Evidence before excitement
Open the original release, documentation, paper, or advisory. Check dates, versions, availability, limits, and whether the claim is more than marketing.
I built the AI Research Command Center to shorten the distance from a meaningful technical development to a tested improvement in my own projects—while keeping every release decision mine.
New models, APIs, security advisories, libraries, and standards arrive faster than I can evaluate them. A headline alone does not tell me whether anything changed in code I actually maintain. I wanted a repeatable way to find the signal, test the project fit, and make progress without handing over the final decision.
Open the original release, documentation, paper, or advisory. Check dates, versions, availability, limits, and whether the claim is more than marketing.
Compare a finding with the current repository—not a stale copy or a broad topic match. Most interesting news should remain research, not become an automatic edit.
Let the system prepare a small local draft and demonstrate it. Then I choose whether to approve, revise, or decline that exact change.
The research conversation finds and remembers what changed. A separate draft chat only starts when there is a verified, bounded improvement. The two paths meet at a review packet, then stop for my decision.
The diagram shows the decision boundaries. Every connection is explained in the steps below.
The arrows show the actual handoffs and stop points. A finding can stay research-only, wait for the draft slot, or become a local proposal. No arrow goes straight from a report to production.
There is one recurring research task. It does not edit projects while it researches. Drafting is a separate, supervised phase after the report has been saved.
At 8:00 AM New York time, Monday through Friday, the same task picks up where the last successful check ended. The local computer and desktop app need to be running for local repository access.
Search AI, software, security, developer tools, APIs, agent systems, startups, and standards. Prefer original documentation and releases; use credible reporting for context. Verify event dates, versions, availability, and limitations.
Refresh the active repository inventory and group related repositories. Check relevant remote default-branch code, dependencies, instructions, commits, and open work before saying a project needs a change.
Publish a dated CRITICAL / IMPORTANT / WATCH report with sources, project connections, and bounded next actions. It closes with attention priorities, one experiment, one project improvement, and an evidence-backed opportunity when justified. Keep a 48-hour search overlap, suppress repeats, retain failed-source gaps, and say when nothing significant qualifies.
Only a CRITICAL or IMPORTANT finding with a verified match to current code and a bounded local change with a measurable check can advance. WATCH, coursework, speculation, and already-addressed work remain research-only.
A separate ledger identifies the underlying finding, material version, and target repository. It prevents duplicate dispatch, allows only one unfinished draft across the system, and starts at most one draft per report. Later candidates wait their turn.
Every selected candidate starts a new chat in my FuegoWork Apps Codex project. The chat is the home for the proposal, implementation, evidence, and approval request.
The target repository gets its own named local branch and worktree. When the target is another repository, its code stays in that separate worktree; the FuegoWork Apps chat's own worktree remains untouched.
The draft runs appropriate local checks with synthetic data and makes a local commit. Its review packet includes the exact diff, results, limits, and before/after screenshots or fixture output.
I can approve the exact reviewed commit, request changes on the same branch, or decline and retain the branch. Approval permits only a guarded local integration if the remote base has not advanced and the target checkout is clean.
Pushing code, opening a pull request, deploying, spending money, running GitHub jobs, and submitting coursework are outside this automatic draft flow. A local approval is not a release.
The gate is selective on purpose. Choose a situation to see what the system does next.
The candidate takes the open ledger slot. A new FuegoWork Apps chat makes the change in an isolated target worktree and returns a packet for your decision.
FuegoWork Apps is the place I review every draft. The project being changed owns the branch and files. That distinction matters when a finding applies to a different repository.
Scope, context, implementation story, evidence, and the final approval question all live together.
A named local branch holds the edit. Tests run here. Other checkouts and active work stay in place.
One pilot examined an AI CLI stream boundary with synthetic events. The earlier behavior accepted extra events after a terminal result. The local draft made those sequences fail as protocol errors, while a valid single result still succeeded.
The packet records the actual fixture output and test results, including limits from the broader suite. Synthetic input demonstrates the adapter rule; it does not claim the real CLI emitted those sequences.
result→late event→acceptedresult→late event→protocol errorThe completed draft asks in its own FuegoWork Apps chat. The decision applies to the exact reviewed work, not an open-ended permission to change or publish later.
Recheck the base and target checkout. If clean and unchanged, integrate the reviewed commit into the local default branch. If conditions changed, return it for review.
The same chat and branch receive the revision. A changed commit gets a new evidence packet and a fresh decision.
The branch is retained for inspection until I choose what to do with it. Nothing is silently merged or removed.
Approval here is not a push, pull request, deployment, paid job, or assignment submission. Those are separate, explicit choices.
This is a working, supervised research-to-draft workflow for my own development projects. It lets me keep learning without turning every announcement into a task—or every draft into a release.