Roadmap

Where fullsend is going, as Now / Next / Later. Those are confidence horizons, not deadlines: Now is what this planning cycle is advancing, Next is likely once Now has room, Later is on the horizon.

This edition is the September 2026 planning snapshot. Themes are how to find work; Now / Next / Later horizons are how to read commitment. Issue tracking spans fullsend-ai/fullsend and fullsend-ai/agents. User support is standing work on every edition until we say otherwise: investigating reports, fixing bugs, and keeping docs accurate. It is real capacity, not the headline of this page.

Earlier published editions live in archived-roadmaps.

At a glance

Theme	Now
Auto-merge	A dedicated opt-in auto-merge agent; multi-runtime lets the multi-model review squad (fullsend#6322) run, which raises trust in that path
Review quality and cost	Compare review approaches using quality and cost together
GitLab	A usable GitLab path, including traces from real runs
Jira	Jira can drive the default agent loop for teams that already live there
Community	Contributor review keeps moving; identify maintainer candidates through the documented process
Platform	Host APIs, public mint, leftover cleanup from deprecated per-org install (ADR 0044), artifacts, behaviour tests
Partners	Tekton and Lightwell stay in motion; not a ship promise
User support	Standing: user reports, bugs, and docs — always on, not the feature headline

Now

What we are actively advancing. Themes are listed in planning-discussion order, not priority rank. Issue links are the work that best matches the September planning discussion, not the full backlog under each theme.

Auto-merge

The deliverable is a fullsend auto-merge agent: dedicated, opt-in, disabled by default, for a narrow class of low-risk work. Repository rules and CI stay the enforcement boundary. CI-aware fixing raises the success rate; it is not a gate on the first foothold.

That agent depends on a trust path the team laid out in planning: multi-runtime so the multi-model review squad (fullsend#6322) can run as the review agent, which should improve trust in auto-merge. Debounce/restart and a tighter triage → code → review → fix loop are the other enablers. Eval measurements and the revert/defect view (next theme) are how we know the path is actually safer. The auto-merge agent will not be recommended for opt-in until revert and defect visibility (fullsend#6892) is operational.

Auto-merge agent

• agents#1132 — Add a dedicated auto-merge agent
• fullsend#3016 — Consider auto-merge for low-risk Renovate patch/minor devDependency updates where CI passes

Depends on: multi-runtime → review squad

• fullsend#587 — Investigate ACP proxy chains for multi-runtime support with consistent security hooks
• fullsend#6322 — Enhance Fullsend review agent with multi-model collaborative review strategy

Depends on: loop hygiene (debounce, review/fix)

• fullsend#1014 — Debounce review dispatch on rapid synchronize events
• fullsend#5666 — Route fix-agent synchronize events to review stage for automated re-review
• agents#343 — Scope re-review to finding verification when push only addresses prior findings
• agents#447 — Review agent should incorporate outstanding human reviews when re-reviewing after PR update
• agents#478 — Post-fix script should reply to review inline threads it addressed

Review quality and cost

Measure the review path that auto-merge will rely on. Run the collaborative review approach beside the current agent (not a day-one replacement). Compare cost over time. Group traces across triage → code → review → fix. Make revert and defect rates visible.

• fullsend#6433 — telemetry: hierarchical work-graph correlation IDs for cost rollup (leaf / parent / feature)
• fullsend#5361 — telemetry: MLflow trace columns (Name/Session/User/Source) blank; token & cost omit cache tokens
• fullsend#6458 — Portable OTLP export for eval measurement scores
• fullsend#6892 — Revert and defect rates are not visible for review and auto-merge outcomes
• agents#209 — Review eval suite has zero coverage — reintroduce review eval cases after post-review 422 fix lands

GitLab

Move from the current foundation through a usable release path. Prove default stages and MLflow export on the same GitLab run, not as separate pieces.

• fullsend#6684 — feat(cli): add --gitlab-url flag to repos install for GitLab bootstrapping
• fullsend#6816 — GitLab MR-event dispatch (fullsend-dispatch.yml) does not pass REPO_FULL_NAME for review stage
• fullsend#6893 — GitLab default agent stages and MLflow trace export are not validated together

Jira

Finish the path from Jira events into default triage and code so teams (for example Konflux) do not copy bugs into GitHub first.

August Jira-poller work that is still open, and how it maps:

• fullsend#4885 — Define Jira authentication strategy: interim service account credentials (issue)
• fullsend#2269 — JIRA identity and credential management (issue; still open, not a September Now item)
• fullsend#3812 — docs(adr): ADR 0079 — Jira project setup via fullsend CLI (open ADR pull request, not a poller tracking issue)
• fullsend#3428 — docs(adr): add ADR 0068 for privacy allowlist on poll input drivers (open ADR pull request, not a poller tracking issue)

The work below is what this edition focuses on.

• fullsend#6672 — jira-poll: built-in harness files lack CEL triggers, causing 0 dispatches
• fullsend#2264 — Add JIRA support to the triage agent
• fullsend#2265 — Add JIRA support to the code agent

Community

Gain additional contributors through the maintainer process. Do not promise specific number of new maintainers. Keep contributor pull requests moving. Progress is tracked through PR review velocity and the maintainer process rather than dedicated issues.

Platform

Host-side APIs for sandboxed subagents, finish public mint, complete leftover cleanup from deprecated per-org installation mode (ADR 0044; this is not new org-mode work), let teams choose artifact storage, and shift tests toward behaviour coverage in the agents repo.

• fullsend#879 — Host-side API servers for sandboxed agents
• fullsend#881 — fullsend run: implement host-side API server lifecycle
• fullsend#5116 — Deploy PROD public CF mint at mint.fullsend.sh
• fullsend#2887 — Remove install_mode input from per-repo shim template
• fullsend#6734 — Configurable artifact storage backend (replace forge-native artifacts)
• fullsend#3786 — feat(behaviour): integrate behaviour test runner in fullsend-ai/agents
• fullsend#3236 — Evaluate overlap between functional tests and behaviour tests + prompt evals

Partners

Keep the Tekton compatibility conversation going. Explore a Lightwell custom-agent workflow after their GA, as a mock/PoC — not a September delivery promise. No public tracking issue yet for either thread.

User support

Standing theme for this edition and later ones until we drop it on purpose. We spend significant time investigating user concerns, fixing bugs, and updating docs. That is real capacity, not a dated deliverable, so it does not get a numbered backlog dump. It still sits last on this page so Auto-merge, GitLab, and the other feature themes remain the planning focus.

Concrete September docs item from planning feedback: land the scheduled / nightly agent example so users are not left guessing how to run a recurring agent.

• fullsend#6601 — docs(#6599): add scheduled agent example guide

Next

Parked or dependency-bound in this planning session. Not dated.

• Isolated sandboxes per subagent (host APIs come first): fullsend#3978 — harness: subagents share the parent sandbox instead of running in isolated sandboxes
• Red Hat AI inference routing — waiting on an inference proxy from another team (no public tracking issue yet)
• Growing eval-measurements past the first review/cost slice: more stages, model experiments, and outcome evals — fullsend#3413 — Design large-scale evaluation experiment for new inference models · fullsend#6384 — Track agents@v0 release cut for eval measurement manifests
• Tool proxies after host APIs land: fullsend#5242 — Tool proxies: transparent CLI interception for sandboxed agents

Later

Direction, not a plan. No dates.

• Kubernetes and OpenShift execution — forge-decoupled agent runtime
• Security hardening, human factors, and production feedback loops — human-factors · governance · production-feedback · tracking: fullsend#172 · fullsend#174 · fullsend#84 · fullsend#159 · fullsend#877 · fullsend#2826 · fullsend#2828
• Agent attestations — cryptographic provenance for agent output: fullsend#267
• Cross-forge orchestration — GitHub + GitLab / multi-org
• Alternative sandbox providers — considered; not worth the refactor now

August 2026 (shipped since last edition)

August planning lived as a flat Now / Next / Later table. The published edition is in archived-roadmaps/2026-08.md. Work that closed in that window includes GitLab polling/dispatch foundations (fullsend#1964, fullsend#5556), removing per-org install mode (fullsend#2302), versioned docs (fullsend#5717, fullsend#5718), public-mint --public support (fullsend#5634), GitLab default-stage environment variables (fullsend#6865), configurable agent status comments (fullsend#3697), multi-endpoint telemetry export (fullsend#5545), and doc-practices guidance (fullsend#5372).

Items from the August edition not listed above were folded into September themes, moved to Next / Later, or deferred — see the archived August edition for the complete table.