internal/

Package a2a models agent-to-agent handoffs with causal ordering.

Package budget tracks per-model context window utilisation and enforces the global 70%/75%/80% budget rules defined in CLAUDE.md.

Package build exposes version metadata injected at build time via -ldflags.

Package capabilities provides model scoring data, embedded at compile time and extensible at runtime via a user overlay (~/.hydra/models.json), so users can add new models (e.g.

Package config defines Hydra's runtime configuration: Cortex, Heads, policies, skills.

Package cost reads ~/.hydra/logs/cost.jsonl and produces spend summaries.

Package diff produces unified diffs in-process.

Package dispatch routes a prompt through policy evaluation, head selection, and execution — with automatic fallback on failure.

Package editor provides atomic, validated, rollback-safe file editing via Hydra Heads.

Package entropy measures the useful information in a context window, not just its length.

Package executor runs prompts against Hydra Heads.

Package formats holds no code.

Package glob implements the single path-glob dialect Hydra uses everywhere: workspace allow/deny rules and MCP ledger resource rules.

Package graph reads a code dependency graph (graph.json, as produced by Graphify or any tree-sitter indexer) and computes a file's blast radius — how much other code transitively depends on it.

Package ledger is Hydra's local MCP accountability ledger: an append-only record of what every agent was allowed to touch and did.

Package optimal computes the provably-best number of parallel agents for a task, per Manifesto Law 4.

Package oracle turns deterministic verifiers — test runners, compilers, linters — into first-class evidence sources for the Trust Control Plane.

Package parallel fans N independent tasks out to N Hydra Heads simultaneously.

Package policy evaluates routing rules before a prompt is dispatched.

Package pricing provides live model cost data fetched from OpenRouter, cached locally with a 24-hour TTL, falling back to the static registry/pricing.yaml when offline or on first run.

Package probe orchestrates all registered providers and returns a ranked list of available Heads with a recommended Cortex.

Package provider defines the interfaces and types every Hydra Head must implement.

Package rank provides ordering and deduplication for discovered Heads.

Package review provides the review surface for files edited by Hydra.

Package runid mints the identifiers that correlate a single logical unit of work across Hydra's logs.

Package runlog is Hydra's append-only record of what happened during a run.

Package security aggregates Hydra's security-relevant data — the ledger's accountability trail, per-head risk, and a short list of honest checks — into one report.

Package swarm implements fan-out dispatch: one prompt fired at multiple Heads simultaneously, with race / best / all response strategies.

Package sysinfo detects hardware specs relevant to AI model selection.

Package testutil holds test fixtures shared across packages.

Package tree turns a run's event log into the two shapes a UI renders: a supervision tree and a chronological timeline.

Package trust is the confidence layer of the Trust Control Plane: it measures how much each evidence source's verdict is actually worth (calibration) and how costly a wrong answer is for a given task (defect cost).

Package tui contains the Hydra init wizard built with Bubbletea.

Package update checks for newer Hydra releases and notifies the user.

Package workspace validates file paths against the workspace registry (registry/workspace.yaml) and provides git root detection and per-extension validator lookup.