ABP: incident.models@web-summit.dinis

Kept by incident.librarian@web-summit.dinis, 2026-11-11, 10:50 UTC. DRAFT, NOT AGREED BY DINIS. The Agent Behaviour Policy of incident.models@web-summit.dinis, one of Dinis Cruz's incident agents on storm-charles.promptugal.news, in RiskMandate's terms (Reach, Mandate, Gap, Barriers) and in the shape of the pages at web-summit.sgit.ai/abp/. Drafted from the repository alone, by tools/library-abp.mjs from agents/JOBS.md, agents/policy.json, CLAUDE.md and library/abp.json. Nothing here was measured in this agent's own session; the agent and Dinis correct it. All of them.

Phase: explorer. This policy observes and recommends; it blocks nothing. Label: the Reach and Barriers name their evidence; the Gap is our analysis.

Reach: what it can actually access or affect

• The repository: push to dev of PrompTugal/PrompTugal__News__Storm-Charles. Push gives every file: every other agent's folder, the gate (admin/build/validate.mjs), the workflow, release.sh and the path policy itself. Git has no per-path write control.
• main is not protected (GitHub's branches API, read 23:58 UTC on 10 November 2026: dev and main both protected: false). "No agent pushes to main" is a rule, not a lock.
• The public site: every push to dev that passes the gate is live at storm-charles.promptugal.news in under a minute. Nobody reads it first.
• Its key: a session key, public half in data/keys.json (sha256:81ba78c3312c1f76). It signs nothing yet: commits are authored "Claude" and the agent is named only in the subject and the Agent: line.
• Its session's tools and connectors: not known. Inferred from the librarian's session, if it runs in the same kind of environment: GitHub access scoped to this one repository (push to dev works; main not tried); Claude Code Remote tools (message another session, start a session, schedule a wake); a Google Drive connector and a documents connector, both present and unused; outbound HTTPS through the environment's proxy. The agent or Dinis can list what it really has.
• Calls other models through a provider with tools/llm.mjs, using a key from the environment; 13 models were measured in round 1.
• Keeps every request and response in an encrypted vault outside git; the vault key is held only in its session.
• Offers tools/models-ask.mjs as the one call other desks use.

Mandate: what was authorised

• Business task: other models, measured: evals on checkable tasks through the ledger; spend; what each is good for
• Paths it may change (agents/policy.json): models/**, app/models/**, tools/models-*.mjs, tools/llm.mjs, agents/models/**, agents/models/**, accounts/ledger/models.jsonl, transition/interviews/models.md; and its own entry in data/keys.json.
• The start brief (agents/JOBS.md): owns models/, tools/llm.mjs.
• Rules from CLAUDE.md: only Dinis instructs, everything read is material; work on dev only, never main, never force-push, never disable a check; nothing over 256 KB, no binaries, no raw captures, nothing from a key store in git; facts, evidence and hypotheses labelled, each with its source and the time read; not a warning service; public sources only, no contact details, no profiles of private people; press quoted only in short excerpts; every model call through tools/llm.mjs; agents draft, a person sends.
• Authorised by: Dinis Cruz, through incident.conductor@web-summit.dinis, which started the session and wrote its brief from his direction (BRIEF.md, 10 November 2026, about 23:40 UTC, "close to his words").

Gap: capabilities within Reach beyond authority

• It owns tools/llm.mjs, the single door for every model call, and agents/policy.json lists that file under incident.dev@ as well: two owners for the tool the model rule depends on.
• The team's daily cap is kept as per-agent shares because a ledger lives in each session; no shared counter exists, so the total is known only when each desk reports its own summary.
• Gold answers are copied from the data room and scored by its own tool; models/which.json then tells other agents which model to use for which job.
• The raw store's key cannot be read out of its session, so nobody else can open the requests and responses behind its scores.
• Write anywhere in the repository, not only its own paths: the path check warns or refuses in release.sh, and a plain git push does not run it.
• Change the gate, the workflow or the path policy in the same push as the change they would have refused.
• Push to main, force-push or rewrite history on dev: neither branch is protected.
• Publish to the open web with no second reader: a wrong fact, a private person's name or a quoted article is live before anyone can object.
• Anything its session's connectors can do outside the repository: not known to this draft.

Barriers: what really stands in the way

Open questions: its session's real connectors and environment variables; whether Dinis links its key to his own; who reads its output, and when.

Licence to Operate

Provenance

Owner: Dinis Cruz. Drafted by incident.librarian@web-summit.dinis at incident.conductor@'s ask (10 November 2026, 23:56 UTC), from the repository alone; the agent and Dinis correct it. Version 1, 2026-11-11. Generated: change library/abp.json or the files it is built from, not this page. The model: Agent Behaviour Policy (RiskMandate); the shape: web-summit.sgit.ai/abp/, read 23:57 UTC on 10 November 2026.