Inner Loop

We built a plan CLI with 40+ commands and 114 plans. The CLI says 74% complete. The UI said 1%. The boss still asks "are we on track?" — and we can't answer without a terminal.

These six projects close that gap. They are not the product. They are the machine that builds the product.

Pain

Agent sessions scatter results. The 'why' dies in Slack. North Stars exist as prose — none queryable. Plans dashboard math is wrong.

Friction

Seven skills span two repos, none chain. Every step needs a human trigger. Memory exists in DB — nothing reads it back into context.

Desire

One trigger runs the loop. Nav profile primes from DB. Decision traces persist automatically. Feature states compute from tests.

Potential

Morning diff in 30 seconds. Overnight loops produce proven demand. Commissioners trust the matrix. Factory measures its own output.

P&ID Role Map

The Control System

Six roles. One loop. Each feeds the next.

CONDUCTOR

Autoresearch Loop

Chains skills into overnight autonomous loops

L0prd_full_cycles_per_month ≥ 1

↓

PUMP

Agent Platform

Identity, memory, comms, dispatch — moves the work

L2agent_first_task_seconds < 30

↓

INSTRUMENT

CLI Platform

drmg binary — one control surface, 60+ commands

L4direct_db_imports_in_scripts = 0

↓

CONVEYOR

WorkChart Bridge

Blueprints become running plans — dream to eng pipeline

L0workchart_stage_automation_rate > 0.6

↓

GAUGE

Decision Tracer

Durable traces of the 'why' behind every change

L0untraced_policy_exceptions = 0

↓

CONTROLLER

Automated Commissioning

Test results compute feature states — closes the loop

L0false_L3_count = 0

Ship Now

≥ 3.02 PRDs

Project Management System

SYSTEM

L2-partial

Track state of things across cadences so we can plan days, weeks, months and know if we're on track.

3/5

Score

can_answer_are_we_on_track = trueCLI works (74%), UI broken (1%). 7 dogfood issues.

Deck

CLI Platform

INSTRUMENT

One binary (drmg) with shared platform lib, repository-only DB access, agent-grade contracts.

3/5

Score

direct_db_imports_in_scripts = 0Completed. 9 namespaces, 60+ commands, unified binary.

Prepare

2.0–2.94 PRDs

WorkChart Bridge

CONVEYOR

Bridge dream repo work charts to engineering execution engine — blueprints become running code.

2.7/5

Score

workchart_stage_automation_rate > 0.6PRD written. 9 work charts, 9 workflows, zero connection.

Agent Platform

PUMP

Identity, memory, scaffold generators, boundary enforcement — the PUMP that powers the factory.

2.6/5

Score

agent_first_task_seconds < 30v2.2: 5 features (PUMP-001-005). 14 absorbed. Plans UI, scaffolds, boundaries, patterns, memory.

Decision Tracer

GAUGE

Durable traces of the 'why' behind every priority change.

2.5/5

Score

untraced_policy_exceptions = 0Spec drafted, ready for engineering.

DeckSpecMaps

Autoresearch Loop

CONDUCTOR

Chain skills into overnight autonomous loops.

2.1/5

Score

prd_full_cycles_per_month >= 17 skills span 2 repos, none chain.

Incubate

1.0–1.91 PRD

Automated Commissioning

CONTROLLER

Test-driven feature matrix — states computed from test results, not hand-edited.

1.6/5

Score

false_L3_count = 0Spec drafted. URGENT: 2 false-L3s found (stub, mocks).

What We Need

Six CLI commands. Each one closes a gap in the loop. Build in order — earlier commands unblock the later ones.

PRD	Command	Success When
Agent Platform (PUMP-005)Register nav in AGENT_PATHS + create semantic/procedural memory JSON files	drmg etl load --agent=nav	Nav profile upserts to DB. drmg etl status --agent=nav shows profile age.
Agent Platform (PUMP-005)Read profile + memories from DB, write nav-profile.md for context-aggregator	drmg agents context --agent=nav	Nav session startup includes DB-backed context. daily-snapshot.md has a Nav profile section.
Autoresearch Loop (LOOP-002)Return North Star metric as scalar (e.g. won_deals/total = 0.28)	drmg measure --prd=<slug>	Every active PRD's North Star is queryable. Returns in <2s with null + reason if unavailable.
WorkChart Bridge (CONVEYOR)Move dream repo work chart stage to plan CLI task automatically	drmg workchart advance --prd=<slug> --stage=<n>	One work chart stage advances without manual CLI call. Plan task created and linked.
Auto Commissioning (CONTROLLER)Compute L-level from test results, detect false-L3s, write receipt	drmg commission --prd=<slug>	false_L3_count = 0 verified by CI. Matches manual assessment on AUTH-* features.
Decision Tracer (GAUGE)Write durable decision trace to IDecisionTracePort	drmg trace log --context='<why>' --ref=<prd>	Untraced policy exceptions = 0. Traces appear in context-graph reality edges automatically.

Build Sequence

Sprint order = pipe sequence. Earlier sprints lay tracks the later ones run on.

Sprint	What Ships	Feeds
N1	Plans UI — math, clickable rows, project grouping	Done
N2	Plan detail page — phases, tasks, progress, evidence	In progress
N3	Project dashboard — things by state across all projects	Next
N4	Cadence views — daily/weekly/monthly planning surfaces	Time+Mind
N5	Computed states — test results update feature matrix	Auto Commissioning
N6	Decision traces — 'why' persisted alongside 'what'	Decision Tracer

Context

Navigation System — Value, Belief, Control
Control System — PID mechanics, pit of success
Progress — Things by state, not tasks by done
Planning — Cadences: daily, weekly, monthly, quarterly
Business Platform Requirements — The external-facing channel
Commissioning Dashboard — L0→L4 states across all features

Questions

If the factory can't measure its own output, how does it know what to build next?

When does the Plans UI replace the priorities markdown page as the primary decision surface?
Which inner loop gap costs the most engineering time each week — measured in repeated manual steps?
If the CONDUCTOR ran overnight, what would you find by morning — and what would be missing?
What's the minimum loop that closes before you can trust the factory to run unsupervised?