[WORKER-FLEET] Zara v2 — staggered research/review/implementation roster and test gates #322

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

Purpose

Authoritative coordination contract for the Zara worker fleet. Individual tracker issues own work; this issue owns stagger, handoff order, collision avoidance and fleet-wide acceptance rules.

Fleet roster and stagger

Use a repeating two-hour cycle. Preserve relative offsets if the launcher cadence is later changed.

Offset	Worker	Tracker	Primary role
+00m	Frontier Research	#645	discover missing capabilities, current Android/voice/model research
+15m	Cross-Repo Reviewer	#646	independent exact-head review across active Zara ecosystem repos
+30m	Core+Android	#623	Android usability, offline/runtime integration, Core hardening
+45m	Voice+Models	zara-plugins#506	voice stack, real tiny models, low-power CPU CI
+60m	Symbolics+Dev	zara-plugins#507	Prolog, coding, memory, expert/sysadmin verified automation
+75m	Providers+Home	zara-plugins#505	SmartThings/home/calendar/comms provider integrations
+90m	Integrations+Registry	zara-plugins#508	remaining integrations, plugin registry/install/package hardening
+105m..+120m	quiet window	—	allow CI, rebases and handoffs to settle before next cycle

Do not launch all workers together. The stagger is intentional to reduce CI/resource spikes, merge/rebase collisions and duplicated issue selection.

Handoff semantics

Frontier Research may create/evolve issues but does not broadly implement them.
Cross-Repo Reviewer reviews active candidate heads independently; it does not become the primary feature author.
Implementation workers consume current source, issue state, latest worker comments, #645 research and #646 findings before choosing a slice.
A worker blocked on another owner's contract files a focused handoff issue; it does not bypass the contract or duplicate another runtime/provider.
Current source and exact issue state win over stale roadmap prose.
Existing active PRs are completed/reviewed before a worker abandons them for its new v2 focus.

Fleet-wide testing law

A feature is not complete because its happy path works.

Use the strongest relevant combination of:

deterministic unit/contract tests;
integration tests across actual ownership boundaries;
property/generative tests for parsers/state machines/budgets;
fuzz/adversarial malformed-input tests;
concurrency/cancellation/stale-generation/race tests;
restart/crash/fault-injection tests;
soak/resource growth tests;
visual/UI/lifecycle tests for desktop/Android;
real recorded speech + real tiny-model smoke for applicable voice/model paths;
exact-head full repo/Nix/Gradle/package gates.

Every concrete bug should gain a deterministic regression fixture before repair when practical.

Forbidden false greens:

xfail/skip used to satisfy acceptance;
timeout/cancel treated as success;
swallowed exceptions or permissive fallback;
fake provider/model reported as live/real;
provider acknowledgement or model assertion reported as verified state;
emulator test reported as real-device acceptance;
CI missing/no checks reported as green.

Android priority

Android usability is a fleet-wide release concern, not only #623's problem. Any change touching shared semantics, voice, providers or plugins must consider Android compatibility where it can affect the client.

The phone bar includes: usable offline composer, deterministic symbolic mode, small-screen/IME behavior, process/activity recreation, permissions, background/foreground, reconnect, stale event fencing, accessibility, resource bounds and installable exact-head APK evidence. Hardware-only claims remain human-gated.

Voice/model priority

The fleet must develop a real low-power model matrix rather than only mocks:

CPU-only and thread-limited profile;
tiny/small real STT or wake path over checked-in WAV fixtures;
sub-1B/similarly light local text model where a Zara path uses model inference;
practical small TTS smoke if research finds a maintained redistributable candidate within the frozen CI artifact/runtime budget;
pinned source/revision/hash/license/quantization metadata;
cold/warm timing, peak RSS, repeated-turn growth and cancellation measurements.

Mocks remain required for edge cases but cannot substitute for the labeled real-model lane.

Merge/collision policy

one focused worktree/branch/PR per implementation slice;
never direct-commit to protected default branches;
before editing a shared hot file/API, check open PR ownership;
avoid simultaneous broad refactors across Core and plugin repos;
cross-repo interface changes require consumer compatibility tests;
#646 should review exact candidate head after substantial updates, not only initial diff.

Worker output minimum

Each tracker retains its detailed WORKER_STATE contract. At minimum every implementation run records target, exact head, tests run/results, blockers/handoffs and next target/mode.

Reconciliation rule

If a worker's old tracker text conflicts with current architecture, update the tracker instead of preserving stale dogma. Do not close unfinished features merely because a newer umbrella epic mentions them.

## Purpose Authoritative coordination contract for the Zara worker fleet. Individual tracker issues own work; this issue owns **stagger, handoff order, collision avoidance and fleet-wide acceptance rules**. ## Fleet roster and stagger Use a repeating two-hour cycle. Preserve relative offsets if the launcher cadence is later changed. | Offset | Worker | Tracker | Primary role | |---:|---|---|---| | +00m | Frontier Research | #645 | discover missing capabilities, current Android/voice/model research | | +15m | Cross-Repo Reviewer | #646 | independent exact-head review across active Zara ecosystem repos | | +30m | Core+Android | #623 | Android usability, offline/runtime integration, Core hardening | | +45m | Voice+Models | zara-plugins#506 | voice stack, real tiny models, low-power CPU CI | | +60m | Symbolics+Dev | zara-plugins#507 | Prolog, coding, memory, expert/sysadmin verified automation | | +75m | Providers+Home | zara-plugins#505 | SmartThings/home/calendar/comms provider integrations | | +90m | Integrations+Registry | zara-plugins#508 | remaining integrations, plugin registry/install/package hardening | | +105m..+120m | quiet window | — | allow CI, rebases and handoffs to settle before next cycle | **Do not launch all workers together.** The stagger is intentional to reduce CI/resource spikes, merge/rebase collisions and duplicated issue selection. ## Handoff semantics - Frontier Research may create/evolve issues but does not broadly implement them. - Cross-Repo Reviewer reviews active candidate heads independently; it does not become the primary feature author. - Implementation workers consume current source, issue state, latest worker comments, #645 research and #646 findings before choosing a slice. - A worker blocked on another owner's contract files a focused handoff issue; it does not bypass the contract or duplicate another runtime/provider. - Current source and exact issue state win over stale roadmap prose. - Existing active PRs are completed/reviewed before a worker abandons them for its new v2 focus. ## Fleet-wide testing law A feature is not complete because its happy path works. Use the strongest relevant combination of: 1. deterministic unit/contract tests; 2. integration tests across actual ownership boundaries; 3. property/generative tests for parsers/state machines/budgets; 4. fuzz/adversarial malformed-input tests; 5. concurrency/cancellation/stale-generation/race tests; 6. restart/crash/fault-injection tests; 7. soak/resource growth tests; 8. visual/UI/lifecycle tests for desktop/Android; 9. real recorded speech + real tiny-model smoke for applicable voice/model paths; 10. exact-head full repo/Nix/Gradle/package gates. Every concrete bug should gain a deterministic regression fixture before repair when practical. Forbidden false greens: - `xfail`/skip used to satisfy acceptance; - timeout/cancel treated as success; - swallowed exceptions or permissive fallback; - fake provider/model reported as live/real; - provider acknowledgement or model assertion reported as verified state; - emulator test reported as real-device acceptance; - CI missing/no checks reported as green. ## Android priority Android usability is a fleet-wide release concern, not only #623's problem. Any change touching shared semantics, voice, providers or plugins must consider Android compatibility where it can affect the client. The phone bar includes: usable offline composer, deterministic symbolic mode, small-screen/IME behavior, process/activity recreation, permissions, background/foreground, reconnect, stale event fencing, accessibility, resource bounds and installable exact-head APK evidence. Hardware-only claims remain human-gated. ## Voice/model priority The fleet must develop a real low-power model matrix rather than only mocks: - CPU-only and thread-limited profile; - tiny/small real STT or wake path over checked-in WAV fixtures; - sub-1B/similarly light local text model where a Zara path uses model inference; - practical small TTS smoke if research finds a maintained redistributable candidate within the frozen CI artifact/runtime budget; - pinned source/revision/hash/license/quantization metadata; - cold/warm timing, peak RSS, repeated-turn growth and cancellation measurements. Mocks remain required for edge cases but cannot substitute for the labeled real-model lane. ## Merge/collision policy - one focused worktree/branch/PR per implementation slice; - never direct-commit to protected default branches; - before editing a shared hot file/API, check open PR ownership; - avoid simultaneous broad refactors across Core and plugin repos; - cross-repo interface changes require consumer compatibility tests; - #646 should review exact candidate head after substantial updates, not only initial diff. ## Worker output minimum Each tracker retains its detailed `WORKER_STATE` contract. At minimum every implementation run records target, exact head, tests run/results, blockers/handoffs and next target/mode. ## Reconciliation rule If a worker's old tracker text conflicts with current architecture, update the tracker instead of preserving stale dogma. Do not close unfinished features merely because a newer umbrella epic mentions them.

nsaspy referenced this issue

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

P0 regression — production zara-server must expose opt-in authenticated CURVE/ZAP listener for Android #173 #355

nsaspy referenced this issue

2026-09-08 02:10:17 +00:00

P2 — implement authenticated Android ZARA/1 client enrollment, reconnect and session continuity #173