Summary
Hat tip to @nethoxa for sharing the Asymmetric Research blog post. 💯
Currently we partition work before agents run — role specialization, lock files, KNOWN_INDEX.md, and SCAN_MODE prevent duplicate effort upfront. But we have no mechanism to verify what agents actually read after a swarm completes. An agent can claim an area via a lock file but silently skip large parts of it.
Inspired by Asymmetric Research's agent-coverage tool which parses session logs to extract per-line read coverage and visualize gaps.
Proposal
Add a post_process step (or standalone script) that:
- Parses all agents' JSONL session logs from
agent_logs/ — extract every Read tool call (file path + line range) and Shell commands that read files (e.g. cat, head, rg)
- Builds a coverage map — which files/line-ranges were read by which agent
- Diffs against the source tree (or a user-supplied target file list) to identify uncovered files and functions
- Emits a coverage report — summary of covered vs. uncovered areas, per-agent breakdown
- (Optional) Generates a follow-up prompt for a cleanup agent pointing at the gaps
Why this matters
- Agents are probabilistic — the blog's data shows repeated runs with identical config produce meaningfully different coverage
- Upfront partitioning (our current approach) and post-hoc gap analysis (their approach) are complementary
- This closes the loop: partition before, verify after, re-prompt for what was missed
References
Motivation
We already partition work before agents run (role specialization, lock files, KNOWN_INDEX). But there's no way to verify what agents actually read after a swarm completes. An agent can claim an area and silently skip large parts of it. Post-hoc coverage analysis would close that gap — partition before, verify after, re-prompt for what was missed.
Summary
Hat tip to @nethoxa for sharing the Asymmetric Research blog post. 💯
Currently we partition work before agents run — role specialization, lock files,
KNOWN_INDEX.md, andSCAN_MODEprevent duplicate effort upfront. But we have no mechanism to verify what agents actually read after a swarm completes. An agent can claim an area via a lock file but silently skip large parts of it.Inspired by Asymmetric Research's agent-coverage tool which parses session logs to extract per-line read coverage and visualize gaps.
Proposal
Add a
post_processstep (or standalone script) that:agent_logs/— extract everyReadtool call (file path + line range) andShellcommands that read files (e.g.cat,head,rg)Why this matters
References
Motivation
We already partition work before agents run (role specialization, lock files, KNOWN_INDEX). But there's no way to verify what agents actually read after a swarm completes. An agent can claim an area and silently skip large parts of it. Post-hoc coverage analysis would close that gap — partition before, verify after, re-prompt for what was missed.