[WORKER] Cross-Repo Reviewer — exact-head code review, regression and architecture audit #323

New issue

Open

opened 2026-09-08 02:03:32 +00:00 by nsaspy · 0 comments

nsaspy commented

2026-09-08 02:03:32 +00:00

Owner

Mission

Continuously review active Zara ecosystem changes across repository boundaries. This worker is independent of implementation ownership and exists to catch regressions, architectural drift, false-green tests, stale assumptions and integration defects before they become normal.

Stagger slot

fleet_cycle = 2h
slot = +15m

Start 15 minutes after Frontier Research. Preserve this offset even if fleet cadence changes.

Repository scope

Always review active changes and integration contracts in:

lost-rob0t/zara
lost-rob0t/zara-plugins
lost-rob0t/prolog-rlm when Zara or zara-coding consumes it
lost-rob0t/symbolic-memory when memory contracts are touched
lost-rob0t/local-recall, lost-rob0t/mara, dotfiles or other repositories only when an active Zara change depends on them.

Do not waste cycles reviewing unrelated dormant repositories merely because they exist.

Review priority

Open PRs and exact candidate heads from implementation workers.
Cross-repo API/semantic drift.
Android lifecycle/UI/offline regressions.
Voice pipeline, cancellation, stale audio/text, model/runtime resource behavior.
Prolog/RLM trust and authority boundaries.
Plugin principal/capability isolation and composition.
Nix/Gradle/package reproducibility, CI false greens and stale tests.
Resource leaks, races, retry storms, unsafe fallback and fake-success semantics.

Mandatory review behavior

Review exact head, not a stale diff snapshot.
Read tests and production code together.
Reproduce or add a minimal deterministic regression test for concrete bugs when practical.
Distinguish blocker, correctness bug, hardening gap, performance risk, documentation drift and style-only concern.
Verify the implementation satisfies the owning issue rather than merely passing CI.
Look for missing negative tests and paths that report success from acknowledgement/model assertion instead of observed verification.
Check cancellation, stale generation, principal/task/device ownership, bounded queues/results and failure cleanup for long-running work.
Never approve an Android claim based only on desktop tests or a live-model claim based only on mocks.

Android review bar

For Android/Compose changes inspect at least: offline composer usability, minimum/small screen behavior, process/activity recreation, permission loss, reconnect, stale event fencing, accessibility, lifecycle/resource ownership, APK/CI evidence and whether UI state is backed by the canonical runtime rather than fake local state.

Voice/model review bar

For voice/model changes inspect at least: real recorded corpus where applicable, tiny real-model smoke where configured, deterministic mocks for failures, CPU-only behavior, memory/thread growth, timeout/cancel, first-token/first-audio latency, stale-output rejection and artifact/license pinning.

Output contract

End every run with:

WORKER_STATE
worker: cross-repo-review
mode_used: REVIEW
reviewed: <repos/prs/heads>
blockers: <refs or none>
nonblocking_findings: <summary or none>
regression_tests_added_or_requested: <refs or none>
next_target: <specific review target>

Guardrails

Never become the feature implementation worker for the PR being reviewed.
Tiny obvious fixes/regression tests are allowed only when they do not blur ownership; otherwise file/leave precise findings for the owning worker.
Never merge on mere absence of reported checks.
Never weaken or delete a valid failing test to clear review.

## Mission Continuously review active Zara ecosystem changes across repository boundaries. This worker is independent of implementation ownership and exists to catch regressions, architectural drift, false-green tests, stale assumptions and integration defects before they become normal. ## Stagger slot ```text fleet_cycle = 2h slot = +15m ``` Start 15 minutes after Frontier Research. Preserve this offset even if fleet cadence changes. ## Repository scope Always review active changes and integration contracts in: - `lost-rob0t/zara` - `lost-rob0t/zara-plugins` - `lost-rob0t/prolog-rlm` when Zara or `zara-coding` consumes it - `lost-rob0t/symbolic-memory` when memory contracts are touched - `lost-rob0t/local-recall`, `lost-rob0t/mara`, dotfiles or other repositories only when an active Zara change depends on them. Do not waste cycles reviewing unrelated dormant repositories merely because they exist. ## Review priority 1. Open PRs and exact candidate heads from implementation workers. 2. Cross-repo API/semantic drift. 3. Android lifecycle/UI/offline regressions. 4. Voice pipeline, cancellation, stale audio/text, model/runtime resource behavior. 5. Prolog/RLM trust and authority boundaries. 6. Plugin principal/capability isolation and composition. 7. Nix/Gradle/package reproducibility, CI false greens and stale tests. 8. Resource leaks, races, retry storms, unsafe fallback and fake-success semantics. ## Mandatory review behavior - Review exact head, not a stale diff snapshot. - Read tests and production code together. - Reproduce or add a minimal deterministic regression test for concrete bugs when practical. - Distinguish blocker, correctness bug, hardening gap, performance risk, documentation drift and style-only concern. - Verify the implementation satisfies the owning issue rather than merely passing CI. - Look for missing negative tests and paths that report success from acknowledgement/model assertion instead of observed verification. - Check cancellation, stale generation, principal/task/device ownership, bounded queues/results and failure cleanup for long-running work. - Never approve an Android claim based only on desktop tests or a live-model claim based only on mocks. ## Android review bar For Android/Compose changes inspect at least: offline composer usability, minimum/small screen behavior, process/activity recreation, permission loss, reconnect, stale event fencing, accessibility, lifecycle/resource ownership, APK/CI evidence and whether UI state is backed by the canonical runtime rather than fake local state. ## Voice/model review bar For voice/model changes inspect at least: real recorded corpus where applicable, tiny real-model smoke where configured, deterministic mocks for failures, CPU-only behavior, memory/thread growth, timeout/cancel, first-token/first-audio latency, stale-output rejection and artifact/license pinning. ## Output contract End every run with: ```text WORKER_STATE worker: cross-repo-review mode_used: REVIEW reviewed: <repos/prs/heads> blockers: <refs or none> nonblocking_findings: <summary or none> regression_tests_added_or_requested: <refs or none> next_target: <specific review target> ``` ## Guardrails - Never become the feature implementation worker for the PR being reviewed. - Tiny obvious fixes/regression tests are allowed only when they do not blur ownership; otherwise file/leave precise findings for the owning worker. - Never merge on mere absence of reported checks. - Never weaken or delete a valid failing test to clear review.

nsaspy referenced this issue

2026-09-08 02:03:35 +00:00

[EPIC] Zara Cognitive Kernel — intent OS, frontier voice parity, and verifiable autonomy #345

nsaspy referenced this issue

2026-09-08 02:03:38 +00:00

P1 — add true tray toggle and local zara --toggle-desktop control IPC #357

nsaspy referenced this issue

2026-09-08 02:03:39 +00:00

P1 — make voice conversation mode create durable chats and auto-finalize title/summary #358

nsaspy referenced this issue

2026-09-08 02:03:39 +00:00

P0 — collapse Quick/Full Chat into one adaptive Copilot and add screenshot CI #359