keelson — agents that ship fast and stay on-spec 🚢
AI agents are fast — but left unguided they drift: one starts on-spec, then quietly makes
a call that touches the original design, plan, or strategy, and carries on. You'd normally
catch it late — at review, after merge, or once it has compounded across three more changes.
keelson keeps them on course. It draws a hard line between the moves an agent should just
make and the ones it must hand back to you:
- Mechanical → the agent just does it. Claim the issue, branch, code + doc + test, open
the change, move status.
- Consequential → escalate to you. Anything that touches the original design, plan, or
strategy.
The payoff: full momentum on the mechanical, a human gate on the consequential — drift surfaces
the moment it happens, not three steps later. Your agents move fast and stay on-spec. ⚓
Port a disciplined, issue-driven agentic flow into any repo, against any tracker.
keelson is a pair of Agent Skills — adopt-keelson, which interviews you and writes a tracker-agnostic operating model into your repo (a questionnaire-organized AGENTS.md plus a selectable lifecycle skill-pack), and tune-gates, which establishes or refines just the quality gates, standalone.
It distills the flow a production repo uses to let AI agents ship safely:
%%{init: {"flowchart": {"diagramPadding": 22, "nodeSpacing": 40, "rankSpacing": 36}}}%%
flowchart TD
S([enter · a change request<br/>or the next backlog item]) --> F
F[frame] --> P[plan] --> I[implement] --> O[open change] --> G{{gate}} --> C[close]
C --> D([done · merged, status set,<br/>handed back])
D -. you start the next .-> S
P -. strategic call .-> H[/escalate to human/]
G -. structural drift .-> H
H -. redirect .-> P
classDef step fill:#e8eef7,stroke:#33415c,color:#0b1324,stroke-width:1px;
classDef enter fill:#dfeee4,stroke:#2f6b4f,color:#0b1f16,stroke-width:1.5px;
classDef done fill:#d7e3f4,stroke:#2c4a73,color:#0b1324,stroke-width:1.5px;
classDef gate fill:#d6ece6,stroke:#2f6b53,color:#0b1f16,stroke-width:1.5px;
classDef human fill:#f7e9d0,stroke:#8a5a16,color:#241a0b,stroke-width:1.5px;
class F,P,I,O,C step;
class S enter;
class D done;
class G gate;
class H human;
linkStyle 7 stroke:#2f6b4f,stroke-width:1.5px,stroke-dasharray:4 3;
linkStyle 8,9,10 stroke:#8a5a16,stroke-width:1.5px,stroke-dasharray:4 3;
Tactical autonomy, human-in-the-loop on drift
Tactical autonomy — the agent just does the mechanical moves (claim the issue, branch, code + doc + test, open the change, move status). Strategic human loop — it escalates the moment a move touches the original design, plan, or strategy. That balance is the whole point: momentum on the mechanical, a human gate on the consequential.
%%{init: {"flowchart": {"diagramPadding": 22, "nodeSpacing": 48, "rankSpacing": 40}}}%%
flowchart TD
M([a move in the loop]) --> Q{touches the original<br/>design, plan, or strategy?}
Q -- no --> A[agent just does it<br/>· tactical autonomy ·]
Q -- yes --> H[escalate to the human<br/>· strategic loop ·]
classDef start fill:#eef1f6,stroke:#33415c,color:#0b1324,stroke-width:1px;
classDef decision fill:#ece8fb,stroke:#5b4b8a,color:#1e1633,stroke-width:1.5px;
classDef agent fill:#e8eef7,stroke:#33415c,color:#0b1324,stroke-width:1.5px;
classDef human fill:#f7e9d0,stroke:#8a5a16,color:#241a0b,stroke-width:1.5px;
class M start;
class Q decision;
class A agent;
class H human;
linkStyle 0 stroke:#33415c,stroke-width:1.5px;
linkStyle 1 stroke:#33415c,stroke-width:1.5px;
linkStyle 2 stroke:#8a5a16,stroke-width:1.5px;
It's provider-neutral. The skills name actions — "create the issue", "open the change", "set status to <column>" — and your agent runs whatever your tracker uses: GitHub, GitLab, Jira, Linear, or anything with an API/CLI/MCP. The only repo-specific values are filled from your interview answers.
What you get
adopt-keelson writes two things into your repo:
AGENTS.md — the substrate every skill defers to: your vocabulary, branch/change rules, status columns, the InnoVestrum engineering standards, and the chosen tracker's traps folded in under Tracker notes.- A lifecycle pack — the ideas you pick (below), written as your repo's own skills in plain language.
Thin templates plus your interview answers become your repo's own artifacts:
