Pavel Kerbel (@kerbelp) profile picture

Building the Repository Context Layer (context.md) + Metatron. The project context AI coding agents miss. Independent research, shared in the open.

Joined April 2014
It’s official: I’m presenting our research on Context Inheritance at AgenticDev @ ASE 2026 in Munich on October 12. How can coding agents inherit a repository’s hard-earned lessons instead of relearning them? Coming to ASE? Let’s meet. conf.researchr.org/home/ase-…
🤖 Made with AI
3
305
I run several Codex sessions on my Mac. When they used the same Chrome setup, their browser work could interfere with each other and with my own browsing. We changed how those sessions connect to the browser. Each browser slot now has its own persistent Chrome user-data directory. I sign in directly in that dedicated browser. My everyday Chrome profile stays separate. Each new Codex session gets one explicit Chrome DevTools MCP connection to its assigned browser. A small custom launcher handles it: codex-browser social-a codex-browser social-b Both can run from the same project directory. The launcher starts or reconnects to the right Chrome instance, verifies its profile and loopback debugging endpoint, and holds an exclusive controller lock. If another session already owns that slot, the next launch is rejected. The configuration is specific to that Codex process. Other browser and desktop-control routes are disabled for those sessions. If the assigned browser is unavailable, the agent should report the blocker and leave my regular Chrome alone. We checked the behavior with two concurrent MCP clients and two actual Codex clients: • Each Codex client saw its assigned browser's tabs. • Creating and navigating tabs through one MCP connection left the other browser unchanged. • A second controller for an occupied slot was rejected. • Separate test cookies survived controlled Chrome restarts. • Ending a controller left the browsers and their saved profiles intact. • Recovery after an interrupted launcher preserved the slot's ownership checks. A few boundaries matter. Existing desktop chats keep their previous connections. The isolated setup applies to fresh sessions started through the launcher. Separate browser profiles also don't coordinate actions on the same social account. Two agents can still try to publish the same thing, so account ownership, approvals and duplicate checks still need coordination. And this is protection against accidental interference on my Mac, not an OS security boundary. Other processes running as my user could deliberately reach the debugging endpoints. Browser state needs an owner, and that ownership needs to be enforced when a session starts.
5
4
285
Metatron's files-first usage ledger separates applied, considered and violated decisions, with a commit reference. I'd review the referenced patch before treating an 'applied' entry as evidence that the decision helped. github.com/kerbelp/metatron/…
1
30
When upgrading a coding agent's tools, I'd check the install method and the running server. Metatron uses the detected installer and prints the command if it can't identify one reliably. Restart the serving process before the next session. github.com/kerbelp/metatron#…
16
Another billing test for a coding agent: process a verified refund that revokes access before the payment callback arrives. The later callback should preserve that purchase's revocation. Check the stored state after both events. github.com/kerbelp/dodo-paym…
1
23
One test I'd ask a coding agent to add to a billing change: deliver the same verified webhook twice, then inspect the entitlement or credit ledger. Assert that the effect happened once, even if both requests returned OK. github.com/kerbelp/dodo-paym…
1
29
I published an independent Dodo Payments skill for AI coding agents. It covers the checks beyond checkout: duplicate webhooks, refunds and uncertain charges. A timeout should trigger reconciliation before another payment attempt. github.com/kerbelp/dodo-paym…
1
2
46
In coding-agent evaluations, I look at code localization and resolved tasks separately. Our Context Inheritance replication materials report both. That makes it easier to see whether a change helped the agent find the code, finish the fix, or neither. github.com/getmetatron/conte…
1
1
33
A repo rule can be valid and still be unhelpful for the current task. Metatron's MCP feedback lets an agent rate the decisions it received and report what was missing. I want the next session to inherit useful context, with a way to flag what misled it. github.com/kerbelp/metatron
1
2
23
Useful tip
44 DR dofollow backlink: 1. Go to AI Collection and register. 2. Tap "Submit Your Tool" from footer. 3. Add your tool and URL. 4. Submit it. This site has 31,800+/m traffic. So, grab a backlink...
11
An agent can draft a verification contract. Running it is a separate step. Metatron's --dry-run prints the ordered plan and assertions without executing them, so you can review the commands before running them locally or in CI. github.com/kerbelp/metatron/…
1
26
Just presented Context Inheritance at AgenticDev / ASE 2026. A git-native way for AI coding agents to consult repository context (and write back what they learn). Slides: getmetatron.com/talks/contex…
21
Our Context Inheritance paper is now in the ACM Digital Library. We studied repository context for coding agents: who writes it, who reads it, and how it’s delivered. The paper covers what we observed and the limits of those findings. dl.acm.org/doi/10.1145/38432…
1
44
Who should write repository context for coding agents? Our Metatron study saw different patterns across authors and readers. It’s worth testing who writes the context in your own setup. Context Inheritance at AgenticDev, Oct 12: getmetatron.com/blog/context…
18