internal/

directory
v0.8.1 Latest Latest
Warning

This package is not in the latest version of its module.

Go to latest
Published: Aug 26, 2026 License: MIT

Directories

Path Synopsis
Package cli carries the shared state every evolve subcommand consumes: the global flags, the layered .evolve configuration (YAML, JSON, JSONC, or TOML — see ConfigExtensions), and the helpers that resolve them into a repository, provider set, and token counter.
Package cli carries the shared state every evolve subcommand consumes: the global flags, the layered .evolve configuration (YAML, JSON, JSONC, or TOML — see ConfigExtensions), and the helpers that resolve them into a repository, provider set, and token counter.
Package configdoc renders the configuration documentation: a Markdown reference describing every option with its default, and annotated example config files (.evolve.yaml, .evolve.jsonc, .evolve.toml) with every default value set.
Package configdoc renders the configuration documentation: a Markdown reference describing every option with its default, and annotated example config files (.evolve.yaml, .evolve.jsonc, .evolve.toml) with every default value set.
Package encfmt reads and writes the supported encoding formats — JSON, JSONC, and YAML — behind one JSON data model.
Package encfmt reads and writes the supported encoding formats — JSON, JSONC, and YAML — behind one JSON data model.
Package evalspec parses the authored eval definitions: triggers and evals files in skill-creator's envelope shape ({skill_name?, evals|triggers: [...]}).
Package evalspec parses the authored eval definitions: triggers and evals files in skill-creator's envelope shape ({skill_name?, evals|triggers: [...]}).
Package grade evaluates case assertions against an agent's workspace and final output: deterministic checks (files, regexes, commands) first, then one LLM-judge session grading all of the case's subjective (llm) assertions at once — the judge runs with the full eval tool posture, so evidence is collected before it can touch the workspace.
Package grade evaluates case assertions against an agent's workspace and final output: deterministic checks (files, regexes, commands) first, then one LLM-judge session grading all of the case's subjective (llm) assertions at once — the judge runs with the full eval tool posture, so evidence is collected before it can touch the workspace.
Package harness defines the agent CLIs evolve can drive (Claude Code, OpenAI Codex, Gemini, Cursor, Copilot, Antigravity, Grok): their runner-binary candidates, credential env vars, skill directories, command construction, and output parsing.
Package harness defines the agent CLIs evolve can drive (Claude Code, OpenAI Codex, Gemini, Cursor, Copilot, Antigravity, Grok): their runner-binary candidates, credential env vars, skill directories, command construction, and output parsing.
Package layout detects which of the three supported repository shapes a root directory has and enumerates its plugins and eval definitions.
Package layout detects which of the three supported repository shapes a root directory has and enumerates its plugins and eval definitions.
Package manifest parses the artifacts a plugin repository ships: plugin and marketplace manifests (as generic JSON, mirroring the loose reads the checks perform) and SKILL.md frontmatter.
Package manifest parses the artifacts a plugin repository ships: plugin and marketplace manifests (as generic JSON, mirroring the loose reads the checks perform) and SKILL.md frontmatter.
Package model defines the model vendors (Anthropic, OpenAI, Google, Cursor, xAI) and their canonical models: provider-qualified ids, pricing, the harnesses each model can be driven by, and the vendor token-counting clients.
Package model defines the model vendors (Anthropic, OpenAI, Google, Cursor, xAI) and their canonical models: provider-qualified ids, pricing, the harnesses each model can be driven by, and the vendor token-counting clients.
Package plan is the single owner of what a sweep runs, for which models, and in what order.
Package plan is the single owner of what a sweep runs, for which models, and in what order.
Package profile turns the hidden --profile flag into pprof output.
Package profile turns the hidden --profile flag into pprof output.
Package remote is evolve's client for patchy's remote-evaluation service: the OIDC login flow and credential store, the HTTP/SSE client, the deterministic workspace bundler, the remote sweep orchestrator, and the bidirectional Reporter seam — EventReporter serializes run.Reporter calls onto the in-pod EVOLVE-EVENT stdout stream, ApplyEvent replays received events back onto a local Reporter, so a remote run's output is indistinguishable from a local one.
Package remote is evolve's client for patchy's remote-evaluation service: the OIDC login flow and credential store, the HTTP/SSE client, the deterministic workspace bundler, the remote sweep orchestrator, and the bidirectional Reporter seam — EventReporter serializes run.Reporter calls onto the in-pod EVOLVE-EVENT stdout stream, ApplyEvent replays received events back onto a local Reporter, so a remote run's output is indistinguishable from a local one.
Package report renders the committed results.json files into EVALUATION.md (root rollup, plus per-plugin detail pages in marketplace/multi layouts) and a machine-readable EVALUATION.json.
Package report renders the committed results.json files into EVALUATION.md (root rollup, plus per-plugin detail pages in marketplace/multi layouts) and a machine-readable EVALUATION.json.
Package results owns the committed results files that live beside each skill's eval definitions (evals/<skill>/results.<ext>, format selected by results_format).
Package results owns the committed results files that live beside each skill's eval definitions (evals/<skill>/results.<ext>, format selected by results_format).
Package run implements the three eval engines behind `evolve run`: static checks (Tier 0), trigger accuracy (Tier 1), and behavioral evals (Tier 2).
Package run implements the three eval engines behind `evolve run`: static checks (Tier 0), trigger accuracy (Tier 1), and behavioral evals (Tier 2).
Package runner executes model.CommandSpecs.
Package runner executes model.CommandSpecs.
Package telemetry wires OpenTelemetry tracing, metrics, and structured logging into evolve and owns their lifecycle.
Package telemetry wires OpenTelemetry tracing, metrics, and structured logging into evolve and owns their lifecycle.
Package tokencount caches provider-reported input-token counts.
Package tokencount caches provider-reported input-token counts.
Package tui renders the interactive selection form and live run dashboard for `evolve run`, plus the fuzzy multi-select picker for `evolve models discover` (see discover.go).
Package tui renders the interactive selection form and live run dashboard for `evolve run`, plus the fuzzy multi-select picker for `evolve models discover` (see discover.go).
Package version exposes build metadata stamped in via -ldflags.
Package version exposes build metadata stamped in via -ldflags.
Package web serves the interactive report viewer: a localhost HTTP server hosting an embedded single-page app (built from internal/web/ui) over a read-only JSON API.
Package web serves the interactive report viewer: a localhost HTTP server hosting an embedded single-page app (built from internal/web/ui) over a read-only JSON API.
Package workspace builds the throwaway project directories agent sessions run in: every skill of the plugin is symlinked into each requested skills dir (so the skill under test has to win against its siblings), case fixture files are written at their relative paths, and the result is committed as the initial state of a fresh git repository — real projects are repos, and both agents and repo-root-detecting tools (actionlint, pre-commit) rely on that.
Package workspace builds the throwaway project directories agent sessions run in: every skill of the plugin is symlinked into each requested skills dir (so the skill under test has to win against its siblings), case fixture files are written at their relative paths, and the result is committed as the initial state of a fresh git repository — real projects are repos, and both agents and repo-root-detecting tools (actionlint, pre-commit) rely on that.

Jump to

Keyboard shortcuts

? : This menu
/ : Search site
f or F : Jump to
y or Y : Canonical URL