Track read:edit ratio per agent from session JSONL logs

[moodmosaic]

Summary

Add a new driver interface function agent_tool_counts that parses
session JSONL and counts research vs mutation tool calls. Append the
counts to the existing stats TSV and display the ratio in both
costs.sh and dashboard.sh.

Design

New driver interface function: agent_tool_counts

Each driver parses its own JSONL format and outputs reads\tedits
(tab-separated).

Driver	Reads (research)	Edits (mutation)
Claude Code	Read, Grep, Glob	Write, Edit, MultiEdit
Codex CLI	command_execution, web_search, mcp_tool_call	file_change
Gemini CLI	read_file, grep_search, list_directory	write_file, edit_file

Shell/Bash is excluded from both categories (ambiguous -- could be
research or mutation).

Extend the stats TSV

Append two columns to the existing 9-column stats_agent_N.tsv:

... | turns | reads | edits

Backward compatible: awk treats missing columns as 0 for older rows.

Display in costs.sh and dashboard.sh

Both scripts already read the TSV via read_agent_stats. Add a R:E
column showing the ratio, e.g.:

#    Model                     Cost      Input     Output  Cache    Turns   R:E      Time
1    claude-opus-4-6 (h)    $0.4200     120k       45k    890k       12   6.2      1h 23m
2    claude-opus-4-6 (h)    $0.3800     105k       38k    820k       10   2.1      1h 05m

Format: ratio as %.1f when edits > 0, - when both are 0.

References

• anthropics/claude-code#42796 -- Claude Code quality regression report
• Boris Cherny's comment documenting read:edit ratio drop from 6.6 to 2.0

Motivation

A recent report on Claude Code quality regression
(anthropics/claude-code#42796)
and Boris Cherny's analysis
showed that a key leading indicator of degraded code quality is the
read:edit ratio -- the number of research tool calls (Read, Grep,
Glob) divided by mutation tool calls (Write, Edit).

In their measurements, the ratio dropped from 6.6 to 2.0, meaning
the agent shifted from carefully reading code before editing to
"edit-happy" behavior -- mutating code without sufficient context
gathering. This correlated directly with worse output quality.

The swarm already records every tool call in its session JSONL logs
(agent_logs/agent_<id>_<commit>_<ts>.log). We should compute and
surface this metric so operators can detect degradation per agent,
per session, and over time.