Speedrift Ecosystem

From repo toolchain to ecosystem operating fabric.

Speedrift now combines repo-local autonomy with a daemonized control plane. The result is continuous multi-repo awareness, bounded autonomous execution, and explicit writeback into local workgraphs.

Workgraph spine Control + repo + worker planes Central register + websocket Bounded safe handlers Claude/Codex handoff prompts
Mental Model Shift

Speedrift fundamentally changed.

The old frame was "run drift lanes in one repo." The new frame is "operate a supervised program across many repos while preserving local agent autonomy and safety boundaries."

Before
repo-local checks, manual portfolio supervision
Now
daemonized control plane + bounded automation loop
Target
autonomous dark-factory workflow with full trail
Old Model run checks per repo manual status polling ad hoc remediation New Model observe all repos continuously plan + execute safe handlers write back corrective tasks
Operating Planes

Repo plane, control plane, worker plane.

Repos own their local graph and execution context. The central daemon supervises, prioritizes, and runs bounded automation. Workers (Claude/Codex/session-driver) execute tasks and push results into local graph state.

local `.workgraph` source of truth central mirror register websocket status stream ledger per cycle
Control Plane ecosystem daemon + factory cycle + policy gates Repo Plane A local graph + hooks + followups Repo Plane B deps + tests + loopbacks Repo Plane C service + registry writes Central Register snapshot + ledgers + queue state Worker Plane Claude/Codex/session-driver execution workers and central decisions write back into repo graphs
Authority Contract

Central automation is explicit and bounded.

Central can

restart services, run deterministic safe handlers, emit corrective tasks, publish dashboard/api/ws updates.

Central cannot

do destructive history rewrites, silently mutate unrelated work, or bypass required verification gates.

plan_only=true: observe mode plan_only=false: execute mode policy budgets + hard stops
Policy Envelope safe handlers only Can bounded execution Cannot unsafe mutation Budget + cooldown Decision ledger
Autonomous Loop

One daemon cycle supervises and advances the whole ecosystem.

1. Observecollect snapshots + service health
2. Prioritizescore stalls, dependency pressure, risk
3. Planselect actions under policy budgets
4. Executerun safe handlers and dispatch workers
5. Recordledger + register + websocket + followups
Observe repo state + signals Prioritize risk and urgency Plan bounded action set Execute handlers + sessiondriver Record ledger + writeback tasks loopback is explicit: create/adjust tasks in local graphs
Dashboard Contract

One operator view, ordered for decisions.

Page order

narrated overview -> operational overview -> repo cards -> graph views -> attention queues.

Interaction contract

click repo card to focus graph; keep all-graph and inter-repo graph views available at all times.

Action format

queue entries provide ready Claude/Codex prompts with why/priority/next-step context.

1) Narrated overview 2) Operational overview 3) Repo cards active/stalled + watch reason priority, dirtiness, risk filters click to focus graph 4) Graph views repo graph + inter-repo graph active nodes/paths highlighted zoom + pan + dependency clarity 5) Attention queues aging gaps, dependency breaks, security/quality, upstream candidates each item includes a runnable Claude/Codex prompt
Repo Awareness

Make stalls explainable and dependencies actionable.

Repos with no activity should state why: service down, blocked dependency, aging in-progress work, or repeated failed loopbacks. This turns passive monitoring into active queue generation.

At-Risk
service stopped while ready work exists
Watch
aging in-progress and dependency uncertainty
Healthy
fresh heartbeat, active progress, passing gates
speedrift-ecosystem driftdriver coredrift meridian paia-os upstream queue edges define semantics; orientation is just layout
Module Stack

Model-mediated modules under one policy envelope.

Core drift lanes enforce repo integrity. Ecosystem modules supervise portfolio execution, quality, security, plan integrity, and cross-repo continuity.

Core Drift Lanes

coredrift, specdrift, datadrift, archdrift, depsdrift keep code/spec/intent/dependency aligned.

Complexity + UX

uxdrift, therapydrift, yagnidrift, redrift catch UX, stability, overbuild, and regression pressure.

Factory Control

factorydrift owns per-cycle action planning and bounded execution management.

Worker Dispatch

sessiondriver uses claude-session-driver to claim and execute ready tasks safely.

Program-Level Auditors

secdrift, qadrift, plandrift emit remediation tasks into local graphs for local agents.

Impact Mechanics

Mechanics only: show formulas, not production numbers.

Metrics are presented as explanatory computation rules with illustrative values. Operators should read trend + tier + next action prompt, not raw one-cycle counts.

coverage flow health execution reliability quality/security pressure
18/20
reporting coverage (illustrative)
24/28
factory success ratio (illustrative)
6/10
plan-integrity coverage (illustrative)
Coverage % reporting = repos reporting / registry repos example: 18 / 20 Flow Health stall_rate + dependency_gap_rate + aging_pressure example tier: watch Execution Reliability factory_success = succeeded_actions / attempted_actions example: 24 / 28 Quality + Security weighted findings + plan_integrity_coverage example: 6 / 10 always attach trend + tier + Claude/Codex prompt
Operations

Run continuously, act deliberately, keep local autonomy.

Start in observe mode, then move to execute mode once queues and policies are stable. Keep the daemon running so portfolio supervision stays live across repos.

Daemon Endpoint

Codified default: dashboard at :8777 and websocket at /ws/status (local + Tailscale).

Observe Mode

plan_only=true for scoring, queues, prompts, and ledgers without auto execution.

Execute Mode

plan_only=false for bounded handler execution and session-driver dispatch.

Local Ownership

Central emits tasks; local agents execute and verify in repo context with full audit trail.

Upstream Leverage

Use candidate queues to package and propose improvements back to external parent repos.

Speedrift Ecosystem Story Deck
1 / 1 1 2 3 4 5 6 7 8 9 10 ← → Space