[dimittal]

Two bugs in automatic (hook) capture, found while running a real Claude Code (Opus 5.5) session against a local Hindsight.

1. Every hook capture after the first was dropped

Without an extraction model, hook captures all send the placeholder summary "Captured agent evidence; requires human review." The duplicate fingerprint covered only the summary, so the server returned extraction_state: duplicate for every capture after the first in a scope. Automatic capture effectively saved one draft, ever.

Fix: hook payloads set evidence_only: true; those drafts are fingerprinted on summary + redacted excerpt. Explicitly proposed facts (MCP propose_clue, CLI) still dedupe on the fact alone, so rejection suppression is unchanged.

2. Harness-injected context leaked into clue evidence

Current Claude Code transcripts contain api-request-blob records that repeat each user turn together with injected <system-reminder> context (account email, workspace details). The hook read them as user messages, so drafts contained that context (the email was redacted, but other workspace details need not be) and each prompt appeared twice.

Fix: the Claude reader keeps only user/assistant records, skips isMeta, and strips <system-reminder> blocks. The server also strips them from captured summaries and excerpts, since any client can send them.

Tests

• New: test_transcript_drafts_dedupe_on_evidence_and_drop_injected_context (integration) and test_claude_capture_skips_request_copies_and_injected_context (agent contract). Both fail on main and pass here.
• Updated the incremental-checkpoint fixture to use real Claude record shape (type: "assistant").
• pytest (87), ruff check, ruff format --check, mypy pass.
• Manual: image built from this branch, Stop hook active, two real Opus 5.5 sessions → two distinct drafts, no injected context, each prompt once.

🤖 Generated with Claude Code