Tacit Adds Recurring Pull Request Review Memory to Claude Code Agents

The Limit of Monolithic Rule Files

Storing developer preferences in a single CLAUDE.md file inevitably fails over extended agentic sessions. When context windows fill with file reads and shell outputs, early guidance becomes background noise. Adding dozens of retrospective rules turns project instructions into an unread checklist, while repetitive prompts cause the model to disregard guidance entirely.

The Tacit Plugin Architecture

The open-source plugin tacit addresses this failure mode by splitting tacit knowledge into two distinct operational stores:

1. Global Rules: Broad architectural heuristics stored with recurrence counters in the format [hits: N, last: YYYY-MM]. 2. Local Checks: Stack-specific command-line invocations stored in files like .claude/papercuts/sql.md with file-path frontmatter, loaded only when matching code paths are written.

When a rule reaches two or more hits (hits >= 2), Tacit flags it for escalation into a deterministic linter or automated verification script.

Self-Auditing and Safety Guardrails

To prevent language models from falsely reporting zero mistakes, Tacit records every captured issue in papercuts-log.jsonl and audits counts with tacit audit. Crucially, hook scripts are configured never to return automated permissions (permissionDecision: allow) to avoid bypassing user confirmations during Git pushes, and all internal helper scripts fail silently without aborting parent workflows.