{"entries":[{"lastCommit":{"author":{"email":"1560508+ryoppippi@users.noreply.github.com","name":"ryoppippi","when":"2026-07-27T12:16:43.000000+01:00"},"committer":{"email":"noreply@github.com","name":"GitHub","when":"2026-07-27T12:16:43.000000+01:00"},"hash":"86788c3a37809103eeab14a2910adf916ce663c1","message":"perf(rust): split the workspace into independently cached crates (#1428)\n\n* refactor(rust): extract shared core crate\n\nMove shared types, pricing, configuration, output, date, and block logic into ccusage-core so adapter crates can depend on an acyclic foundation. Keep the ccusage package focused on command dispatch and move Codex-owned types into its adapter boundary.\n\nCo-authored-by: ryoppippi \n\n* test(core): move snapshots with extracted modules\n\nKeep Insta snapshot ownership aligned with the new ccusage-core crate so its existing characterization tests continue to compare the same outputs.\n\nCo-authored-by: ryoppippi \n\n* refactor(adapters): extract shared adapter helpers\n\nMove JSONL parsing, file walking, date filtering, size-balanced chunking, and ordered parallel reads into ccusage-adapter-common. This removes the Claude adapter as an infrastructure dependency before agent crates are split.\n\nCo-authored-by: ryoppippi \n\n* refactor(adapters): split agents into workspace crates\n\nMove every agent adapter and the unified all-agent aggregator into dedicated ccusage-adapter-* crates. Preserve their tests and snapshots with each owning crate, keep shared infrastructure in ccusage-adapter-common, and leave ccusage as thin command wiring.\n\nCo-authored-by: ryoppippi \n\n* perf(nix): cache real adapter artifacts\n\nBuild the shared foundation and each adapter as filtered Crane cargoBuild derivations, merge their real target artifacts, and compile only the unified adapter plus thin binary at the final layers. Apply the same target-specific graph to native, musl, and Darwin cross packages.\n\nCo-authored-by: ryoppippi \n\n* fix(core): keep test-only pricing length private\n\nCo-authored-by: ryoppippi \n\n* docs(rust): route adapter work to crates\n\nUpdate repository guidance and source references for the new ccusage-adapter-* workspace layout and shared core/common boundaries.\n\nCo-authored-by: ryoppippi \n\n* fix(nix): use dependency artifacts on Darwin\n\nCrane emits full artifact archives on macOS, so merging sibling adapter archives can overwrite shared crate variants. Keep per-adapter layers on Linux and fall back to the safe dependency-only cache for native and cross Darwin packages.\n\nCo-authored-by: ryoppippi \n\n* refactor(core): own shared adapter reporting\n\nMove generic JSON row and period helpers from the OpenCode adapter into ccusage-core. Remove the resulting OpenCode dependency from leaf adapters and trim obsolete direct dependencies from the binary crate.\n\n* fix(pricing): follow extracted asset paths\n\nRegenerate, format, diff, and stage models.dev and Codex fallback snapshots at their new owning crate paths. Watch the moved config schema source in treefmt as well.\n\n* perf(nix): isolate and retain current adapter artifacts\n\nBuild independent adapters directly from the shared foundation. Root the current Linux native artifact graph and garbage-collect superseded store generations before saving that cache, preventing per-commit artifact accumulation.\n\n* refactor(adapters): keep implementation modules private\n\nExpose only each adapter's intentional command and loader surface while keeping parser, path, report, and implementation modules crate-private.\n\n* fix(core): harden date parsing against overflow and invalid offsets\n\nUse checked arithmetic in parse_digits so multi-digit inputs cannot silently overflow, and reject timezone offsets whose hours exceed 23 or minutes exceed 59.\n\nCo-authored-by: Codesmith \n\n* refactor(rust): centralize adapter reporting and prune unused deps\n\nMove print_table_for_agent into ccusage-adapter-common so goose and codebuff no longer depend on the amp crate, and update the Nix artifact layering accordingly. Remove unused hermes dependencies and drop the bundled sqlite dependency from the JSONL-only copilot, gemini, openclaw, and qwen adapters.\n\nCo-authored-by: Codesmith \n\n* refactor(claude): reuse shared parallel file reader\n\nLoad Claude usage files through ccusage_adapter_common::read_files_parallel and drop the crate-local read_usage_files_parallel duplicate, so improvements to the shared JSONL parallelism apply to Claude too.\n\nCo-authored-by: Codesmith \n\n* chore(rust): drop unused adapter deps and imports\n\nRemove the unused ccusage-terminal dependency from ccusage-adapter-all, prune ccusage-adapter-droid down to the crates it actually uses (dropping bundled sqlite, compact_str, memchr, rustc-hash, smallvec, ccusage-terminal, serde, and the insta dev-dependency), and remove the stale allow(unused_imports) plus dead common-helper imports from the Pi adapter.\n\nCo-authored-by: Codesmith \n\n* fix(report): align agent table totals and weekly labels with JSON\n\nThe shared agent usage table now counts extra_total_tokens by using row.total_tokens() per row and reading totalTokens from the computed totals for the footer, so reasoning and fallback tokens no longer under-report. It also includes row.week in the label fallback chain so --weekly rows keep their period identifiers, matching the JSON report and the core table renderer.\n\nCo-authored-by: Codesmith \n\n* perf(nix): collapse the Crane artifact layers into a linear chain\n\nThe per-adapter fan-out could not be used on Darwin at all, because Crane\ninstalls full rather than incremental archives there\n(rust-lang/rust#115982) and merging sibling full archives can overwrite\nshared path-crate artifacts with incompatible variants. macOS therefore fell\nback to the dependency-only cache and recompiled the whole workspace on\nevery build.\n\nA linear chain only ever extends the previous layer, so it needs no merging\nand works on every platform. The adapters share one layer because they are\nmutually independent: one cargoBuild with a -p per crate lets Cargo compile\nthem concurrently, so the layer costs about as much as its slowest adapter,\nwhile the fan-out added five serial Nix round-trips that each unpack and\nre-archive the dependency target directory.\n\nLinux, macOS arm64, and the cross-compiled macOS x64 build now all reuse the\nsame two layers.\n\n* perf(release): fold identical code in the Linux link\n\nSplitting the runtime into per-adapter crates costs binary size: each crate\ninstantiates the shared generic machinery for its own types and stable Rust\ncannot share those instantiations across crates, so the aarch64-darwin\nbinary grew 2,930,240 -> 3,096,896 bytes, all of it in __text. Linker\ndead-stripping recovers none of it — the duplicates are reachable code, not\ndead code — so mold's identical-code folding is the lever that applies.\n\n* perf(ci): build Nix derivations in parallel\n\nnix-quick-install-action writes the nix.conf it is given verbatim, so Nix\nkept its documented default of max-jobs = 1 and built every derivation one\nat a time — the independent flake checks as well as the artifact layers — on\n32-vCPU runners.\n\ncores stays at its default so the dependency derivation, which compiles the\nwhole third-party graph in a single build, still gets every core.\n\n* perf(nix): scope the clippy and rustfmt checks to the Rust source\n\nccusage-clippy compiles the entire workspace with --all-targets, and its\nsource was the whole repository, so a README, docs page, TypeScript file, or\nworkflow edit invalidated the most expensive check in `nix flake check`.\n\ntests.nix already had the right include list; it now lives in one place so\nthe two cannot drift.\n\n* perf(build): make the LiteLLM pricing download optional\n\nccusage-core's build script reached minreq with https-rustls, which enabled\nrustls's aws-lc-rs provider on top of the ring provider ureq already needs.\nCompiling aws-lc's C and assembly sources took longer than any other unit in\nthe dependency graph, and every workspace crate waited on it, because a build\nscript and its dependencies must be compiled and run before the crate that\nowns them is compiled.\n\nNo Nix build ever used that code path: the sandbox has no network, and every\nNix package hands the snapshot to the build through\nCCUSAGE_PRICING_JSON_PATH. The dev shell now exports the same pinned\nsnapshot, and the Windows release build — the one platform Nix cannot target\n— enables the feature to keep downloading it.\n\n* perf(adapters): finish pruning adapter imports and dependencies\n\nThe manifests and root imports still carried helpers and crates the adapters\nnever reach: sqlite stayed in 6 crates that open no database, insta in\nadapters with no snapshot test, ccusage-terminal in every adapter even though\nthey only see it through ccusage-core's re-exports, and the shared-loader\nimports were still wrapped in a module-wide unused_imports allow that hid\nwhich loaders each adapter depends on.\n\n* docs(rust): note the fetch-litellm-pricing build feature\n\n* perf(cli): split the argument parser out of ccusage-cli\n\nccusage-cli was the root of the dependency graph — ccusage-core and all 15\nadapters depend on it — and it also owned the parser, the help renderer, the\nbuild script, and the 30 KB of embedded help JSON. Touching help text\ntherefore recompiled 20 of 20 workspace units, and every crate waited on a\nbuild script before it could start.\n\nThe adapters and core only ever use the plain argument types, which stay in\nccusage-cli. The new ccusage-cli-parser owns Cli, the parser, the help\nrenderer, and the assets, and only the ccusage binary depends on it, so it\nsits outside the Crane artifact layers entirely: editing cli-help.json now\nleaves the foundation and adapter layers byte-identical (verified by\ncomparing the adapters layer derivation path before and after).\n\nThe config-merge tests move with the parser, since they assert how a parsed\ncommand line and a config file combine.\n\n* docs(rust): give every workspace crate a README\n\nThe split produced 23 crates whose boundaries are only visible in the\nmanifests. Each README now states what the crate owns, what belongs elsewhere,\nits public surface, and which Crane artifact layer it lands in — the last one\nbeing what determines how much a change to it costs. Adapter READMEs also name\nthe data source and environment variables, and link the existing `src/README.md`\nwhere one documents the record shapes.\n\n* fix(adapter-common): treat a zero chunk count as one chunk\n\nchunk_file_indexes_by_size became a cross-crate helper, and with a non-empty\nfile list and a chunk count of zero it indexed an empty vector. One chunk still\nreturns every index, which is what a caller asking for no parallelism wants.\n\nAlso corrects the build-script path in nix/rust-src.nix: the file that reads\nflake.lock is ccusage-core/build.rs.\n\n* docs(rust): derive each crate README's module list from its sources\n\nThe first pass used one template for every adapter, so READMEs listed modules\nthe crate does not have — the Claude adapter has no parser.rs or report.rs, and\nthe Codex adapter's aggregate, replay, speed, and types modules were missing.\nEach list now comes from the crate's own src directory, and the Pi default store\npath includes the sessions directory it actually reads.\n\n* docs(rust): point the skill at the split crate layout\n\nThe skill still described the Rust CLI as two crates. It now lists what each\ncrate owns, points at the per-crate READMEs for the artifact-layer cost, and\nstates the rule the amp/goose/codebuff coupling broke: adapters do not depend on\neach other, shared code goes to ccusage-adapter-common.\n\n* perf(core): move the blocks report into the binary\n\nblocks.rs and its SessionBlock, BurnRate, and Projection types had exactly one\nconsumer, the ccusage binary, but sat in ccusage-core, so 582 lines of\nbilling-block rendering recompiled every adapter whenever they changed. The two\nblock-only thresholds move with them; DEFAULT_SESSION_DURATION_HOURS stays in\ncore because the config layer defaults to it.\n\n* perf(core): extract ccusage-config from core\n\nconfig.rs and config_schema.rs were 28% of ccusage-core's source with exactly one\nconsumer, the binary, and config_schema is where `schemars` and its derive enter\nthe graph. Because core sits in the foundation artifact layer, that meant every\n`ccusage.json` field or schema change recompiled core, adapter-common, and all 15\nadapters, and every build compiled schemars before core could start.\n\n`cargo tree -p ccusage-core` no longer mentions schemars. The config-merge tests\nin ccusage-cli-parser reach the new crate as a dev-dependency.\n\n* refactor(core): let adapters name their own load-progress label\n\nprogress::UsageLoadAgent was a closed enum listing all 16 agents plus a match\nmapping each to a display string, so adding or renaming an agent meant editing\nccusage-core — the crate every adapter depends on and the one in the foundation\nartifact layer. It is now a newtype around the label, and each adapter passes its\nown, keeping the roster with the adapters. Labels are unchanged, including\npi-agent, GitHub Copilot CLI, and Gemini CLI.\n\n* perf(core): inject the pricing HTTP client from the binary\n\nureq entered ccusage-core through one 25-line function, and brought rustls, ring,\nwebpki, and about twenty more crates with it — all of which had to compile before\ncore, and therefore before adapter-common and every adapter. Core now takes the\nfetcher as a function pointer that main installs, and the ureq implementation\nlives with the binary.\n\n`cargo tree -p ccusage-core` no longer mentions ureq, rustls, ring, or webpki.\nWithout a fetcher installed a refresh reports that it is unavailable and the\nembedded snapshots are used, which is the existing --offline behaviour; the\nshipped binary installs one, verified by a live run still taking the network path.\n\n* refactor(rust): group the adapter crates and share manifest settings\n\nEach adapter manifest repeated the same package fields and dependency specs, so\nadding a crate meant copying a block and a version bump meant editing 23 files.\n`[workspace.package]` and `[workspace.dependencies]` now hold both, and each crate\nlists only what it uses — `jiff.workspace = true` rather than the four-line\nfeature spec. It also means moving a crate no longer touches its dependents,\nsince path dependencies live in one place.\n\nThe per-source adapters move to `rust/adapters/`, leaving `rust/crates`\nfor the crates that are not tied to a single source. `ccusage-adapter-common`\nlives with the adapters it serves, which also keeps the `adapters/*` member glob\nmatching in every filtered source tree a Crane layer builds from — a glob that\nmatches nothing is an error, and the foundation layer contains no other adapter.\n\n* perf(ci): stop the schema generator from pulling the adapter graph\n\ngenerate-config-schema was a second bin of the ccusage package, so building it\ncompiled every adapter even though it only calls into the config layer. It moves\nto ccusage-config, where a clean build is 3 units, and the Nix derivation starts\nfrom the foundation artifacts instead of the adapter ones.\n\nThat matters for wall clock because treefmt depends on this generator and the\npreflight job builds treefmt, and preflight gates the whole build matrix by\ndesign — so every build job was waiting on 15 adapters that the formatter never\nneeded.\n\n* perf(release): fold identical code in the macOS link too\n\nMeasured on the shipped Linux binary, mold's identical-code folding recovers\n158,720 of the 240,768 bytes the crate split costs (4,482,112 -> 4,323,392\nagainst 4,241,344 on main). ld64 has no equivalent, so the macOS binaries kept\nthe full cost; lld does, and the Nix build with it produces 3,232,928 bytes\nrather than 3,299,504.\n\nBoth macOS packages still link only /usr/lib and /System/Library dylibs, which\nthe postInstall gate enforces, and `--help` and the JSON reports are byte\nidentical to the ld64 build.\n\n* build(rust): move to 1.97.1 and configure hawk\n\nhawk lints a Cargo workspace for `pub` that the workspace itself never needs,\nwhich is exactly the surface a 24-crate split creates. It uses `rustc_private`\nand must run on the toolchain it was built against, so the pin moves to 1.97.1;\nthe dev shell picks that up from rust-toolchain.toml with no flake change.\n\nhawk.toml names the two production entry points rather than letting hawk assume\nevery binary target is shipped, so the reason each one is an entry point is\nwritten down. 1.97.1's clippy also wants a byte-string literal in\nparse_timezone_offset.\n\n* refactor(rust): narrow the visibility the crate split widened\n\nSplitting one crate into 24 turned a lot of module-private items into `pub`,\nbecause that is what crossing a crate boundary needs — but only some of them are\nactually crossed. hawk's closed-world analysis over the two production entry\npoints found 85 such declarations; this applies the 68 machine-applicable ones,\nturning `pub` into `pub(crate)` and dropping restricted modifiers that no longer\nrestrict anything.\n\n* refactor(rust): narrow the re-exports the crate split widened\n\nThe remaining hawk findings were public re-exports that no other crate reaches:\neach adapter re-exported report_from_rows even though only its own run() calls\nit, claude re-exported three path helpers used only inside itself, and codex\nre-exported two types its own parser and replay modules use. Splitting each\nstatement keeps every name visible where it is used and stops advertising it\npast the crate boundary.\n\nCli::parse had no callers at all — the binary goes through\nparse_from_with_config because it must pass a config context — so it is gone and\nits parse_from shorthand is now test-only. hawk reports no findings.\n\n* docs(rust): write down the workspace visibility rule and how to check it\n\nA 24-crate workspace needs a stated rule for `pub`, or the next crate will export\nwhatever compiles. The rust skill now says what the rule is, how `just hawk`\nchecks it, that hawk must be installed separately because it is not in nixpkgs,\nwhy the toolchain is pinned to 1.97.1, and that its findings are suggestions\nrather than a gate — it is explicitly experimental.\n\n* fix(nix): point the schema formatter at the moved generator\n\nThe schema-gen treefmt formatter still listed config_schema.rs under\nccusage-core and the generator under the ccusage package, so after both moved to\nccusage-config it never ran on the files whose changes should regenerate the\nschema. The config-schema flake check would still have caught the drift, but only\nafter a failing build instead of a formatting pass.\n\n---------\n\nCo-authored-by: Cursor Agent \nCo-authored-by: ryoppippi \nCo-authored-by: Codesmith ","tree":"6ecc56f4b9aff6d73b1c56750b102c17033d8c77"},"mode":"0040000","name":"build-linux-native-package","size":0},{"lastCommit":{"author":{"email":"1560508+ryoppippi@users.noreply.github.com","name":"ryoppippi","when":"2026-06-13T12:28:27.000000+01:00"},"committer":{"email":"noreply@github.com","name":"GitHub","when":"2026-06-13T12:28:27.000000+01:00"},"hash":"ed48f25e7a0a022f11d82f2f05c988cf72934e98","message":"build: migrate package tooling to nushell and node test (#1306)\n\n* build(ccusage): run native binary check with nushell\n\nReplace the Bun ensure-native-binary script with a Nushell script that keeps the same native package fast path, portable binary validation, cargo fallback build, and version check behavior.\n\nThis removes Bun as a required runtime for the package prepack native-binary check. The dev shell now provides Nushell, and the CI/release packaging jobs install nushell instead of bun for that prepack path while keeping Bun available for the remaining benchmark and fixture TypeScript scripts.\n\n* fix(ccusage): derive native package target from metadata\n\nAvoid duplicating the native package platform map in the Nushell prepack check. The script now scans the native package manifests and selects the package whose os and cpu metadata match the current platform.\n\nThis addresses cubic feedback about drift risk while preserving the same fast path and cargo fallback behavior.\n\n* build(ccusage): run nushell check from nix shebang\n\nMake the native-binary check executable and let the script acquire Nushell through a Nix shebang pinned to the repository flake inputs.\n\nThe package build no longer needs nushell installed into the lightweight CI or release profiles, and the dev shell does not need to add nushell just for this prepack script. The cargo fallback still works when the script is invoked from the project dev shell.\n\n* ci: upsert performance comments with gh api\n\nReplace the Bun-based upsert-pr-comment helper with a Nushell script that uses gh api through a Nix shebang. The perf jobs can now update PR comments without invoking Bun for the GitHub API wrapper.\n\nThe script preserves the existing update-or-create behavior and keeps 401/403 comment write failures non-fatal for restricted tokens.\n\n* fix(ci): update latest paginated PR comment\n\nFetch all issue comment pages before looking for the marker so busy PRs do not create duplicate performance comments after the first 100 comments. When several matching github-actions comments exist, choose the newest one so the visible current performance report is updated instead of an older historical comment.\n\n* build(ccusage): move performance comparison to nushell\n\nReplace the Bun/TypeScript PR performance comparison script with a Nushell script that carries its own Nix shebang. The perf workflow no longer needs repo-level pnpm install or Node for URL parsing; it installs only bun and just for fixture generation, while the comparison script resolves nu, bun, curl, git, hyperfine, just, and pnpm itself for the paths it owns.\n\nMove developer command bodies into apps/ccusage/justfile and keep package.json scripts as thin compatibility shims, with build and prepack left as direct lifecycle commands so pnpm pack and publish flows do not depend on just being visible inside npm lifecycle PATH. Drop gunshi after replacing the fixture generator CLI parser with a small local parser.\n\n* fix(ccusage): harden cli arg parsing and just recipe PATH\n\nUse pnpm exec for tsdown and clean-pkg-json in the package justfile so\njust build-package and just prepack work outside the pnpm lifecycle PATH.\n\nIn generate-large-fixture parseCliArgs, reject a flag-shaped next token\nso '--output-dir --codex-output-dir foo' fails fast instead of running\nwith '--codex-output-dir' as the output path.\n\nCo-authored-by: Codesmith \n\n* fix(ccusage): accurate head runtime description and drop dead helper\n\nWhen the PR head package is installed and exposes a native binary,\nthe rust-mode benchmark already uses that installed native binary, not\nrust/target/release/ccusage. Reflect that in the runtime description so\nthe perf comment does not claim the workspace release binary was used.\n\nAlso drop the unused run_command helper in compare-pr-performance.nu.\n\nCo-authored-by: Codesmith \n\n* fix(ccusage): preserve nested record columns in diagnostic results\n\ncompare_runtime_diagnostic_section built its results list with 'append'\nfollowed by '| flatten'. In Nushell, 'flatten' on a list of records\nexpands nested record columns, which turned each diagnostic result from\n{command, label, measurement: {median, ...}} into {command, label,\nmedian, ...} and dropped the 'measurement' column. Rendering then\nfailed with 'Cannot find column \\'measurement\\'' on the runtime\ndiagnostics section, breaking both ccusage perf comment and ccusage\nrust perf comment CI jobs.\n\nSwitch to list concatenation with '++' to merge per-command result\nlists without touching record structure, and remove the redundant\n'| flatten' on the section result.\n\nCo-authored-by: Codesmith \n\n* build: format nushell scripts with nufmt\n\nAdd nufmt to the treefmt formatter set so executable Nushell helpers are checked by the normal fmt/pre-commit path. Run the new formatter over the existing Nushell scripts and adjust the performance helper's native package helper to use optional parameters instead of empty-string defaults, because nufmt normalizes empty string defaults in a way Nushell cannot parse.\n\n* fix(ccusage): only fall back to comment create on 404 PATCH\n\ntry_update_comment previously collapsed every non-zero gh api PATCH\nexit into a single 'failed' signal, so any transient failure (5xx,\nrate limit, validation) silently triggered a POST and posted a\nduplicate bot comment instead of updating in place.\n\nInspect the gh api stderr instead: treat HTTP 404 as the only case\nthat should fall through to create_comment, mirror the existing\n401/403 skip behavior used by create_comment, and propagate any other\nerror so the workflow surfaces it rather than masking it with a\nduplicate comment.\n\nCo-authored-by: Codesmith \n\n* chore(ci): format upsert comment script\n\n* build: add nushell to dev shell\n\n* fix(ccusage): accurate base runtime description on install fallback\n\nbase_runtime_description was gated on base_package_url alone, but when\na base package URL fails to install in time and --base-dir is provided\nthe script falls back to building from base_dir. In that fallback case\nthe description still claimed the base was running the published\nccusage package from pkg.pr.new even though the local base_dir binary\nwas actually executing.\n\nGate the description on base_package_install so the perf comment only\nclaims the pkg.pr.new install path when an install actually succeeded;\notherwise let render_fixture_section fall back to the local-build\ndescription, matching the head-side fix in 068e481.\n\nCo-authored-by: Codesmith \n\n* build: remove unused fs-fixture dependency\n\n* build(ccusage): keep package scripts lifecycle-only\n\n* build: migrate action scripts to nushell\n\nReplace the remaining Action-facing JavaScript helper scripts with Nushell shebang scripts so CI can invoke them without a workspace JS install. Keep E2E coverage by generating the preview fixture from Nushell and keep package URL materialization in compare-pr-performance on Bun because pkg.pr.new install is the remaining Bun-specific path.\n\nRemove the unused mitata bench script, drop @types/bun and the repo-local Bun API skill, and keep Bun out of the dev shell and GitHub Actions profile installs. Package execution benchmarks now run the built JS launcher with Node while package installation still runs Bun from the script shebang-provided Nix shell.\n\n* fix(ccusage): quote benchmark command text\n\n* build: replace vitest with node test\n\nMove the remaining TypeScript package tests to Node built-in test runner so Vitest and its transitive dependency tree can leave the workspace.\n\nUse oxlint type-aware type checking as the root typecheck entry point and remove tsgo from the dev shell. Keep hermetic Nix formatter/check paths on a lightweight oxlint config because pure Nix checks do not include npm dependencies.\n\nRefresh repo-local skills and agent guidance for Node tests and Nushell scripts.\n\n* fix(ci): convert merged models tests to node assertions\n\norigin/main added models-dev compact tests while this branch removed Vitest. Convert those tests to node:test assertions so the PR merge commit passes oxlint type-aware checking.\n\n* fix(ccusage): run launcher through package bin symlinks\n\nNode test migration replaced the Vitest guard with a direct argv entry comparison. npm and npx execute package bins through node_modules/.bin symlinks, so the wrapper skipped runCli and produced empty output in E2E JSON checks.\n\nCompare real paths for the module URL and argv entry, and cover the symlink case in the launcher tests.\n\n---------\n\nCo-authored-by: Codesmith ","tree":"2ee53e592286285c44919ac0c5bf0ee173a3e7ab"},"mode":"0040000","name":"build-macos-nix-native-package","size":0},{"lastCommit":{"author":{"email":"1560508+ryoppippi@users.noreply.github.com","name":"ryoppippi","when":"2026-07-27T12:16:43.000000+01:00"},"committer":{"email":"noreply@github.com","name":"GitHub","when":"2026-07-27T12:16:43.000000+01:00"},"hash":"86788c3a37809103eeab14a2910adf916ce663c1","message":"perf(rust): split the workspace into independently cached crates (#1428)\n\n* refactor(rust): extract shared core crate\n\nMove shared types, pricing, configuration, output, date, and block logic into ccusage-core so adapter crates can depend on an acyclic foundation. Keep the ccusage package focused on command dispatch and move Codex-owned types into its adapter boundary.\n\nCo-authored-by: ryoppippi \n\n* test(core): move snapshots with extracted modules\n\nKeep Insta snapshot ownership aligned with the new ccusage-core crate so its existing characterization tests continue to compare the same outputs.\n\nCo-authored-by: ryoppippi \n\n* refactor(adapters): extract shared adapter helpers\n\nMove JSONL parsing, file walking, date filtering, size-balanced chunking, and ordered parallel reads into ccusage-adapter-common. This removes the Claude adapter as an infrastructure dependency before agent crates are split.\n\nCo-authored-by: ryoppippi \n\n* refactor(adapters): split agents into workspace crates\n\nMove every agent adapter and the unified all-agent aggregator into dedicated ccusage-adapter-* crates. Preserve their tests and snapshots with each owning crate, keep shared infrastructure in ccusage-adapter-common, and leave ccusage as thin command wiring.\n\nCo-authored-by: ryoppippi \n\n* perf(nix): cache real adapter artifacts\n\nBuild the shared foundation and each adapter as filtered Crane cargoBuild derivations, merge their real target artifacts, and compile only the unified adapter plus thin binary at the final layers. Apply the same target-specific graph to native, musl, and Darwin cross packages.\n\nCo-authored-by: ryoppippi \n\n* fix(core): keep test-only pricing length private\n\nCo-authored-by: ryoppippi \n\n* docs(rust): route adapter work to crates\n\nUpdate repository guidance and source references for the new ccusage-adapter-* workspace layout and shared core/common boundaries.\n\nCo-authored-by: ryoppippi \n\n* fix(nix): use dependency artifacts on Darwin\n\nCrane emits full artifact archives on macOS, so merging sibling adapter archives can overwrite shared crate variants. Keep per-adapter layers on Linux and fall back to the safe dependency-only cache for native and cross Darwin packages.\n\nCo-authored-by: ryoppippi \n\n* refactor(core): own shared adapter reporting\n\nMove generic JSON row and period helpers from the OpenCode adapter into ccusage-core. Remove the resulting OpenCode dependency from leaf adapters and trim obsolete direct dependencies from the binary crate.\n\n* fix(pricing): follow extracted asset paths\n\nRegenerate, format, diff, and stage models.dev and Codex fallback snapshots at their new owning crate paths. Watch the moved config schema source in treefmt as well.\n\n* perf(nix): isolate and retain current adapter artifacts\n\nBuild independent adapters directly from the shared foundation. Root the current Linux native artifact graph and garbage-collect superseded store generations before saving that cache, preventing per-commit artifact accumulation.\n\n* refactor(adapters): keep implementation modules private\n\nExpose only each adapter's intentional command and loader surface while keeping parser, path, report, and implementation modules crate-private.\n\n* fix(core): harden date parsing against overflow and invalid offsets\n\nUse checked arithmetic in parse_digits so multi-digit inputs cannot silently overflow, and reject timezone offsets whose hours exceed 23 or minutes exceed 59.\n\nCo-authored-by: Codesmith \n\n* refactor(rust): centralize adapter reporting and prune unused deps\n\nMove print_table_for_agent into ccusage-adapter-common so goose and codebuff no longer depend on the amp crate, and update the Nix artifact layering accordingly. Remove unused hermes dependencies and drop the bundled sqlite dependency from the JSONL-only copilot, gemini, openclaw, and qwen adapters.\n\nCo-authored-by: Codesmith \n\n* refactor(claude): reuse shared parallel file reader\n\nLoad Claude usage files through ccusage_adapter_common::read_files_parallel and drop the crate-local read_usage_files_parallel duplicate, so improvements to the shared JSONL parallelism apply to Claude too.\n\nCo-authored-by: Codesmith \n\n* chore(rust): drop unused adapter deps and imports\n\nRemove the unused ccusage-terminal dependency from ccusage-adapter-all, prune ccusage-adapter-droid down to the crates it actually uses (dropping bundled sqlite, compact_str, memchr, rustc-hash, smallvec, ccusage-terminal, serde, and the insta dev-dependency), and remove the stale allow(unused_imports) plus dead common-helper imports from the Pi adapter.\n\nCo-authored-by: Codesmith \n\n* fix(report): align agent table totals and weekly labels with JSON\n\nThe shared agent usage table now counts extra_total_tokens by using row.total_tokens() per row and reading totalTokens from the computed totals for the footer, so reasoning and fallback tokens no longer under-report. It also includes row.week in the label fallback chain so --weekly rows keep their period identifiers, matching the JSON report and the core table renderer.\n\nCo-authored-by: Codesmith \n\n* perf(nix): collapse the Crane artifact layers into a linear chain\n\nThe per-adapter fan-out could not be used on Darwin at all, because Crane\ninstalls full rather than incremental archives there\n(rust-lang/rust#115982) and merging sibling full archives can overwrite\nshared path-crate artifacts with incompatible variants. macOS therefore fell\nback to the dependency-only cache and recompiled the whole workspace on\nevery build.\n\nA linear chain only ever extends the previous layer, so it needs no merging\nand works on every platform. The adapters share one layer because they are\nmutually independent: one cargoBuild with a -p per crate lets Cargo compile\nthem concurrently, so the layer costs about as much as its slowest adapter,\nwhile the fan-out added five serial Nix round-trips that each unpack and\nre-archive the dependency target directory.\n\nLinux, macOS arm64, and the cross-compiled macOS x64 build now all reuse the\nsame two layers.\n\n* perf(release): fold identical code in the Linux link\n\nSplitting the runtime into per-adapter crates costs binary size: each crate\ninstantiates the shared generic machinery for its own types and stable Rust\ncannot share those instantiations across crates, so the aarch64-darwin\nbinary grew 2,930,240 -> 3,096,896 bytes, all of it in __text. Linker\ndead-stripping recovers none of it — the duplicates are reachable code, not\ndead code — so mold's identical-code folding is the lever that applies.\n\n* perf(ci): build Nix derivations in parallel\n\nnix-quick-install-action writes the nix.conf it is given verbatim, so Nix\nkept its documented default of max-jobs = 1 and built every derivation one\nat a time — the independent flake checks as well as the artifact layers — on\n32-vCPU runners.\n\ncores stays at its default so the dependency derivation, which compiles the\nwhole third-party graph in a single build, still gets every core.\n\n* perf(nix): scope the clippy and rustfmt checks to the Rust source\n\nccusage-clippy compiles the entire workspace with --all-targets, and its\nsource was the whole repository, so a README, docs page, TypeScript file, or\nworkflow edit invalidated the most expensive check in `nix flake check`.\n\ntests.nix already had the right include list; it now lives in one place so\nthe two cannot drift.\n\n* perf(build): make the LiteLLM pricing download optional\n\nccusage-core's build script reached minreq with https-rustls, which enabled\nrustls's aws-lc-rs provider on top of the ring provider ureq already needs.\nCompiling aws-lc's C and assembly sources took longer than any other unit in\nthe dependency graph, and every workspace crate waited on it, because a build\nscript and its dependencies must be compiled and run before the crate that\nowns them is compiled.\n\nNo Nix build ever used that code path: the sandbox has no network, and every\nNix package hands the snapshot to the build through\nCCUSAGE_PRICING_JSON_PATH. The dev shell now exports the same pinned\nsnapshot, and the Windows release build — the one platform Nix cannot target\n— enables the feature to keep downloading it.\n\n* perf(adapters): finish pruning adapter imports and dependencies\n\nThe manifests and root imports still carried helpers and crates the adapters\nnever reach: sqlite stayed in 6 crates that open no database, insta in\nadapters with no snapshot test, ccusage-terminal in every adapter even though\nthey only see it through ccusage-core's re-exports, and the shared-loader\nimports were still wrapped in a module-wide unused_imports allow that hid\nwhich loaders each adapter depends on.\n\n* docs(rust): note the fetch-litellm-pricing build feature\n\n* perf(cli): split the argument parser out of ccusage-cli\n\nccusage-cli was the root of the dependency graph — ccusage-core and all 15\nadapters depend on it — and it also owned the parser, the help renderer, the\nbuild script, and the 30 KB of embedded help JSON. Touching help text\ntherefore recompiled 20 of 20 workspace units, and every crate waited on a\nbuild script before it could start.\n\nThe adapters and core only ever use the plain argument types, which stay in\nccusage-cli. The new ccusage-cli-parser owns Cli, the parser, the help\nrenderer, and the assets, and only the ccusage binary depends on it, so it\nsits outside the Crane artifact layers entirely: editing cli-help.json now\nleaves the foundation and adapter layers byte-identical (verified by\ncomparing the adapters layer derivation path before and after).\n\nThe config-merge tests move with the parser, since they assert how a parsed\ncommand line and a config file combine.\n\n* docs(rust): give every workspace crate a README\n\nThe split produced 23 crates whose boundaries are only visible in the\nmanifests. Each README now states what the crate owns, what belongs elsewhere,\nits public surface, and which Crane artifact layer it lands in — the last one\nbeing what determines how much a change to it costs. Adapter READMEs also name\nthe data source and environment variables, and link the existing `src/README.md`\nwhere one documents the record shapes.\n\n* fix(adapter-common): treat a zero chunk count as one chunk\n\nchunk_file_indexes_by_size became a cross-crate helper, and with a non-empty\nfile list and a chunk count of zero it indexed an empty vector. One chunk still\nreturns every index, which is what a caller asking for no parallelism wants.\n\nAlso corrects the build-script path in nix/rust-src.nix: the file that reads\nflake.lock is ccusage-core/build.rs.\n\n* docs(rust): derive each crate README's module list from its sources\n\nThe first pass used one template for every adapter, so READMEs listed modules\nthe crate does not have — the Claude adapter has no parser.rs or report.rs, and\nthe Codex adapter's aggregate, replay, speed, and types modules were missing.\nEach list now comes from the crate's own src directory, and the Pi default store\npath includes the sessions directory it actually reads.\n\n* docs(rust): point the skill at the split crate layout\n\nThe skill still described the Rust CLI as two crates. It now lists what each\ncrate owns, points at the per-crate READMEs for the artifact-layer cost, and\nstates the rule the amp/goose/codebuff coupling broke: adapters do not depend on\neach other, shared code goes to ccusage-adapter-common.\n\n* perf(core): move the blocks report into the binary\n\nblocks.rs and its SessionBlock, BurnRate, and Projection types had exactly one\nconsumer, the ccusage binary, but sat in ccusage-core, so 582 lines of\nbilling-block rendering recompiled every adapter whenever they changed. The two\nblock-only thresholds move with them; DEFAULT_SESSION_DURATION_HOURS stays in\ncore because the config layer defaults to it.\n\n* perf(core): extract ccusage-config from core\n\nconfig.rs and config_schema.rs were 28% of ccusage-core's source with exactly one\nconsumer, the binary, and config_schema is where `schemars` and its derive enter\nthe graph. Because core sits in the foundation artifact layer, that meant every\n`ccusage.json` field or schema change recompiled core, adapter-common, and all 15\nadapters, and every build compiled schemars before core could start.\n\n`cargo tree -p ccusage-core` no longer mentions schemars. The config-merge tests\nin ccusage-cli-parser reach the new crate as a dev-dependency.\n\n* refactor(core): let adapters name their own load-progress label\n\nprogress::UsageLoadAgent was a closed enum listing all 16 agents plus a match\nmapping each to a display string, so adding or renaming an agent meant editing\nccusage-core — the crate every adapter depends on and the one in the foundation\nartifact layer. It is now a newtype around the label, and each adapter passes its\nown, keeping the roster with the adapters. Labels are unchanged, including\npi-agent, GitHub Copilot CLI, and Gemini CLI.\n\n* perf(core): inject the pricing HTTP client from the binary\n\nureq entered ccusage-core through one 25-line function, and brought rustls, ring,\nwebpki, and about twenty more crates with it — all of which had to compile before\ncore, and therefore before adapter-common and every adapter. Core now takes the\nfetcher as a function pointer that main installs, and the ureq implementation\nlives with the binary.\n\n`cargo tree -p ccusage-core` no longer mentions ureq, rustls, ring, or webpki.\nWithout a fetcher installed a refresh reports that it is unavailable and the\nembedded snapshots are used, which is the existing --offline behaviour; the\nshipped binary installs one, verified by a live run still taking the network path.\n\n* refactor(rust): group the adapter crates and share manifest settings\n\nEach adapter manifest repeated the same package fields and dependency specs, so\nadding a crate meant copying a block and a version bump meant editing 23 files.\n`[workspace.package]` and `[workspace.dependencies]` now hold both, and each crate\nlists only what it uses — `jiff.workspace = true` rather than the four-line\nfeature spec. It also means moving a crate no longer touches its dependents,\nsince path dependencies live in one place.\n\nThe per-source adapters move to `rust/adapters/`, leaving `rust/crates`\nfor the crates that are not tied to a single source. `ccusage-adapter-common`\nlives with the adapters it serves, which also keeps the `adapters/*` member glob\nmatching in every filtered source tree a Crane layer builds from — a glob that\nmatches nothing is an error, and the foundation layer contains no other adapter.\n\n* perf(ci): stop the schema generator from pulling the adapter graph\n\ngenerate-config-schema was a second bin of the ccusage package, so building it\ncompiled every adapter even though it only calls into the config layer. It moves\nto ccusage-config, where a clean build is 3 units, and the Nix derivation starts\nfrom the foundation artifacts instead of the adapter ones.\n\nThat matters for wall clock because treefmt depends on this generator and the\npreflight job builds treefmt, and preflight gates the whole build matrix by\ndesign — so every build job was waiting on 15 adapters that the formatter never\nneeded.\n\n* perf(release): fold identical code in the macOS link too\n\nMeasured on the shipped Linux binary, mold's identical-code folding recovers\n158,720 of the 240,768 bytes the crate split costs (4,482,112 -> 4,323,392\nagainst 4,241,344 on main). ld64 has no equivalent, so the macOS binaries kept\nthe full cost; lld does, and the Nix build with it produces 3,232,928 bytes\nrather than 3,299,504.\n\nBoth macOS packages still link only /usr/lib and /System/Library dylibs, which\nthe postInstall gate enforces, and `--help` and the JSON reports are byte\nidentical to the ld64 build.\n\n* build(rust): move to 1.97.1 and configure hawk\n\nhawk lints a Cargo workspace for `pub` that the workspace itself never needs,\nwhich is exactly the surface a 24-crate split creates. It uses `rustc_private`\nand must run on the toolchain it was built against, so the pin moves to 1.97.1;\nthe dev shell picks that up from rust-toolchain.toml with no flake change.\n\nhawk.toml names the two production entry points rather than letting hawk assume\nevery binary target is shipped, so the reason each one is an entry point is\nwritten down. 1.97.1's clippy also wants a byte-string literal in\nparse_timezone_offset.\n\n* refactor(rust): narrow the visibility the crate split widened\n\nSplitting one crate into 24 turned a lot of module-private items into `pub`,\nbecause that is what crossing a crate boundary needs — but only some of them are\nactually crossed. hawk's closed-world analysis over the two production entry\npoints found 85 such declarations; this applies the 68 machine-applicable ones,\nturning `pub` into `pub(crate)` and dropping restricted modifiers that no longer\nrestrict anything.\n\n* refactor(rust): narrow the re-exports the crate split widened\n\nThe remaining hawk findings were public re-exports that no other crate reaches:\neach adapter re-exported report_from_rows even though only its own run() calls\nit, claude re-exported three path helpers used only inside itself, and codex\nre-exported two types its own parser and replay modules use. Splitting each\nstatement keeps every name visible where it is used and stops advertising it\npast the crate boundary.\n\nCli::parse had no callers at all — the binary goes through\nparse_from_with_config because it must pass a config context — so it is gone and\nits parse_from shorthand is now test-only. hawk reports no findings.\n\n* docs(rust): write down the workspace visibility rule and how to check it\n\nA 24-crate workspace needs a stated rule for `pub`, or the next crate will export\nwhatever compiles. The rust skill now says what the rule is, how `just hawk`\nchecks it, that hawk must be installed separately because it is not in nixpkgs,\nwhy the toolchain is pinned to 1.97.1, and that its findings are suggestions\nrather than a gate — it is explicitly experimental.\n\n* fix(nix): point the schema formatter at the moved generator\n\nThe schema-gen treefmt formatter still listed config_schema.rs under\nccusage-core and the generator under the ccusage package, so after both moved to\nccusage-config it never ran on the files whose changes should regenerate the\nschema. The config-schema flake check would still have caught the drift, but only\nafter a failing build instead of a formatting pass.\n\n---------\n\nCo-authored-by: Cursor Agent \nCo-authored-by: ryoppippi \nCo-authored-by: Codesmith ","tree":"6ecc56f4b9aff6d73b1c56750b102c17033d8c77"},"mode":"0040000","name":"build-windows-native-package","size":0},{"lastCommit":{"author":{"email":"1560508+ryoppippi@users.noreply.github.com","name":"ryoppippi","when":"2026-07-28T02:05:08.000000+01:00"},"committer":{"email":"noreply@github.com","name":"GitHub","when":"2026-07-28T02:05:08.000000+01:00"},"hash":"b184bdce4dafa3204398a324444fb4f53a34dcf1","message":"ci(pricing): land pricing updates on main instead of opening PRs (#1509)\n\n* ci(pricing): land pricing updates on main instead of opening PRs\n\nThe hourly pricing refresh opened a PR with github.token, which cannot\ntrigger ci.yaml, so the PR arrived with no checks and nothing merged it.\nPricing data was therefore only as fresh as the last time someone merged\nthat PR by hand, even though the workflow already accepts a\nworkflow_dispatch.\n\nBoth jobs now run the checks that a pull request would have run - flake\ncheck, Rust tests and the Node tests, in a parallel step - and push the\nresult straight to main. `just gen-models-dev-pricing` already runs before\nvalidation, so the committed snapshots never lag the locked revision.\n\nThe commit and rebase-onto-main dance is shared by both jobs, so it lives in\na composite action. A concurrency group keeps a manual dispatch from racing\nthe hourly run's push.\n\n* refactor(pricing): move the lock update logic into nushell scripts\n\nThe two update steps were shell pipelines around jq, curl and cmp writing\nJSON to /tmp. Nushell handles the same work as structured data: the lock\nfile is read with `open`, the pricing document is fetched with `http get`\nand compared as a record, so no temporary files or byte-level formatting\nassumptions are involved.\n\nThe scripts carry the repository's Nix shebang, so they need no extra\nnushell setup step in the workflow and stay runnable by hand.\n\nmodels.dev now decides through a `match` over {snapshots, lock}, which\nnames the three states the step can end in - snapshots moved, only the lock\nmoved, nothing moved - instead of inferring them from consecutive `git\ndiff` calls.\n\n* ci(pricing): run the full check set for both pricing sources\n\nThe validation legs only covered flake check, the Rust tests and the Node\ntests, so the Babashka harness test from ci.yaml's test job was missing. Now\nthat the push lands on main without any pull request checks behind it, the\nworkflow has to run what a PR would have run: `nix flake check` already\ncovers the preflight job's treefmt and gitleaks checks, and the remaining\nlegs match the test job leg for leg.\n\nThe two sources were separate jobs with a duplicated leg list, which is how\nthe Babashka leg went missing in the first place. They are a matrix now, so\nthe checks are written once and both sources cannot drift apart.\n`max-parallel: 1` preserves the sequential push the previous `needs:`\nordering gave.","tree":"a334d7f98cccac3a31c264eb6f4b242aabb9419c"},"mode":"0040000","name":"commit-and-push-main","size":0},{"lastCommit":{"author":{"email":"1560508+ryoppippi@users.noreply.github.com","name":"ryoppippi","when":"2026-05-19T23:06:16.000000+01:00"},"committer":{"email":"noreply@github.com","name":"GitHub","when":"2026-05-19T23:06:16.000000+01:00"},"hash":"82c1194580ae9f4d3ac88b5a085ca735daa38edc","message":"ci: skip heavy checks for docs-only changes\n\nSkip heavyweight CI jobs for documentation-only changes by detecting whether a PR or push touched code paths. Keep lightweight spelling checks running, and preserve full CI for workflow or source changes.\\n\\nAlso harden the detection checkout and workflow token permissions based on CodeRabbit review feedback.","tree":"3489219c1e092c24a8097a4bd2323a947fad25d3"},"mode":"0040000","name":"detect-code-changes","size":0},{"lastCommit":{"author":{"email":"1560508+ryoppippi@users.noreply.github.com","name":"ryoppippi","when":"2026-07-20T16:10:40.000000+01:00"},"committer":{"email":"1560508+ryoppippi@users.noreply.github.com","name":"ryoppippi","when":"2026-07-20T16:11:20.000000+01:00"},"hash":"d0e753e28935a33b0b58a475e3187fc6bc5dc434","message":"chore: udpate github actions sha\n","tree":"29b57f30be3f8680fe2f3c7fab9861d116103a33"},"mode":"0040000","name":"setup-native-build-cache","size":0},{"lastCommit":{"author":{"email":"1560508+ryoppippi@users.noreply.github.com","name":"ryoppippi","when":"2026-07-27T12:16:43.000000+01:00"},"committer":{"email":"noreply@github.com","name":"GitHub","when":"2026-07-27T12:16:43.000000+01:00"},"hash":"86788c3a37809103eeab14a2910adf916ce663c1","message":"perf(rust): split the workspace into independently cached crates (#1428)\n\n* refactor(rust): extract shared core crate\n\nMove shared types, pricing, configuration, output, date, and block logic into ccusage-core so adapter crates can depend on an acyclic foundation. Keep the ccusage package focused on command dispatch and move Codex-owned types into its adapter boundary.\n\nCo-authored-by: ryoppippi \n\n* test(core): move snapshots with extracted modules\n\nKeep Insta snapshot ownership aligned with the new ccusage-core crate so its existing characterization tests continue to compare the same outputs.\n\nCo-authored-by: ryoppippi \n\n* refactor(adapters): extract shared adapter helpers\n\nMove JSONL parsing, file walking, date filtering, size-balanced chunking, and ordered parallel reads into ccusage-adapter-common. This removes the Claude adapter as an infrastructure dependency before agent crates are split.\n\nCo-authored-by: ryoppippi \n\n* refactor(adapters): split agents into workspace crates\n\nMove every agent adapter and the unified all-agent aggregator into dedicated ccusage-adapter-* crates. Preserve their tests and snapshots with each owning crate, keep shared infrastructure in ccusage-adapter-common, and leave ccusage as thin command wiring.\n\nCo-authored-by: ryoppippi \n\n* perf(nix): cache real adapter artifacts\n\nBuild the shared foundation and each adapter as filtered Crane cargoBuild derivations, merge their real target artifacts, and compile only the unified adapter plus thin binary at the final layers. Apply the same target-specific graph to native, musl, and Darwin cross packages.\n\nCo-authored-by: ryoppippi \n\n* fix(core): keep test-only pricing length private\n\nCo-authored-by: ryoppippi \n\n* docs(rust): route adapter work to crates\n\nUpdate repository guidance and source references for the new ccusage-adapter-* workspace layout and shared core/common boundaries.\n\nCo-authored-by: ryoppippi \n\n* fix(nix): use dependency artifacts on Darwin\n\nCrane emits full artifact archives on macOS, so merging sibling adapter archives can overwrite shared crate variants. Keep per-adapter layers on Linux and fall back to the safe dependency-only cache for native and cross Darwin packages.\n\nCo-authored-by: ryoppippi \n\n* refactor(core): own shared adapter reporting\n\nMove generic JSON row and period helpers from the OpenCode adapter into ccusage-core. Remove the resulting OpenCode dependency from leaf adapters and trim obsolete direct dependencies from the binary crate.\n\n* fix(pricing): follow extracted asset paths\n\nRegenerate, format, diff, and stage models.dev and Codex fallback snapshots at their new owning crate paths. Watch the moved config schema source in treefmt as well.\n\n* perf(nix): isolate and retain current adapter artifacts\n\nBuild independent adapters directly from the shared foundation. Root the current Linux native artifact graph and garbage-collect superseded store generations before saving that cache, preventing per-commit artifact accumulation.\n\n* refactor(adapters): keep implementation modules private\n\nExpose only each adapter's intentional command and loader surface while keeping parser, path, report, and implementation modules crate-private.\n\n* fix(core): harden date parsing against overflow and invalid offsets\n\nUse checked arithmetic in parse_digits so multi-digit inputs cannot silently overflow, and reject timezone offsets whose hours exceed 23 or minutes exceed 59.\n\nCo-authored-by: Codesmith \n\n* refactor(rust): centralize adapter reporting and prune unused deps\n\nMove print_table_for_agent into ccusage-adapter-common so goose and codebuff no longer depend on the amp crate, and update the Nix artifact layering accordingly. Remove unused hermes dependencies and drop the bundled sqlite dependency from the JSONL-only copilot, gemini, openclaw, and qwen adapters.\n\nCo-authored-by: Codesmith \n\n* refactor(claude): reuse shared parallel file reader\n\nLoad Claude usage files through ccusage_adapter_common::read_files_parallel and drop the crate-local read_usage_files_parallel duplicate, so improvements to the shared JSONL parallelism apply to Claude too.\n\nCo-authored-by: Codesmith \n\n* chore(rust): drop unused adapter deps and imports\n\nRemove the unused ccusage-terminal dependency from ccusage-adapter-all, prune ccusage-adapter-droid down to the crates it actually uses (dropping bundled sqlite, compact_str, memchr, rustc-hash, smallvec, ccusage-terminal, serde, and the insta dev-dependency), and remove the stale allow(unused_imports) plus dead common-helper imports from the Pi adapter.\n\nCo-authored-by: Codesmith \n\n* fix(report): align agent table totals and weekly labels with JSON\n\nThe shared agent usage table now counts extra_total_tokens by using row.total_tokens() per row and reading totalTokens from the computed totals for the footer, so reasoning and fallback tokens no longer under-report. It also includes row.week in the label fallback chain so --weekly rows keep their period identifiers, matching the JSON report and the core table renderer.\n\nCo-authored-by: Codesmith \n\n* perf(nix): collapse the Crane artifact layers into a linear chain\n\nThe per-adapter fan-out could not be used on Darwin at all, because Crane\ninstalls full rather than incremental archives there\n(rust-lang/rust#115982) and merging sibling full archives can overwrite\nshared path-crate artifacts with incompatible variants. macOS therefore fell\nback to the dependency-only cache and recompiled the whole workspace on\nevery build.\n\nA linear chain only ever extends the previous layer, so it needs no merging\nand works on every platform. The adapters share one layer because they are\nmutually independent: one cargoBuild with a -p per crate lets Cargo compile\nthem concurrently, so the layer costs about as much as its slowest adapter,\nwhile the fan-out added five serial Nix round-trips that each unpack and\nre-archive the dependency target directory.\n\nLinux, macOS arm64, and the cross-compiled macOS x64 build now all reuse the\nsame two layers.\n\n* perf(release): fold identical code in the Linux link\n\nSplitting the runtime into per-adapter crates costs binary size: each crate\ninstantiates the shared generic machinery for its own types and stable Rust\ncannot share those instantiations across crates, so the aarch64-darwin\nbinary grew 2,930,240 -> 3,096,896 bytes, all of it in __text. Linker\ndead-stripping recovers none of it — the duplicates are reachable code, not\ndead code — so mold's identical-code folding is the lever that applies.\n\n* perf(ci): build Nix derivations in parallel\n\nnix-quick-install-action writes the nix.conf it is given verbatim, so Nix\nkept its documented default of max-jobs = 1 and built every derivation one\nat a time — the independent flake checks as well as the artifact layers — on\n32-vCPU runners.\n\ncores stays at its default so the dependency derivation, which compiles the\nwhole third-party graph in a single build, still gets every core.\n\n* perf(nix): scope the clippy and rustfmt checks to the Rust source\n\nccusage-clippy compiles the entire workspace with --all-targets, and its\nsource was the whole repository, so a README, docs page, TypeScript file, or\nworkflow edit invalidated the most expensive check in `nix flake check`.\n\ntests.nix already had the right include list; it now lives in one place so\nthe two cannot drift.\n\n* perf(build): make the LiteLLM pricing download optional\n\nccusage-core's build script reached minreq with https-rustls, which enabled\nrustls's aws-lc-rs provider on top of the ring provider ureq already needs.\nCompiling aws-lc's C and assembly sources took longer than any other unit in\nthe dependency graph, and every workspace crate waited on it, because a build\nscript and its dependencies must be compiled and run before the crate that\nowns them is compiled.\n\nNo Nix build ever used that code path: the sandbox has no network, and every\nNix package hands the snapshot to the build through\nCCUSAGE_PRICING_JSON_PATH. The dev shell now exports the same pinned\nsnapshot, and the Windows release build — the one platform Nix cannot target\n— enables the feature to keep downloading it.\n\n* perf(adapters): finish pruning adapter imports and dependencies\n\nThe manifests and root imports still carried helpers and crates the adapters\nnever reach: sqlite stayed in 6 crates that open no database, insta in\nadapters with no snapshot test, ccusage-terminal in every adapter even though\nthey only see it through ccusage-core's re-exports, and the shared-loader\nimports were still wrapped in a module-wide unused_imports allow that hid\nwhich loaders each adapter depends on.\n\n* docs(rust): note the fetch-litellm-pricing build feature\n\n* perf(cli): split the argument parser out of ccusage-cli\n\nccusage-cli was the root of the dependency graph — ccusage-core and all 15\nadapters depend on it — and it also owned the parser, the help renderer, the\nbuild script, and the 30 KB of embedded help JSON. Touching help text\ntherefore recompiled 20 of 20 workspace units, and every crate waited on a\nbuild script before it could start.\n\nThe adapters and core only ever use the plain argument types, which stay in\nccusage-cli. The new ccusage-cli-parser owns Cli, the parser, the help\nrenderer, and the assets, and only the ccusage binary depends on it, so it\nsits outside the Crane artifact layers entirely: editing cli-help.json now\nleaves the foundation and adapter layers byte-identical (verified by\ncomparing the adapters layer derivation path before and after).\n\nThe config-merge tests move with the parser, since they assert how a parsed\ncommand line and a config file combine.\n\n* docs(rust): give every workspace crate a README\n\nThe split produced 23 crates whose boundaries are only visible in the\nmanifests. Each README now states what the crate owns, what belongs elsewhere,\nits public surface, and which Crane artifact layer it lands in — the last one\nbeing what determines how much a change to it costs. Adapter READMEs also name\nthe data source and environment variables, and link the existing `src/README.md`\nwhere one documents the record shapes.\n\n* fix(adapter-common): treat a zero chunk count as one chunk\n\nchunk_file_indexes_by_size became a cross-crate helper, and with a non-empty\nfile list and a chunk count of zero it indexed an empty vector. One chunk still\nreturns every index, which is what a caller asking for no parallelism wants.\n\nAlso corrects the build-script path in nix/rust-src.nix: the file that reads\nflake.lock is ccusage-core/build.rs.\n\n* docs(rust): derive each crate README's module list from its sources\n\nThe first pass used one template for every adapter, so READMEs listed modules\nthe crate does not have — the Claude adapter has no parser.rs or report.rs, and\nthe Codex adapter's aggregate, replay, speed, and types modules were missing.\nEach list now comes from the crate's own src directory, and the Pi default store\npath includes the sessions directory it actually reads.\n\n* docs(rust): point the skill at the split crate layout\n\nThe skill still described the Rust CLI as two crates. It now lists what each\ncrate owns, points at the per-crate READMEs for the artifact-layer cost, and\nstates the rule the amp/goose/codebuff coupling broke: adapters do not depend on\neach other, shared code goes to ccusage-adapter-common.\n\n* perf(core): move the blocks report into the binary\n\nblocks.rs and its SessionBlock, BurnRate, and Projection types had exactly one\nconsumer, the ccusage binary, but sat in ccusage-core, so 582 lines of\nbilling-block rendering recompiled every adapter whenever they changed. The two\nblock-only thresholds move with them; DEFAULT_SESSION_DURATION_HOURS stays in\ncore because the config layer defaults to it.\n\n* perf(core): extract ccusage-config from core\n\nconfig.rs and config_schema.rs were 28% of ccusage-core's source with exactly one\nconsumer, the binary, and config_schema is where `schemars` and its derive enter\nthe graph. Because core sits in the foundation artifact layer, that meant every\n`ccusage.json` field or schema change recompiled core, adapter-common, and all 15\nadapters, and every build compiled schemars before core could start.\n\n`cargo tree -p ccusage-core` no longer mentions schemars. The config-merge tests\nin ccusage-cli-parser reach the new crate as a dev-dependency.\n\n* refactor(core): let adapters name their own load-progress label\n\nprogress::UsageLoadAgent was a closed enum listing all 16 agents plus a match\nmapping each to a display string, so adding or renaming an agent meant editing\nccusage-core — the crate every adapter depends on and the one in the foundation\nartifact layer. It is now a newtype around the label, and each adapter passes its\nown, keeping the roster with the adapters. Labels are unchanged, including\npi-agent, GitHub Copilot CLI, and Gemini CLI.\n\n* perf(core): inject the pricing HTTP client from the binary\n\nureq entered ccusage-core through one 25-line function, and brought rustls, ring,\nwebpki, and about twenty more crates with it — all of which had to compile before\ncore, and therefore before adapter-common and every adapter. Core now takes the\nfetcher as a function pointer that main installs, and the ureq implementation\nlives with the binary.\n\n`cargo tree -p ccusage-core` no longer mentions ureq, rustls, ring, or webpki.\nWithout a fetcher installed a refresh reports that it is unavailable and the\nembedded snapshots are used, which is the existing --offline behaviour; the\nshipped binary installs one, verified by a live run still taking the network path.\n\n* refactor(rust): group the adapter crates and share manifest settings\n\nEach adapter manifest repeated the same package fields and dependency specs, so\nadding a crate meant copying a block and a version bump meant editing 23 files.\n`[workspace.package]` and `[workspace.dependencies]` now hold both, and each crate\nlists only what it uses — `jiff.workspace = true` rather than the four-line\nfeature spec. It also means moving a crate no longer touches its dependents,\nsince path dependencies live in one place.\n\nThe per-source adapters move to `rust/adapters/`, leaving `rust/crates`\nfor the crates that are not tied to a single source. `ccusage-adapter-common`\nlives with the adapters it serves, which also keeps the `adapters/*` member glob\nmatching in every filtered source tree a Crane layer builds from — a glob that\nmatches nothing is an error, and the foundation layer contains no other adapter.\n\n* perf(ci): stop the schema generator from pulling the adapter graph\n\ngenerate-config-schema was a second bin of the ccusage package, so building it\ncompiled every adapter even though it only calls into the config layer. It moves\nto ccusage-config, where a clean build is 3 units, and the Nix derivation starts\nfrom the foundation artifacts instead of the adapter ones.\n\nThat matters for wall clock because treefmt depends on this generator and the\npreflight job builds treefmt, and preflight gates the whole build matrix by\ndesign — so every build job was waiting on 15 adapters that the formatter never\nneeded.\n\n* perf(release): fold identical code in the macOS link too\n\nMeasured on the shipped Linux binary, mold's identical-code folding recovers\n158,720 of the 240,768 bytes the crate split costs (4,482,112 -> 4,323,392\nagainst 4,241,344 on main). ld64 has no equivalent, so the macOS binaries kept\nthe full cost; lld does, and the Nix build with it produces 3,232,928 bytes\nrather than 3,299,504.\n\nBoth macOS packages still link only /usr/lib and /System/Library dylibs, which\nthe postInstall gate enforces, and `--help` and the JSON reports are byte\nidentical to the ld64 build.\n\n* build(rust): move to 1.97.1 and configure hawk\n\nhawk lints a Cargo workspace for `pub` that the workspace itself never needs,\nwhich is exactly the surface a 24-crate split creates. It uses `rustc_private`\nand must run on the toolchain it was built against, so the pin moves to 1.97.1;\nthe dev shell picks that up from rust-toolchain.toml with no flake change.\n\nhawk.toml names the two production entry points rather than letting hawk assume\nevery binary target is shipped, so the reason each one is an entry point is\nwritten down. 1.97.1's clippy also wants a byte-string literal in\nparse_timezone_offset.\n\n* refactor(rust): narrow the visibility the crate split widened\n\nSplitting one crate into 24 turned a lot of module-private items into `pub`,\nbecause that is what crossing a crate boundary needs — but only some of them are\nactually crossed. hawk's closed-world analysis over the two production entry\npoints found 85 such declarations; this applies the 68 machine-applicable ones,\nturning `pub` into `pub(crate)` and dropping restricted modifiers that no longer\nrestrict anything.\n\n* refactor(rust): narrow the re-exports the crate split widened\n\nThe remaining hawk findings were public re-exports that no other crate reaches:\neach adapter re-exported report_from_rows even though only its own run() calls\nit, claude re-exported three path helpers used only inside itself, and codex\nre-exported two types its own parser and replay modules use. Splitting each\nstatement keeps every name visible where it is used and stops advertising it\npast the crate boundary.\n\nCli::parse had no callers at all — the binary goes through\nparse_from_with_config because it must pass a config context — so it is gone and\nits parse_from shorthand is now test-only. hawk reports no findings.\n\n* docs(rust): write down the workspace visibility rule and how to check it\n\nA 24-crate workspace needs a stated rule for `pub`, or the next crate will export\nwhatever compiles. The rust skill now says what the rule is, how `just hawk`\nchecks it, that hawk must be installed separately because it is not in nixpkgs,\nwhy the toolchain is pinned to 1.97.1, and that its findings are suggestions\nrather than a gate — it is explicitly experimental.\n\n* fix(nix): point the schema formatter at the moved generator\n\nThe schema-gen treefmt formatter still listed config_schema.rs under\nccusage-core and the generator under the ccusage package, so after both moved to\nccusage-config it never ran on the files whose changes should regenerate the\nschema. The config-schema flake check would still have caught the drift, but only\nafter a failing build instead of a formatting pass.\n\n---------\n\nCo-authored-by: Cursor Agent \nCo-authored-by: ryoppippi \nCo-authored-by: Codesmith ","tree":"6ecc56f4b9aff6d73b1c56750b102c17033d8c77"},"mode":"0040000","name":"setup-nix-cache","size":0}],"lastCommit":{"author":{"email":"1560508+ryoppippi@users.noreply.github.com","name":"ryoppippi","when":"2026-07-28T02:05:08.000000+01:00"},"committer":{"email":"noreply@github.com","name":"GitHub","when":"2026-07-28T02:05:08.000000+01:00"},"hash":"b184bdce4dafa3204398a324444fb4f53a34dcf1","message":"ci(pricing): land pricing updates on main instead of opening PRs (#1509)\n\n* ci(pricing): land pricing updates on main instead of opening PRs\n\nThe hourly pricing refresh opened a PR with github.token, which cannot\ntrigger ci.yaml, so the PR arrived with no checks and nothing merged it.\nPricing data was therefore only as fresh as the last time someone merged\nthat PR by hand, even though the workflow already accepts a\nworkflow_dispatch.\n\nBoth jobs now run the checks that a pull request would have run - flake\ncheck, Rust tests and the Node tests, in a parallel step - and push the\nresult straight to main. `just gen-models-dev-pricing` already runs before\nvalidation, so the committed snapshots never lag the locked revision.\n\nThe commit and rebase-onto-main dance is shared by both jobs, so it lives in\na composite action. A concurrency group keeps a manual dispatch from racing\nthe hourly run's push.\n\n* refactor(pricing): move the lock update logic into nushell scripts\n\nThe two update steps were shell pipelines around jq, curl and cmp writing\nJSON to /tmp. Nushell handles the same work as structured data: the lock\nfile is read with `open`, the pricing document is fetched with `http get`\nand compared as a record, so no temporary files or byte-level formatting\nassumptions are involved.\n\nThe scripts carry the repository's Nix shebang, so they need no extra\nnushell setup step in the workflow and stay runnable by hand.\n\nmodels.dev now decides through a `match` over {snapshots, lock}, which\nnames the three states the step can end in - snapshots moved, only the lock\nmoved, nothing moved - instead of inferring them from consecutive `git\ndiff` calls.\n\n* ci(pricing): run the full check set for both pricing sources\n\nThe validation legs only covered flake check, the Rust tests and the Node\ntests, so the Babashka harness test from ci.yaml's test job was missing. Now\nthat the push lands on main without any pull request checks behind it, the\nworkflow has to run what a PR would have run: `nix flake check` already\ncovers the preflight job's treefmt and gitleaks checks, and the remaining\nlegs match the test job leg for leg.\n\nThe two sources were separate jobs with a duplicated leg list, which is how\nthe Babashka leg went missing in the first place. They are a matrix now, so\nthe checks are written once and both sources cannot drift apart.\n`max-parallel: 1` preserves the sequential push the previous `needs:`\nordering gave.","tree":"a334d7f98cccac3a31c264eb6f4b242aabb9419c"}}