v1.11.0

writ token-audit now bills each API response once at its own model's rates and includes subagent spend, and the sub-agent roles carry an explicit model, effort and status contract that the orchestrator routes on. The old audit counted most responses about twice and priced every model at one rate; subagent transcripts were not counted at all. Writ no longer runs headless Claude Code anywhere.

Changed

• token-audit bills per API response. Claude Code writes one transcript record per content block, so the audit summed each response about twice. Records sharing a message.id now count once; streaming snapshots (only output_tokens grows) collapse silently, the last record wins, and any other difference raises a duplicate_id_conflict warning.
• Real per-model prices. A RATE_CARD holds input, output, cache read and 5-minute and 1-hour cache write rates for each model, including Opus 5.5's 0.05x cache read. Model ids match exactly (a trailing date suffix is stripped); an unknown model is reported as unpriced and the dollar figure is labelled partial. --model is a fallback for records that carry no model and defaults to none.
• Subagent spend is included. The audit links each Agent dispatch to its transcript and reports cost by model, role and dispatch, dispatch coverage, missing and orphan transcripts, and a reconciliation against Claude Code's own cost-state total. Treat the audit as a floor: transcripts omit some billed calls Claude Code makes.
• Subagent usage survives a deleted transcript. The SubagentStop hook appends a token-only subagent_usage row to the metrics stream; the audit prices a dispatch from it when the transcript is gone. writ token-audit gains --project.
• Per-role effort. SubagentRole.effort_preference is exported as effort: frontmatter: planner and implementer high; explorer, test writer and reviewer medium, matching what each ran at before. HANDBOOK section 8 documents the role table.
• Status contracts for every worker. The explorer, test writer and planner end with one status (COMPLETE, INSUFFICIENT_CONTEXT, a REQUIRES_* decision, or CONFLICTING_EVIDENCE); the reviewer's JSON status gains insufficient_context and conflicting_evidence, and those two block a commit even with an empty critical list, in any letter case.
• Dispatch policy. PBK-PROC-ORCHESTRATOR-001 says when to work directly instead of delegating, to retrieve missing facts before reaching for a stronger model, when to override a dispatch's model (haiku for bounded lookups only; fable only on the user's request), and how to route each returned status.
• The fix loop escalates by evidence. Rounds 4-5 dispatch a fresh implementer on the same model with the plan, failing test output, every open finding and prior fix reports, instead of an unspecified "more capable model".
• writ/analysis loads pydantic lazily, so the stop hook imports the token audit without it.

Removed

• The efficacy-ab harness (writ/analysis/efficacy_ab.py, variants.py, the command, its task suite and runbook). It was Writ's only headless claude -p use; tests/test_no_headless_claude.py now fails on any new one.