Fullsend

Appearance

73. Named mint privilege levels under agent roles ​

Date: 2026-07-19

Status ​

Accepted

Context ​

The mint maps each agent role to a single hardcoded permission set (ADR 0007). Every token minted for a role gets the same ceiling. The threat model establishes least privilege as a cross-cutting principle (see security-threat-model.md), but the current model cannot differentiate between a phase that only needs read access and one that needs write — both receive write-level tokens.

#2821 (human-gated permission adjustments) and the least-privilege topic (#5312) depend on a general mechanism for minting tokens at differentiated privilege levels within a role. Without this mechanism, downstream work risks encoding use-case-specific designs instead of a reusable model.

Decision ​

Role levels ​

Each role defines an ordered set of named privilege levels. Each level's permissions are a superset of all preceding levels. The first level is always read — generally read-only — and is the default when no level is requested.

The read level for each built-in role is derived by taking the role's current permission set and downgrading all *:write permissions to their read counterparts (e.g., contents: write becomes contents: read). The write level is defined as the full permission set currently granted by each built-in role. For roles where write is not explicitly defined, requesting write falls back to read automatically.

Mint API ​

Token requests accept an optional level field:

json
{"role": "coder", "level": "write"}

When level is omitted, the mint defaults to read. The mint validates that the requested level exists for the role and returns an error for unknown levels.

Custom roles ​

CUSTOM_ROLE_PERMISSIONS continues to accept the current JSON shape — a flat role-to-permissions map — interpreted as the read permissions for each role:

json
{"my-role": {"contents": "read", "issues": "write"}}

An alternate shape specifies multiple named levels per role. The mint auto-detects the format per role by checking whether the role's value contains a levels key (reserved for format discrimination — role definitions must not use levels as a permission name):

json
{"my-role": {"levels": {"read": {"contents": "read"}, "write": {"contents": "read", "issues": "write"}}}}

Mixed format is allowed within a single CUSTOM_ROLE_PERMISSIONS value — some roles may use the flat format while others use the multi-level format. The mint detects the format independently for each role.

Clients ​

mintclient and CLI paths that mint tokens accept an optional level parameter, defaulting to read when unset. The harness passes the level derived from the privilege_levels configuration (see below) to the mint client for each phase.

Acquisition semantics ​

The harness is the sole caller that selects privilege levels. It determines the level for each run phase from the privilege_levels configuration and requests the corresponding token from the mint before that phase begins. Agents and scripts within a phase do not choose or escalate their own level — they receive the token the harness minted for their phase.

Harness privilege_levels flag ​

Harness YAML supports a privilege_levels field mapping run phases to levels:

yaml
privilege_levels:
  pre_script: write
  runtime: read
  post_script: write

A default key specifies the level for phases not explicitly listed. Phases not shown above — such as validation_loop — inherit the runtime level, since they execute within the runtime phase and share its security context.

When privilege_levels is omitted entirely, the harness behaves as if:

yaml
privilege_levels:
  default: write

This preserves backward compatibility — existing harnesses continue to receive write-level tokens for all phases.

Consequences ​

  • Agents running in the LLM sandbox can receive read-only tokens, reducing blast radius if the sandbox is compromised.
  • The mint API defaults to read when level is omitted, producing narrower tokens than the current behavior. Non-harness callers that omit level — including reconcileMintToken, the mint-token CLI, and e2e test helpers — will receive read-level tokens instead of write-level tokens. Migration of these callers is coordinated in #2823 and #2826.
  • CUSTOM_ROLE_PERMISSIONS supports both the existing flat format and the new multi-level format — including mixed format within a single value — without a breaking change.
  • Harnesses that omit privilege_levels default to write for all phases, so existing harness configurations are unaffected — the harness requests write from the mint, preserving current token scope.
  • Implementation of role levels in the mint, clients, and harness is tracked in #2823 and #2826.
  • #2821 (human-gated permission adjustments) is unblocked and can build on this mechanism.