Skip to content

151. Use harness human-role minimums for actor admission ​

Date: 2026-10-23

Status ​

Accepted

This decision applies only to CEL/entity routes and migrated v2 model-bound admission. Legacy non-CEL, non-entity workflows remain under Authorization v1 and their existing workflow guards.

ADR 0170 and entity-context-snapshot v2 are the versioned snapshot contract that represents the human_min_role values.

Earlier ADREffect of this decision
ADR 0054Apply the harness minimum before CEL on migrated routes.
ADR 0098Apply it to decisive actor-attributed entity facts.
ADR 0133 and ADR 0141Apply it at model-record admission and exposure gates.
ADR 0061Resolve the minimum before the harness actor gate and CEL.
ADR 0107A human label grant meets triage only; bots remain separate.

Context ​

ADR 0061 puts platform authorization before harness CEL. Authorization v1 sets different human thresholds for observation and mutation routes, while a harness's role selects its run credentials. Credential scope does not decide which people may trigger a harness or contribute actor-authored data to it. A single triage threshold could let lower-privilege actors influence a write-capable harness. A per-harness minimum addresses that gap; its trusted value must be resolved before it gates admission. ADR 0112 allows trusted-fact-guarded overlays to set such fields.

Options ​

Keep one triage threshold for every harness and rely on CEL for stricter routing; derive a threshold from token permissions; or use an explicit default mapping from mint role with a configurable, platform-enforced harness minimum.

Decision ​

  1. Fullsend enforces a non-waivable triage floor for ordinary human actors. Bots use their separate registered-role and label-handoff rules.
  2. Each harness has an effective human_min_role. It defaults to triage for role: triage or role: review, and to write for every other role, including custom roles. An explicit value may be triage, write, maintain, or admin; an invalid value makes the harness ineligible.
  3. Trusted top-level values and guarded overlays may set the minimum. An overlay may preserve or lower its layer's non-overlay baseline, never raise it. A missing optional guard keeps the baseline; unverifiable required evidence or an evaluation error fails closed. CEL and deprecated forge: blocks cannot set or relax the minimum.
  4. The same effective minimum gates the human prompting event, decisive actor-originated entity facts, and Fullsend-controlled model-bound records. CEL may narrow admission further but cannot relax these gates.
  5. A verified human label addition grants triage only for its exact accepted transition and proven projection. It does not create a verified role; a stronger harness minimum requires the labeler's verified effective role.

Authorization Contract v2 defines the policy sources, overlay-resolution and provenance rules, per-record gates, auditing, and run-time rechecks. ADR 0144 defines the bound-run lifecycle. These v2 rules do not extend to legacy non-CEL, non-entity workflows.

Forge scope ​

GitHub, GitLab, and Jira use their verified normalized-role mappings. Jira Developer and Administrator satisfy triage because Jira has no separate triage level. Forgejo must provide a verified effective role or deny the role-based path. Bot and label grants require separate provider evidence.

Consequences ​

  • The human_min_role field is target behavior until the loader validates and enforces it. Existing legacy routes keep their Authorization v1 guards.
  • A harness author may deliberately set human_min_role: triage for a write-capable role; that permits triage actors and records to guide that harness without changing its minted capabilities.
  • Each harness applies its own minimum to shared verified role evidence; authorization decisions and filtered views are not shared across minima.
  • Entity-context-snapshot v1 retains its observation/mutation-derived triage/write threshold. A different snapshot threshold requires a versioned contract.