pi-operator-extension issues

/op:supervisor-review promises session or claim-set review, while review-delegate currently accepts one claim and needs a verify command.

source:

next: Narrow wave-1 supervisor-review to one claim, or explicitly model claim-set review as repeated single-claim bundles.

task hint: none · claims: no matching claims

/op:delegate has no general CLI route to an implementer that is not already assigned or reviewing the task.

source:

next: Restrict chooser targets to routable harnesses or make new implementer delegation create a scoped child task.

task hint: none · claims: no matching claims

Session-derived harness ids, carrier ids, and authority identities are separate axes and current adapters do not map all of them cleanly.

source:

next: Decide the minimum identity/target config before implementing delegate launch.

task hint: none · claims: no matching claims

review-delegate may launch reviewing Pi with --approve, which can trust the project-local extension being reviewed.

source:

next: Open a separate Operator task for the review-delegate approval/isolation defect.

task hint: none · claims: no matching claims

Dogfood acceptance must be falsifiable, not just a prose judgment that the ledger is clearer.

source:

next: Define a binary dogfood gate such as non-self-verification plus required ledger artifacts.

task hint: none · claims: no matching claims

Distinct-UID verification currently blocks on ad hoc sudo/auth state; the correct UX is a human-visible authorization prompt, not a hidden agent retry or same-UID downgrade.

source:

next: Add a supervisor-review/verifier-auth design slice that prompts the user explicitly when a verifier UID run is needed and records advisory vs trusted outcome distinctly.

task hint: none · claims: no matching claims

Verifier-only identities cannot attach draft/no-status evidence because evidence attachment without verification requires builder authority; verifier rerun logs must be attached as verified evidence or supplied through a builder-produced artifact path.

source:

next: Make /op:supervisor-review distinguish advisory review notes, builder-owned draft artifacts, and verifier-owned status-setting evidence so the UI does not offer an impossible no-status verifier attach path.

task hint: none · claims: no matching claims

/op:handoff treated a one-word user input "go" as a literal next_action and recorded an empty handoff with null fields instead of generating a useful closeout from current task context.

source:

next: Change /op:handoff UX so it generates a deterministic draft from task/status/recent handoffs first, defaults to generated closeout without requiring "go", and refuses near-empty handoffs unless the user explicitly uses m

task hint: none · claims: no matching claims

Generated review_delegations scripts are not yet reliable end-to-end: operator-verifier may lack Pi provider credentials/home setup, sudo resets PATH so pytest may be unavailable, pytest cache writes warn under the verifier UID, and the generated prompt still suggests --verified-

source:

next: Implement /op:verify-run or repair review-delegate generation so the verifier launch validates home/provider/PATH first, uses a writable cache or disables pytest cache, and instructs verifier evidence attachment with the

task hint: none · claims: no matching claims

Review/delegation UI exposed too much harness plumbing: users reason in terms of model/persona targets such as Claude/Luna/Grok, while Pi is a carrier runtime and operator-verifier is a separate Unix authority identity.

source:

next: Rename prompts/docs around model/persona targets, remove Pi as a default delegation target, and keep reviewer labels distinct from verifier Unix users in generated review instructions.

task hint: none · claims: no matching claims

review-delegate resolved the Grok persona to the wrong provider/model: the intended xai/grok-4.x target was generated as openai-codex/gpt-5.6-luna (the Luna seat's cloud model), so pi hunted for a missing openai-codex credential instead of the present xai grant. Same GroK-vs-Luna

source:

next: Make review-delegate persona→provider resolution bind Grok to the configured xai provider/grok model (never a different seat's cloud command), keep review_harness and the generated run_command consistent, and have doctor

task hint: none · claims: no matching claims