Module context_usage 

Harness transcript token and context-window policy.

This module owns the pure half of pre-emptive context clearing: parse token usage from harness transcript content, map model names to context windows, compute context percentage, and decide whether an opted-in caller should request a destructive context clear. Filesystem transcript discovery and reads stay in orchestration adapters.

Structs

ClearDecision

Outcome of the dispatch-time pre-emptive clear gate.

TranscriptContextPct

Pure transcript context-% policy result.

UsedTokens

Cumulative token usage read from a harness session transcript.

Enums

Harness

Harness whose session transcript token usage can be interpreted.

TranscriptContextPctDiagnostic

Diagnostic reason for a transcript context-% miss.

Constants

CLAUDE_CONTEXT_WINDOW

Default context window in tokens for the current Claude model family.

Functions

claude_project_hash

Claude Code encodes a project directory into its transcript project subdirectory by replacing every / and . with -.

claude_projects_subdir

Compose the ~/.claude/projects/<project-hash>/ transcript directory.

claude_transcript_path

Locate a Claude Code session transcript by known session id.

clear_decision

Decide whether an opted-in caller should pre-emptively clear context before dispatching a queue head.

context_pct

Compute context-usage percentage for used tokens against model’s window, clamped to [0, 100].

context_window_for_model

Map a resolved model id to its context window in tokens.

parse_claude_jsonl_used_tokens

Parse cumulative token usage from Claude Code JSONL transcript content.

parse_codex_jsonl_context_pct

Parse Codex TUI session JSONL and compute context usage from the latest token_count event.

parse_codex_jsonl_session_meta_cwd

Parse the payload.cwd from a Codex TUI session_meta record.

transcript_context_pct_from_content

Parse transcript content and compute context-usage percentage for a harness.