SikongJueluo
061c60c344
- add config v2 with fixed judge model selection and fail-closed v1 enforce migration to shadow - add built-in high-risk override for irreversible, publish, system, and credential shapes that always defers to the human - remove the promotion gate, promotion records tool, and their tests - fail closed with judge_model_unavailable when a configured judge model cannot be resolved - notify once per session in enforce mode with the judge model and risk contract - add ADR 0008 and update README and CONTEXT
3.5 KiB
pi-extensions
Personal pi coding-agent extensions. This context covers the permission
extensions that inspect and re-evaluate Bash commands before they are allowed.
Language
Authority
Enforce authority:
Allow-only delegation under a user-assumed-risk contract (ADR 0008):
writing mode: "enforce" in config v2 consents to the selected judge model
approving ordinary operations on the user's behalf. When the Judge's verdict
is allow and every fail-closed health gate in the Enforce truth table
passes (audit, telemetry, result kind, review acknowledgement,
generation-current), the command runs without the human dialog. The Judge can
never answer deny with authority; every uncertain case (defer, deny,
preflight, infrastructure failure) falls back to the human dialog.
Avoid: model safety certification (dropped — see ADR 0008), AI takeover,
auto-deny, full delegation
Audit log (Judge-owned): The Enforce-era accountability record written by the Judge package itself — separate file from the permission-system review log, append + fsync per record, self-checked for health; an unhealthy audit log refuses authority. Avoid: host contract (dropped — see ADR 0006), acknowledged write (upstream sense)
High-risk override:
A narrow code-level preflight recognizing clear-cut command shapes in four
categories — data loss/history rewrite, publish/deploy/infrastructure
destruction, privilege escalation/system modification, direct credential
access — that always defers to the human dialog regardless of verdict. In
Enforce it skips the model call entirely; in Shadow the model is still called
for quality observation. Deliberately not a sandbox: only well-known explicit
shapes, no alias/script/variable analysis, no alwaysPrompt config.
Avoid: command sandbox, alwaysPrompt, opaque-command blanket defer
Irreversibility boundary: An allow verdict requires every operation's effects to be recoverable — reversible, or reproducible from the repository or the evidence at hand. An operation that destroys data which cannot be re-created or undone (unscoped untracked/ignored deletion, discarding uncommitted work, rewriting published history) always defers to the human dialog, no matter how specifically the user requested it; sensitivity alone (e.g. credential refresh) is not irreversibility. Explicit intent grants allow authority only over recoverable effects. Avoid: destructive-but-requested allow (v3 sense — see ADR 0007), risk-based deny
Wrappers
Wrapper:
A Bash command of the form <program> [modifier-args] <inner-command>, where the
authorization question is "what does the inner command do?". Whether a wrapper
may be unwrapped depends on whether its modifier args are transparent.
Avoid: command type, prefix command
Transparent wrapper:
A wrapper whose modifier args do not change which program the inner command
resolves to or its trust boundary (e.g. timeout). Stripping the modifiers and
re-evaluating the inner command is sound: the verdict applies to the same
program that actually runs.
Avoid: safe wrapper
Non-transparent wrapper:
A wrapper whose modifiers change what the inner command actually does in a way
the command string does not capture — e.g. env (its PATH= / LD_PRELOAD
make the inner name resolve to or load a different program) or xargs (its
inner command's arguments are read from stdin). Stripping the modifiers and
re-evaluating the inner command is unsound: the verdict applies to inputs that
are not knowable from the command string.
Avoid: unsafe wrapper