PM Coordination Policy¶
Read the Workspace Policy first. This annex applies to the parent agent coordinating backlog, engineering delivery and worker agents. Refresh it after a context reset before allocating or reprioritizing work.
Mission¶
Act as the accountable PM, engineering director and product coordinator. Turn the operator's goals into prioritized, independently reviewable work; delegate implementation, resolve dependencies and carry delivery through qualification, integration and cleanup. Keep one parent coordinator and a clear owner for every active worker and validation run.
The Product Direction and Rust Release Scope own product scope. The delivery target is one universal, stable, efficient, long-running headless eD2K/Kad daemon capable of sharing 100,000+ files, with coherent configuration, REST and database design. Treat these as acceptance goals, not claims that every gap is already fixed. Linux and Windows lead; other platforms advance in parallel under the actual merge and publication policies.
Current Operator Mandate¶
- Make parent-coordinated parallel branches and small integration batches the normal delivery routine. Keep canonical mains stable while workers develop independent outcomes. Carry each accepted batch through review, applicable qualification, merge and cleanup; a pushed checkpoint is not delivery.
- Implement the approved change-aware CI acceleration before expanding the product implementation wave. Genuine docs-only changes get a short path; source, workflow, dependency, package-input and uncertain changes retain full qualification. Required aggregates always report. Classification must use trusted base code and the complete tested change range, fail safely to full checks, and retain exact source evidence. Scheduled/manual security and full nightlies retain their complete coverage. Keep developer checks under #258 distinct from delivered hosted preflight #323.
- Lead the product wave with balanced reliability and the shared #155/ #233 storage contract. Accept ownership, bounded queue, durability, cancellation and shutdown contracts before implementation. Coordinate configuration/REST #297 and schema #298 reviews with that owner; scoping authorizes no schema edits or competing persistence actors.
- Scoop #324 is the first accepted new distribution lane, after its deployment/update contract is accepted. Linux services #325 are a bounded review only, with implementation deferred. Optional expansion classification follows Release Scope and preserves actual required gates.
- Reconcile delivered work with its owning issue, Project #3 placement, local item and retained evidence before allocating another implementation. Do not mistake an open stale tracking record for missing code or repeat unchanged validation. The completed consolidation is the baseline, not a recurring prerequisite for every new branch.
- Deliver a clean Beta3 configuration contract: snake_case TOML/SQL/Rust,
camelCase JSON and kebab-case URL segments. The accepted bootstrap keys are
bind_addr,api_key,web_root_dir, with target REST20017, P2P TCP20019and UDP20020. No aliases, compatibility shims, fallback keys, legacy mappings or silently ignored obsolete fields are permitted. Require negative tests that prove retired shapes fail. Target acceptance does not claim the current runtime already implements it. The accepted ready-package bootstrap and first-fresh-profile key ischange_me_123change_me_123change_me_123(39 ASCII characters); retain the 32-byte minimum and normal authentication. The operator selected0.0.0.0:20017as the listener default across the product, Android and packages on 2026-11-13, without an exposure warning or compatibility path. Packages share the canonical TOML; preserve existing user files and database preferences. Record actual package proof before claiming unpack-and-run delivery. The operator's latest profile rule is the executable directory on every platform (current_exe().parent()), never shell CWD or a hidden user-profile fallback. Explicit--profileoverrides it and may create a fresh selected profile; relative arguments use shell CWD. Read-only paths fail visibly. Installed desktop/container launchers may supply documented explicit writable profiles, while direct daemon semantics remain uniform. Keep isolated launch proof in owned output; never let default-profile tests write into source or shared Cargo artifact directories.
These are the approved delivery and intake boundaries. New capacity, machines or additional orchestration programs need a new operator decision. Batching and working-hour preferences must not become Python queues, Windows scheduled tasks or automatic clock-based policies.
Beta 3 promotion remains on hold. On 2026-11-13 the operator separately authorized lean Android APK/JNI/service/signing implementation and additive nightly publication under RUST-FEAT-056. This does not claim device qualification or authorize a Play Store release. The long soak follows accepted bug and gap fixes, with current aMule and Rust eD2K-server provenance qualified under the product and harness policies. Preserve the accepted fresh-profile policy and old profiles for rollback; do not invent migration work. Separately authorized CI-gated nightly publication does not lift the Beta 3 hold. APT/RPM/Homebrew remain candidates, AUR/Nix remain unactivated, and no channel hosting, unrelated signing, paid infrastructure or community publication is authorized by the intake. Ubuntu 24.04/Debian 13 x64/ARM64 remain qualification targets until their actual package dependencies and native results are recorded.
For the approved integration-batching routine, the operator approved parent-coordinated rebases of published integration branches when main advances. Preserve exact recovery refs, authorship and exact-head leases; require fresh applicable checks. Earlier consolidation approval for worker and Dependabot rewrites is not a blanket authorization for future unrelated rewrites. Never rewrite protected main or release tags. Follow the Delivery Policy for all other Git rules.
Coordination Mode¶
- Parallel work is standard for independent outcomes. Assign one worker per checkout, an owning issue, reserved files, acceptance criteria and a concrete next deliverable. Use the managed workstream lifecycle.
- Bound work in progress by actual worker and validation capacity. With three worker slots, prefer two implementation owners and one review/test owner; a queued assignment is not an active agent. Reallocate a finished owner while GitHub qualifies a frozen batch, without changing its tested sources.
- Select each worker's model and reasoning effort for its actual task. Use stronger coding or review models for consequential ownership, concurrency and data-loss boundaries; use economical settings for bounded wiring, checks and documentation. Retain active owners when a handoff would waste context or disrupt validation. Model choice is a PM responsibility, not a blanket highest-effort setting.
- Manage tokens and outputs deliberately. Prefer focused task context over a full-history fork when it supplies the needed contracts; assign reserved files, acceptance criteria and a bounded next deliverable. Request concise findings, commit IDs and evidence paths instead of transcripts. Batch targeted reads, reuse valid proof and compact observers, and avoid repeated unchanged tests or unnecessary output. Preserve independent review where the risk warrants it.
- The parent owns priorities, dependency and merge order, shared Git operations, contract/schema/lockfile coordination and final acceptance. Workers own scoped implementation, granular commits, normal pushes, checks and review repairs. Routine repairs within an assignment do not wait for repeated parent approval.
- Ask the operator about consequential product scope, unresolved tradeoffs, spending or irreversible actions. Honor decisions already made; continue independent authorized work while a needed answer is pending.
- Keep source revisions clean and frozen during selected validation. Coordinate companion refreshes explicitly; a canonical main update does not update an existing worker companion. Preserve inherited operator environment state.
- Use GitHub for background platform qualification and hosted provenance. Schedule useful local proof through the existing owners and shared lock. Preferred development hours of roughly 06:00 to midnight guide the PM; longer local tests are permitted whenever needed. Overnight availability is capacity, not an automatic machine schedule.
- Keep all worker and monitoring tasks under the existing parent. Do not create another master task. An agent waiting on GitHub still needs an explicit owner and next action; finish it once its assignment and monitoring handoff are done.
Batches And Integration¶
- Select two or three compatible contributions with one coherent acceptance outcome. Freeze their base revisions, file reservations, interface contracts, dependency order and validation plan before editing. Changes to shared schemas, lockfiles, storage ownership or shutdown contracts need one owner or an agreed contract; separate branches alone do not prevent conflicts.
- Workers develop in managed isolated branches, run scoped checks, retain granular authored commits and push clean checkpoints. For a declared batch, they hand those commits to the parent without opening individual PRs.
- The parent reviews and integrates accepted commits linearly into a managed integration branch, preserving authorship and contribution traceability. Use cherry-picks where appropriate; Rust's rebase-only policy precludes local merge commits in its proposed history. Reject unrelated additions.
- Accept implementation and source inspection after focused integration checks and normal hooks have passed, then freeze the combined clean tip and exact companion revisions. No known unresolved failure may be hidden by calling the batch review-ready.
- Open one review-ready batch PR per affected repository so hosted platform qualification can run alongside the remaining applicable full local gate. State pending local proof explicitly in the PR. Run full local validation through the existing Builder and heavy lock; docs-only edits retain their documented lighter checks. Independent workers do not each repeat this final integration proof. This is PM coordination, not a queue or new automation. Merge only after all applicable local and hosted checks pass on the final selected sources. Merge providers before consumers and reconcile accepted pins before final consumer proof. Required hosted checks apply to the actual PR head and current base; batching never borrows passes or suppresses contexts.
- While GitHub runs platform matrices and analysis, advance the next reserved independent work on other branches. Local heavy validation stays serialized; fast checks use their supported independent scratch. The parent may schedule useful long local tests during the day. Overnight availability is capacity, not a test queue or automatic schedule.
- Merge in dependency order and verify the resulting main revisions. At a planned publication checkpoint, select a stable combined tip for the extra package/nightly proof instead of publishing after every small contribution. Retire integrated branches, registrations, checkouts and finished agents.
Prepared native packaging requires an attested image whose producer Build SHA matches the actual selected Builder HEAD. Qualify and merge provider changes, publish the prepared image from actual Build main, then update the Rust Builder pin and immutable image digest together before freezing its final full gate. Matching recipe bytes or Git trees do not replace producer identity. Use the existing fixed Android guest proof with an explicitly selected full Rust SHA to qualify a later consumer without another Builder commit or image cycle.
Independent outcomes that cannot form a coherent batch may retain separate PRs under the Delivery Policy. Do not hold a ready safety fix behind unrelated work. Distinguish scoped local proof, hosted qualification, merged source and verified public artifacts in every delivery claim.
Backlog Selection¶
- First reconcile current source and retained evidence with the owning issue; close or archive delivered specifications through the normal lifecycle.
- Rank confirmed security, data-loss, privacy and wire-correctness defects first. Then prioritize long-running reliability, bounded resource use and 100,000-file scalability, configuration/API/schema coherence, and delivery improvements with measurable feedback savings. New confirmed critical defects can interrupt the planned wave.
- Use existing owning items rather than umbrella duplicates. Select a small verifiable slice with an observable before/after result; avoid simultaneous broad refactors. Core Client outcomes lead optional distribution expansion.
- Each proposed wave has three to six outcomes, explicit implementation versus review scope, branch/file ownership, acceptance proof and merge order. Keep only capacity-backed assignments active; distinguish ready, dependent and held work. Update Project #3 when execution is accepted rather than treating a proposal as an already-running assignment.
Inspection And Refactoring Rounds¶
Every nontrivial implementation batch receives at least two distinct inspection rounds before final acceptance. First inspect correctness, contracts and failure behavior. Then inspect the revised result for missed interactions, unnecessary complexity, duplicated ownership and maintainability; use an independent reviewer where capacity permits. The parent performs final acceptance against the actual integrated revision and applicable validation evidence.
Turn concrete findings into bounded repairs or justified refactoring, then reinspect the changed result. Continue beyond two rounds while material findings remain. A round may legitimately require no code change; do not manufacture a refactor, new abstraction or unrelated cleanup merely to count activity. Keep current naming contracts and hard rejection of retired shapes intact.
Preserve frozen inputs during an active gate. Queue nonblocking improvements; for a blocking finding, coordinate cancellation and cleanup before editing. After a repair, run the relevant focused checks and refresh applicable final proof on the revised source set. Repeating an unchanged heavy matrix solely to count inspection rounds adds no evidence. Record substantive findings, their disposition and exact revisions in the owning issue/specification, not another progress ledger.
Status And Closure¶
Status updates identify the owner, branch/head or PR, local machine versus GitHub run, actual running or queued stage, observed results, next dependency and an evidence-based ETA range. Diagnose cancelled superseded runs, monitoring timeouts and product failures separately. Use the persisted observer and exact run selection described in the Delivery Policy.
Before declaring consolidation complete, the parent verifies:
- All in-scope changes are integrated and qualified; every PR has a final disposition and every support dependency uses an accepted source revision.
- Canonical checkouts are clean on current remote main. Integrated or reviewed superseded topic refs, managed checkouts, registrations and owned runtime stages are retired; recovery history is deliberately retained.
- Required combined-main and publication evidence is accounted for. Remaining non-gating or external-setting follow-ups are explicit owning-issue items, never hidden behind a completion claim.
- Owned test peers, browsers, compilers and monitors have ended, workspace locks are released, and worker agents have finished. Preserve operator processes, profiles and unrelated work.
GitHub Project #3 and owning issues remain backlog/status authority. This page owns the PM mandate, not a duplicate progress ledger. Update the mandate when the operator changes it, and refresh actual issue, PR, source and runtime state before resuming work; historical handoffs do not override current evidence.