Specialist review agents on a shared message board with Claude Code, Codex and MCP
In Triglav the code is written by Senior agents. Before a Senior starts, the plan goes to a group of researchers. These are read-only agents that advise on a shared message board. Triglav does not come with a fixed team of them. You create the researcher personas yourself and decide what each one focuses on, so the group can fit your own codebase and your own worries. My group has five: a bug hunter, a security reviewer, a planner, a detailed challenger and a checker of the agent instructions. They read the plan, open the repositories, and post their answers on the board, where I can read every word.
Between 7 October and 12 November, Triglav ran 184 jobs on my machine and 142 are done. The board holds 4,344 messages. This post shows how the specialists work together, using four real threads.
The advisors
A researcher in Triglav is a persona that you add on the Researchers page. It has a name, a focus of up to 500 characters, a repository, a model and an effort level. The runtime is Claude Code or Codex, and mine all run on Claude Code. Nothing is built in, so Triglav has no default personas. The focus line is free text, and it is the only thing that tells a persona what to look for.
A persona has no process of its own. Triglav sends every question and proposal to every enabled persona. A mention goes only to the persona it names, a decision goes only to the personas that objected on that thread, and a review round goes to the personas that already took part in that job. When something is due, Triglav starts a short read-only run. The personas cannot change job state. They give advice and Staff decides.
I created these five between 28 October and 11 November. The focus column is my own focus text, shortened.
| Persona | Model | Focus |
|---|---|---|
| BugHunter | Claude Sonnet 5.5, high effort | Bugs and side effects. Looks wider than the diff: the PR and the code around it. |
| Sonnet Medium Security | Claude Sonnet 5.5, medium effort | Application security. Challenges plans, messages and Client decisions as a security specialist. |
| Opus High Researcher | Claude Opus 5.5, high effort | Helps Staff plan and decide. Challenges Senior decisions. |
| HaikuHigh2 | Claude Haiku 5.5, high effort | Detailed repo analysis. Challenges other agents when it has good arguments. Reports only major and critical findings. |
| Fable | Claude Fable 5.2, low effort | Checks Triglav’s own agent instructions against the code. |
Every plan goes to all five, and each has its own model and focus line, so they do not read it the same way. The next threads show what that gives.
One plan, five readers
Staff opens a plan review as a timed question. The post has the problem, the change, the alternatives it rejected, the premises it wants checked and the acceptance criteria. The researcher rules tell each advisor to write down how it would solve the problem before reading the plan, and then to check the premises in the code.
One plan on 12 November changed a function in a calculation library, so that it returns the error of each failed statement. One acceptance criterion said that a circular reference must return its error with a source tag and an error code, on every statement in the cycle.
Fable and Opus both opened the engine source and found that this error has no source and no code. It is a plain message. They also found that the engine puts it on only two statements of a longer cycle, so “every statement in the cycle” could never pass. BugHunter found that the app detects cycles by matching the message text, so the plan needed a boolean flag. Fable added that the match should be done once, inside the package. HaikuHigh2 found that a second criterion, about statements left to the server, would fail unless errors are filtered one by one, and that the existing comparison helper could not give the plan’s true, false and undefined results.
Staff did not accept this on trust. It read the engine at the pinned version tag, confirmed the lines, and changed the plan: a cycle flag set once inside the package, the criterion rewritten to name the pair of statements the engine marks, and the existing reviver function exported, so no second one is written. The whole thread took under three minutes.
Security and bug hunting on the same thread
A second plan added an allowlist of one command to a shared option of a rendering library. Every advisor added something.
Security noticed that allowing a command still allows any value after it, so raw input could set a CSS class and cover the page. Opus read the library source, found that the library passes the class value to the check, and proposed restricting the value as well, which is one more line. Staff opened the same source file, confirmed it, and accepted. BugHunter confirmed the call sites and found one more renderer with its own default that the plan had not listed. HaikuHigh2 asked what the change does to content that is already stored, which a repository search cannot answer.
Later a Senior who was implementing the plan found that one of Staff’s own premises was wrong. Staff said so on the board, raised a decision for me with a recommendation, and told the Senior to keep implementing while it waited. Three advisors supported the recommendation, each from a different angle, and HaikuHigh2 disagreed. I made the call.
Changing a position on evidence
A third thread was a design for deleting and restoring a page across several services. Staff posted a proposal and asked for it to be challenged.
Opus found a sequence in the proposed design that loses data. The sync cursor can move past a restore marker while the page row is still marked deleted, and then the calculation never comes back. Security asked that the worker take the workspace id from the stored row and never from the request payload. HaikuHigh2 first demanded that the server refuse writes to deleted items, Then it read the client code that sends offline edits, found that a refused write would lose them, and withdrew the demand with the lines as evidence.
Staff recorded seven decisions and then amendments, each pointing to the message that caused it. The dissents that stayed open went into the report.
Agents audit their own instructions
The fourth thread is about Triglav itself. An earlier scan of the agent instructions claimed that a line in AGENTS.md was wrong, and Staff asked the advisors to verify it. The answer depends on which CLI loads which file, and most advisors had no shell to run the CLIs. They said so in the replies and marked what came from memory. Security read the code and corrected BugHunter on which settings source the review seats use. Opus found the relevant strings in the installed Codex binary. Fable reported what its own run had loaded.
Staff’s decision keeps the three levels apart. The line is wrong for review seats, which was verified in code because the review prompt injects the file. It is probably wrong for Codex runs, from memory. The claim about an empty settings source rests on one observed session. Staff also wrote that a one-command check has to run before any file is deleted.
What keeps the board useful
Every advisor gets the same guide. These are the rules from it that matter most here:
- Form your own view before you read the others. If you agree, say why in your own words.
- Say what you checked, in the form “ran X”, “read file:line” or “not checked”. Most advisor replies in the threads above end with a line like that.
- Agreement is not authority. Many agents repeating a claim does not make it true.
- At most three messages per thread unless someone mentions you, and no “+1”.
- An objection about safety or data loss stops that line of work until Staff answers it.
- If an objection stands after Staff’s decision,
file_dissentputs it in front of me, once.
After the plan, the code goes through the local review panel before any PR. The default panel is a strong Codex seat, a cheap Claude Sonnet seat and two Claude Haiku lens seats, for contracts and for integrity, that run in shadow mode and never block. On review rounds the advisors read board text only and never open the diff.
Try it
Triglav is source-available on GitHub at github.com/adam-drag/triglav. Personas are added on the Researchers page with a name, a focus, a repository, a model and an effort. The five above are only my choice. A different codebase may want a performance reviewer, an accessibility reviewer or a database specialist instead.