tests (loads_billable_hermes_sessions_from_state_db, report_includes_message_count_and_reasoning_total) cover the happy path

ccusage performance comparison

PR SHA: 6b46b1ddbfb8
Base SHA: 1727a60f7df5

This compares the PR package against the configured base package on the same CI runner.

Package runner startup

Execution setup measures any pre-benchmark package materialization used by the execution benchmark. Bunx temp cache measures one bunx -p <url> ccusage --version run with an empty Bun install cache. Warm reuses that cache and reports the median of repeated runs.

Cached bunx execution performance

Runs the same large fixture through bunx -p <pkg.pr.new URL> ccusage after the Bun install cache has already been populated by the startup measurement. This separates cached package-runner execution from first-fetch package materialization.

Fixtures: Claude /home/runner/work/_temp/ccusage-large-fixture (1.01 GiB, 2,597 files), Codex /home/runner/work/_temp/ccusage-large-codex-fixture (1.01 GiB, 2,597 files)
Base package: 1727a60f7df5; PR package: 6b46b1ddbfb8. Both run through bunx -p <pkg.pr.new URL> ccusage using the warmed Bun install cache from package runner startup, measured by hyperfine with 0 warmups and 1 runs.

Package runtime diagnostics

Compares the PR package wrapper, the installed native optional dependency binary, and the workspace release binary on the same large fixture. This identifies whether slow package results come from JavaScript wrapper overhead, the published native binary build, or the Rust core itself.

Fixtures: Claude /home/runner/work/_temp/ccusage-large-fixture (1.01 GiB, 2,597 files), Codex /home/runner/work/_temp/ccusage-large-codex-fixture (1.01 GiB, 2,597 files)
All rows run --offline --json, measured by hyperfine with 0 warmups and 1 runs. This isolates wrapper overhead from the installed native optional dependency and the workspace release binary built on the runner.

Committed fixture performance

Committed small fixtures for stable PR-to-PR feedback and explicit Claude/Codex command coverage.

Fixtures: Claude apps/ccusage/test/fixtures/claude (0.00 MiB, 2 files), Codex apps/ccusage/test/fixtures/codex (0.00 MiB, 1 files)
Base runs the published ccusage package from pkg.pr.new, installed before measurement; PR runs the published ccusage package from pkg.pr.new, installed before measurement. Both run --offline --json, measured by hyperfine with 2 warmups and 7 runs.

Large real-world-shaped fixture performance

Generated fixtures shaped from aggregate local log statistics: thousands of JSONL files, many small sessions, and a long tail of larger sessions. No real prompts, paths, or outputs are stored in the fixtures.

Fixtures: Claude /home/runner/work/_temp/ccusage-large-fixture (1.01 GiB, 2,597 files), Codex /home/runner/work/_temp/ccusage-large-codex-fixture (1.01 GiB, 2,597 files)
Base runs the published ccusage package from pkg.pr.new, installed before measurement; PR runs the published ccusage package from pkg.pr.new, installed before measurement. Both run --offline --json, measured by hyperfine with 0 warmups and 1 runs.

Artifact size

Lower medians and smaller artifacts are better. CI runner noise still applies; use same-run ratios as directional PR feedback, not release guarantees.

@coderabbitai review

Pushed the Hermes branch again after #1032 Kilo was merged into main.

Changes in this push:

• merged the latest origin/main into the Hermes adapter branch
• kept the Hermes Rust adapter wired into CLI/all-report/config schema/progress code alongside Kilo
• kept Hermes documented as an experimental data source
• added Kilo compatibility for the Hermes branch's message_count field on LoadedEntry
• regenerated and formatted the config schema

Hermes-specific fix still included:

• bundled fallback pricing covers gpt-5.5 and grok-4.3, so Hermes rows for those models no longer collapse to $0.00 just because LiteLLM returns zero loaded pricing

Local validation:

• nix develop --command cargo test --manifest-path rust/Cargo.toml -p ccusage hermes
• nix develop --command pnpm run format
• nix develop --command cargo test --manifest-path rust/Cargo.toml --workspace
• nix develop --command pnpm typecheck
• nix develop --command pnpm run test

Pre-push validation also passed clippy/oxfmt/eslint hooks.

#!/bin/bash
# Check calculate_hermes_cost for any debug/warn logging
rg -n -A 15 "fn calculate_hermes_cost" rust/crates/ccusage/src/adapter/hermes.rs
echo "---FILE_BOUNDARY---"
# Check agent_label in all.rs
rg -n -A 25 'fn agent_label' rust/crates/ccusage/src/adapter/all.rs
echo "---FILE_BOUNDARY---"
# Check Kilo adapter for message_count compatibility
rg -n "message_count" rust/crates/ccusage/src/adapter/kilo.rs 2>/dev/null || echo "kilo.rs not found or no message_count"
echo "---FILE_BOUNDARY---"
# Check all "n" artifacts remain gone
rg -n '"n"' rust/crates/ccusage/src/ --type=rust | grep -v "//\|test\|\"n/a\"\|\"n \"\|n_count\|file_n\|spawn\|stdin\|return\|\"\\\\n\"" | head -30
echo "---FILE_BOUNDARY---"
# Check progress.rs for Hermes and Kilo
rg -n "Hermes\|Kilo\|agent_label" rust/crates/ccusage/src/progress.rs
echo "---FILE_BOUNDARY---"
# Check config_schema.rs for hermes field
rg -n "hermes\|HermesConfig" rust/crates/ccusage/src/config_schema.rs | head -20
echo "---FILE_BOUNDARY---"
# Check load_hermes_rows in all.rs
rg -n -A 20 "fn load_hermes_rows" rust/crates/ccusage/src/adapter/all.rs
echo "---FILE_BOUNDARY---"
# Check for Kilo in all.rs to confirm both are wired
rg -n "kilo\|hermes" rust/crates/ccusage/src/adapter/all.rs | head -30
echo "---FILE_BOUNDARY---"
# Check cli.rs for both hermes and kilo commands
rg -n "hermes\|kilo\|Hermes\|Kilo" rust/crates/ccusage/src/cli.rs | head -30

#!/bin/bash
# Get the tail of calculate_hermes_cost to see the return 0.0 path and any logging
sed -n '375,410p' rust/crates/ccusage/src/adapter/hermes.rs
echo "---FILE_BOUNDARY---"
# Check progress.rs directly
cat rust/crates/ccusage/src/progress.rs
echo "---FILE_BOUNDARY---"
# Check config_schema.rs for hermes entries
grep -n "hermes\|HermesConfig\|Hermes" rust/crates/ccusage/src/config_schema.rs | head -30
echo "---FILE_BOUNDARY---"
# Check cli.rs for hermes command
grep -n "hermes\|Hermes" rust/crates/ccusage/src/cli.rs | head -30
echo "---FILE_BOUNDARY---"
# Check all.rs for both hermes and kilo imports
grep -n "use crate\|hermes\|kilo" rust/crates/ccusage/src/adapter/all.rs | head -20

@ryoppippi The Kilo merge looks clean — all wiring is intact and consistent. Here's the updated status:

✅ Everything looks good

🟡 One persistent minor nit

calculate_hermes_cost returns 0.0 silently (lines 393–396):

    for candidate in model_candidates(entry) {
        // ...
        if cost >= 0.0 && cost.is_finite() && cost > 0.0 { return cost; }
    }
    0.0  // ← still no diagnostic

The embedded pricing now covers the main frontier models, so this is low-impact in practice. If you'd like to make unknown future models diagnosable without users needing to know about LOG_LEVEL, a single call consistent with the rest of the file would do it:

    crate::debug_log(|| format!("hermes: no pricing found for model '{}' (provider: '{}'); cost=0", entry.model, entry.provider));
    0.0

Otherwise the PR is in great shape and ready to go. Up to you whether to add the diagnostic before merging.

@coderabbitai review

Pushed a follow-up for the remaining Hermes review thread.

Changes in this push:

• now tries provider-qualified Hermes pricing keys before the bare model
• added as a regression

Validation:

• nix develop --command cargo test --manifest-path rust/Cargo.toml -p ccusage hermes
• pre-push clippy/oxfmt hooks passed

@coderabbitai review

Correcting the previous review request formatting: pushed a follow-up for the remaining Hermes review thread.

Changes in this push:

• model_candidates() now tries provider-qualified Hermes pricing keys before the bare model
• added tries_provider_qualified_model_candidate_first as a regression

Validation:

• nix develop --command cargo test --manifest-path rust/Cargo.toml -p ccusage hermes
• pre-push clippy/oxfmt hooks passed

#!/bin/bash
# Check calculate_hermes_cost and model_candidates
rg -n -A 40 "fn calculate_hermes_cost\|fn model_candidates" rust/crates/ccusage/src/adapter/hermes.rs
echo "---FILE_BOUNDARY---"
# Check for regression test
rg -n -A 30 "provider_qualified\|regression\|provider.*model.*candidate\|candidate.*provider" rust/crates/ccusage/src/adapter/hermes.rs | head -80
echo "---FILE_BOUNDARY---"
# Check debug_log calls in hermes.rs
rg -n "debug_log\|warn\|eprintln" rust/crates/ccusage/src/adapter/hermes.rs
echo "---FILE_BOUNDARY---"
# See the full bottom of hermes.rs (where tests and cost functions usually are)
wc -l rust/crates/ccusage/src/adapter/hermes.rs
tail -120 rust/crates/ccusage/src/adapter/hermes.rs

@ryoppippi The fix looks correct and the regression test is solid. Here's the final status:

✅ All issues resolved

🟡 One persistent minor nit (acknowledged, not blocking)

calculate_hermes_cost still returns 0.0 silently for truly unknown future models — no debug_log call before the zero return. You've consistently chosen not to add this diagnostic, which is totally fine given the embedded pricing now covers the main frontier models. Just flagging it one last time in case you want to add it before merge:

crate::debug_log(|| format!("hermes: no pricing found for '{}'/'{}'", entry.provider, entry.model));
0.0

Otherwise, the PR is in great shape — the implementation is clean, all renames are correct, tests cover happy paths and regression cases, and Kilo/main parity is maintained. 🐇

#!/bin/bash
# Check the updated model_candidates function
rg -n -A 20 "fn model_candidates" rust/crates/ccusage/src/adapter/hermes.rs
echo "---FILE_BOUNDARY---"
# Check calculate_hermes_cost for any debug logging
rg -n -A 20 "fn calculate_hermes_cost" rust/crates/ccusage/src/adapter/hermes.rs
echo "---FILE_BOUNDARY---"
# Check the new regression test
rg -n -A 30 "tries_provider_qualified_model_candidate_first" rust/crates/ccusage/src/adapter/hermes.rs
echo "---FILE_BOUNDARY---"
# Check all tests in hermes.rs to understand coverage
rg -n "#\[test\]\|fn test\|fn loads_\|fn report_\|fn calculates_\|fn embedded_\|fn tries_" rust/crates/ccusage/src/adapter/hermes.rs

@ryoppippi The model_candidates() fix looks correct and the regression test is well-targeted. Here's the verification:

✅ model_candidates() ordering — FIXED

fn model_candidates(entry: &HermesEntry) -> Vec<String> {
    let mut candidates = Vec::new();
    if entry.provider != "hermes" {
        candidates.push(format!("{}/{}", entry.provider, entry.model));  // ← tried first
    }
    candidates.push(entry.model.clone());  // ← bare model as fallback
    // dedup via HashSet
}

Provider-qualified form is now the first candidate for known providers, which means openai/gpt-5.5 gets looked up before gpt-5.5. The HashSet deduplication correctly handles the edge case where provider and model prefix coincide.

✅ tries_provider_qualified_model_candidate_first — correct regression test

assert_eq!(
    model_candidates(&entry),
    vec!["openai/gpt-5.5".to_string(), "gpt-5.5".to_string()]
);

Exactly the right assertion — verifies both ordering and completeness.

🟡 One persistent minor nit (same as prior reviews)

calculate_hermes_cost returns 0.0 silently at line 395. With embedded pricing now covering the main frontier models and provider-qualified lookup now tried first, the practical impact is low. But a crate::debug_log(...) before the 0.0 would make future unknown-model cases diagnosable via LOG_LEVEL:

    crate::debug_log(|| format!("hermes: no pricing found for model '{}' (provider: '{}'); cost=0", entry.model, entry.provider));
    0.0

Entirely up to you — the PR is in great shape otherwise.

ccusage performance comparison

PR SHA: 6a2c6e02d790
Base SHA: 35cb05e38459

This compares the PR package against the configured base package on the same CI runner.

Package runner startup

Execution setup measures any pre-benchmark package materialization used by the execution benchmark. Bunx temp cache measures one bunx -p <url> ccusage --version run with an empty Bun install cache. Warm reuses that cache and reports the median of repeated runs.

Cached bunx execution performance

Runs the same large fixture through bunx -p <pkg.pr.new URL> ccusage after the Bun install cache has already been populated by the startup measurement. This separates cached package-runner execution from first-fetch package materialization.

Fixtures: Claude /home/runner/work/_temp/ccusage-large-fixture (1.01 GiB, 2,597 files), Codex /home/runner/work/_temp/ccusage-large-codex-fixture (1.01 GiB, 2,597 files)
Base package: 35cb05e38459; PR package: 6a2c6e02d790. Both run through bunx -p <pkg.pr.new URL> ccusage using the warmed Bun install cache from package runner startup, measured by hyperfine with 0 warmups and 1 runs.

Package runtime diagnostics

Compares the PR package wrapper, the installed native optional dependency binary, and the workspace release binary on the same large fixture. This identifies whether slow package results come from JavaScript wrapper overhead, the published native binary build, or the Rust core itself.

Fixtures: Claude /home/runner/work/_temp/ccusage-large-fixture (1.01 GiB, 2,597 files), Codex /home/runner/work/_temp/ccusage-large-codex-fixture (1.01 GiB, 2,597 files)
All rows run --offline --json, measured by hyperfine with 0 warmups and 1 runs. This isolates wrapper overhead from the installed native optional dependency and the workspace release binary built on the runner.

Committed fixture performance

Committed small fixtures for stable PR-to-PR feedback and explicit Claude/Codex command coverage.

Fixtures: Claude apps/ccusage/test/fixtures/claude (0.00 MiB, 2 files), Codex apps/ccusage/test/fixtures/codex (0.00 MiB, 1 files)
Base runs the published ccusage package from pkg.pr.new, installed before measurement; PR runs the published ccusage package from pkg.pr.new, installed before measurement. Both run --offline --json, measured by hyperfine with 2 warmups and 7 runs.

Large real-world-shaped fixture performance

Generated fixtures shaped from aggregate local log statistics: thousands of JSONL files, many small sessions, and a long tail of larger sessions. No real prompts, paths, or outputs are stored in the fixtures.

Fixtures: Claude /home/runner/work/_temp/ccusage-large-fixture (1.01 GiB, 2,597 files), Codex /home/runner/work/_temp/ccusage-large-codex-fixture (1.01 GiB, 2,597 files)
Base runs the published ccusage package from pkg.pr.new, installed before measurement; PR runs the published ccusage package from pkg.pr.new, installed before measurement. Both run --offline --json, measured by hyperfine with 0 warmups and 1 runs.

Artifact size

Lower medians and smaller artifacts are better. CI runner noise still applies; use same-run ratios as directional PR feedback, not release guarantees.

ccusage performance comparison

PR SHA: 6a2c6e02d790
Base SHA: 35cb05e38459

This compares the Rust PR release binary against the configured base package on the same CI runner.

Package runner startup

Execution setup measures any pre-benchmark package materialization used by the execution benchmark. Bunx temp cache measures one bunx -p <url> ccusage --version run with an empty Bun install cache. Warm reuses that cache and reports the median of repeated runs.

Cached bunx execution performance

Runs the same large fixture through bunx -p <pkg.pr.new URL> ccusage after the Bun install cache has already been populated by the startup measurement. This separates cached package-runner execution from first-fetch package materialization.

Fixtures: Claude /home/runner/work/_temp/ccusage-large-fixture (1.01 GiB, 2,597 files), Codex /home/runner/work/_temp/ccusage-large-codex-fixture (1.01 GiB, 2,597 files)
Base package: 35cb05e38459; PR package: 6a2c6e02d790. Both run through bunx -p <pkg.pr.new URL> ccusage using the warmed Bun install cache from package runner startup, measured by hyperfine with 0 warmups and 1 runs.

Package runtime diagnostics

Compares the PR package wrapper, the installed native optional dependency binary, and the workspace release binary on the same large fixture. This identifies whether slow package results come from JavaScript wrapper overhead, the published native binary build, or the Rust core itself.

Fixtures: Claude /home/runner/work/_temp/ccusage-large-fixture (1.01 GiB, 2,597 files), Codex /home/runner/work/_temp/ccusage-large-codex-fixture (1.01 GiB, 2,597 files)
All rows run --offline --json, measured by hyperfine with 0 warmups and 1 runs. This isolates wrapper overhead from the installed native optional dependency and the workspace release binary built on the runner.

Committed fixture performance

Committed small fixtures for stable PR-to-PR feedback and explicit Claude/Codex command coverage.

Fixtures: Claude apps/ccusage/test/fixtures/claude (0.00 MiB, 2 files), Codex apps/ccusage/test/fixtures/codex (0.00 MiB, 1 files)
Base runs the published ccusage package from pkg.pr.new, installed before measurement; PR runs rust/target/release/ccusage directly. Both run --offline --json, measured by hyperfine with 2 warmups and 7 runs.

Large real-world-shaped fixture performance

Generated fixtures shaped from aggregate local log statistics: thousands of JSONL files, many small sessions, and a long tail of larger sessions. No real prompts, paths, or outputs are stored in the fixtures.

Fixtures: Claude /home/runner/work/_temp/ccusage-large-fixture (1.01 GiB, 2,597 files), Codex /home/runner/work/_temp/ccusage-large-codex-fixture (1.01 GiB, 2,597 files)
Base runs the published ccusage package from pkg.pr.new, installed before measurement; PR runs rust/target/release/ccusage directly. Both run --offline --json, measured by hyperfine with 0 warmups and 1 runs.

Artifact size

Lower medians and smaller artifacts are better. CI runner noise still applies; use same-run ratios as directional PR feedback, not release guarantees.

ccusage performance comparison

PR SHA: 338d94cf1345
Base SHA: 35cb05e38459

This compares the Rust PR release binary against the configured base package on the same CI runner.

Package runner startup

Execution setup measures any pre-benchmark package materialization used by the execution benchmark. Bunx temp cache measures one bunx -p <url> ccusage --version run with an empty Bun install cache. Warm reuses that cache and reports the median of repeated runs.

Cached bunx execution performance

Runs the same large fixture through bunx -p <pkg.pr.new URL> ccusage after the Bun install cache has already been populated by the startup measurement. This separates cached package-runner execution from first-fetch package materialization.

Fixtures: Claude /home/runner/work/_temp/ccusage-large-fixture (1.01 GiB, 2,597 files), Codex /home/runner/work/_temp/ccusage-large-codex-fixture (1.01 GiB, 2,597 files)
Base package: 35cb05e38459; PR package: 338d94cf1345. Both run through bunx -p <pkg.pr.new URL> ccusage using the warmed Bun install cache from package runner startup, measured by hyperfine with 0 warmups and 1 runs.

Package runtime diagnostics

Compares the PR package wrapper, the installed native optional dependency binary, and the workspace release binary on the same large fixture. This identifies whether slow package results come from JavaScript wrapper overhead, the published native binary build, or the Rust core itself.

Fixtures: Claude /home/runner/work/_temp/ccusage-large-fixture (1.01 GiB, 2,597 files), Codex /home/runner/work/_temp/ccusage-large-codex-fixture (1.01 GiB, 2,597 files)
All rows run --offline --json, measured by hyperfine with 0 warmups and 1 runs. This isolates wrapper overhead from the installed native optional dependency and the workspace release binary built on the runner.

Committed fixture performance

Committed small fixtures for stable PR-to-PR feedback and explicit Claude/Codex command coverage.

Fixtures: Claude apps/ccusage/test/fixtures/claude (0.00 MiB, 2 files), Codex apps/ccusage/test/fixtures/codex (0.00 MiB, 1 files)
Base runs the published ccusage package from pkg.pr.new, installed before measurement; PR runs rust/target/release/ccusage directly. Both run --offline --json, measured by hyperfine with 2 warmups and 7 runs.

Large real-world-shaped fixture performance

Generated fixtures shaped from aggregate local log statistics: thousands of JSONL files, many small sessions, and a long tail of larger sessions. No real prompts, paths, or outputs are stored in the fixtures.

Fixtures: Claude /home/runner/work/_temp/ccusage-large-fixture (1.01 GiB, 2,597 files), Codex /home/runner/work/_temp/ccusage-large-codex-fixture (1.01 GiB, 2,597 files)
Base runs the published ccusage package from pkg.pr.new, installed before measurement; PR runs rust/target/release/ccusage directly. Both run --offline --json, measured by hyperfine with 0 warmups and 1 runs.

Artifact size

Lower medians and smaller artifacts are better. CI runner noise still applies; use same-run ratios as directional PR feedback, not release guarantees.

ccusage performance comparison

PR SHA: 338d94cf1345
Base SHA: 35cb05e38459

This compares the PR package against the configured base package on the same CI runner.

Package runner startup

Execution setup measures any pre-benchmark package materialization used by the execution benchmark. Bunx temp cache measures one bunx -p <url> ccusage --version run with an empty Bun install cache. Warm reuses that cache and reports the median of repeated runs.

Cached bunx execution performance

Runs the same large fixture through bunx -p <pkg.pr.new URL> ccusage after the Bun install cache has already been populated by the startup measurement. This separates cached package-runner execution from first-fetch package materialization.

Fixtures: Claude /home/runner/work/_temp/ccusage-large-fixture (1.01 GiB, 2,597 files), Codex /home/runner/work/_temp/ccusage-large-codex-fixture (1.01 GiB, 2,597 files)
Base package: 35cb05e38459; PR package: 338d94cf1345. Both run through bunx -p <pkg.pr.new URL> ccusage using the warmed Bun install cache from package runner startup, measured by hyperfine with 0 warmups and 1 runs.

Package runtime diagnostics

Compares the PR package wrapper, the installed native optional dependency binary, and the workspace release binary on the same large fixture. This identifies whether slow package results come from JavaScript wrapper overhead, the published native binary build, or the Rust core itself.

Fixtures: Claude /home/runner/work/_temp/ccusage-large-fixture (1.01 GiB, 2,597 files), Codex /home/runner/work/_temp/ccusage-large-codex-fixture (1.01 GiB, 2,597 files)
All rows run --offline --json, measured by hyperfine with 0 warmups and 1 runs. This isolates wrapper overhead from the installed native optional dependency and the workspace release binary built on the runner.

Committed fixture performance

Committed small fixtures for stable PR-to-PR feedback and explicit Claude/Codex command coverage.

Fixtures: Claude apps/ccusage/test/fixtures/claude (0.00 MiB, 2 files), Codex apps/ccusage/test/fixtures/codex (0.00 MiB, 1 files)
Base runs the published ccusage package from pkg.pr.new, installed before measurement; PR runs the published ccusage package from pkg.pr.new, installed before measurement. Both run --offline --json, measured by hyperfine with 2 warmups and 7 runs.

Large real-world-shaped fixture performance

Generated fixtures shaped from aggregate local log statistics: thousands of JSONL files, many small sessions, and a long tail of larger sessions. No real prompts, paths, or outputs are stored in the fixtures.

Fixtures: Claude /home/runner/work/_temp/ccusage-large-fixture (1.01 GiB, 2,597 files), Codex /home/runner/work/_temp/ccusage-large-codex-fixture (1.01 GiB, 2,597 files)
Base runs the published ccusage package from pkg.pr.new, installed before measurement; PR runs the published ccusage package from pkg.pr.new, installed before measurement. Both run --offline --json, measured by hyperfine with 0 warmups and 1 runs.

Artifact size