Skip to content
Merged
9 changes: 9 additions & 0 deletions AGENTS.md
Original file line number Diff line number Diff line change
@@ -1,5 +1,14 @@
# Forge agent instructions

For current lane allocation, first read
[FOUR_LANE_DEVELOPMENT_V1](docs/roadmap/FOUR_LANE_DEVELOPMENT_V1.md) and
[its machine-readable plan](docs/roadmap/four-lane-development-v1.json).
They supersede only the older two-lane concurrency/default-pool instructions:
LANE_1 owns Forge Platform, LANE_2 EP, LANE_3 Forge and LANE_4 Workspace.
Existing #141 installer and #142 r30 assignments and resource boundaries remain
unchanged. Older two-lane documents remain historical inventory/navigation,
not a competing active allocation. Reading a plan never starts a Work session.

Read [BOOTSTRAP.md](BOOTSTRAP.md), [ENGINEERING_METHOD.md](ENGINEERING_METHOD.md),
[PROMPT_INITIALIZATION.md](PROMPT_INITIALIZATION.md), and the generated
[AI-development projection](docs/ai-development/GENERATED_PROJECTION.md)
Expand Down
77 changes: 77 additions & 0 deletions docs/roadmap/FOUR_LANE_DEVELOPMENT_V1.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,77 @@
# Four product lanes — portfolio allocation and delivery V1

Owner decision: 1 October 2026. Owning transition: [LANE_3 #207](https://github.com/pcvantol/forge/issues/207). **DESIGNED; canonical after protected merge. NO_BUMP.** This changes the temporary operator allocation, not product execution authority or the Forge/EP runtime scheduler.

## One product, one source writer

| Lane | Owning repository | Register | Immediate outcome |
|---|---|---|---|
| LANE_1 / ARCHITECT_1 | autonomous-engineering-system/forge-platform | [#141](https://github.com/pcvantol/forge/issues/141) | Finish the existing first-production installer and unblock its exact qualification handoffs. |
| LANE_2 / ARCHITECT_2 | pcvantol/engineering-platform | [#142](https://github.com/pcvantol/forge/issues/142) | Retain r30 credential qualification unchanged; thereafter select bounded EP efficiency or necessary producer work. |
| LANE_3 / ARCHITECT_3 | pcvantol/forge | [#207](https://github.com/pcvantol/forge/issues/207) | Complete this planning transition; next unblock a real consumer with the smallest qualified Forge slice. |
| LANE_4 / ARCHITECT_4 | pcvantol/workspace | [#208](https://github.com/pcvantol/forge/issues/208) | Build the selected own Server/HTTP/CLI/read-only browser foundation and prove the installed wheel. |

Lane numbers are intentionally not reassigned. The existing LANE_1 installer assignment and LANE_2 r30 keep their IDs, revisions, tests, repair budgets, leases and terminal criteria. LANE_2's exceptional r30 consumer qualification is grandfathered; it does not give that lane Forge/Platform source rights. A waiting assignment is not idle. No second executor or automatic replacement assignment is allowed.

This plan supersedes **scheduling and default repository pools only** in `DUAL_LANE_DEVELOPMENT_V1.md` and `dual-lane-development-v1.json`. Those files remain historical source data and compatibility navigation, not a competing current scheduler. Generic delivery, protections, private-evidence rules and [VERTICAL_SLICE_DELIVERY_V1](VERTICAL_SLICE_DELIVERY_V1.md) continue to apply to all four lanes. The new active entrypoint is this document and [four-lane-development-v1.json](four-lane-development-v1.json).

## The whole portfolio without a second backlog

The machine-readable plan inherits all **34 historical families, 158 exclusive rollup IDs and seven shared IDs** from the existing allocation, and the **31 source references** from the dated inventory. The original 194 named records, 29 rollup relationships and 165 analytical units remain unchanged. These are a historical subset count, not all current work, work remaining, Missions, PRs or hours.

Actual nodes, statuses, hard dependencies, external evidence gates and refinements stay in their owning product DAGs. Resolve their current source and exact evidence when selecting work; follow linked owning graphs. Include later nodes such as installer lifecycle and producer refinements by exact repository ownership even when absent from the September inventory. Unknown owners/edges block selection until reconciled. Do not drop them or label them satisfied. Do not overwrite historical release receipts or imply an old PLANNED parent proves no implementation exists.

| Product | Families retained in its single lane |
|---|---|
| Forge | HTTP/API/CLI; Console/hosting/data administration; inner-loop CI; parallel Actions; project loop/roadmaps; outer-loop CI; governed progression; policy/release; adaptive sizing; hygiene; Genesis/Managed bootstrap; role conversations. |
| EP | Subagent efficiency/assurance; parallel Actions; HTTP/ingress; progression; policy/release; execution envelope; role/task/model policy; hygiene; project bootstrap. |
| Workspace | HTTP/API/CLI; package distribution; project loop/roadmaps; progression; policy/release; hygiene; adaptive sizing; role conversations; project bootstrap. |
| Forge Platform | First-production installer/UX/release, including all later lifecycle refinements; policy/release; progression; hygiene. |

The seven shared nodes keep their owners and are counted once; LANE_3 coordinates their evidence joins, never writes another product. Additional strategy horizons inherit their owning product lane but remain undecomposed/unselected, not silently converted into implementation assignments. This is a federated allocation review, **not a claim that every current edge or implementation in all 31 sources was re-audited**. A global critical-path calculation requires that fresh union, cycle/unresolved-reference checks and actual remaining-duration inputs.

## Execution order: ready work, no four-lane wave barrier

Eligibility precedes priority: an owner-selected bounded outcome, exact qualified prerequisites, real authority, an available repository writer slot and resource/provider budget. Among eligible work, use this order:

1. Finish or unblock an active installer acceptance or safety-critical defect. A required producer handoff inherits the blocked outcome's priority. Deliver the exact test candidate/resources early; do not make native credential qualification wait for full installer acceptance.
2. Deliver the smallest qualified producer outcome that unblocks another active lane. Publish a versioned contract/evidence subset within the same vertical assignment, then continue its implementation/delivery; contract-only is not overall completion.
3. Protect source capacity for the explicitly selected Workspace foundation. Its own server, storage, UI shell and installed-wheel proof do not wait for the installer, a physical reboot, rich Forge Console or complete multi-Action execution.
4. Reduce a measured critical path or repeated AI overhead. EP can select context/isolation/ledger/deterministic-validation slices; Forge can select the parallel frontier or read-only project projection. Follow their existing dependencies, not just their family rank.
5. Select other ready product work by priority, comparable measured effort, ready-age and stable ID. Unknown effort is unknown, not zero. Do not promise a fourfold speedup or infer acceptance from PR count.

For Workspace integration, Forge's narrow `PRM-F-CONTRACT` / `PRM-F-PROJECTION` with the required HTTP gate can precede the whole project-loop family. Workspace can meanwhile deliver useful own behavior and honest UNCONFIGURED/UNAVAILABLE peer states. Real peer views wait for exact authenticated HTTP evidence. No imports, database access, peer CLI, File Inbox or fake live fallback.

The existing concurrency chain remains: Forge `PA-F0` enables EP `PA-E0`; EP `PA-E1/E2/E3` enable Forge `PA-F3`; EP `PA-E4` enables Forge `PA-F5`; EP `PA-EQ` joins Forge `PA-FQ`. Forge continues its independent `PA-F1/F2` work while EP implements its side. Do not replace these fine-grained gates with “finish all Forge, then all EP”, or treat four Codex chats as qualified runtime parallelism.

Workspace public packages/deployment contracts later feed a **separately selected** Forge Platform adapter. Workspace is not added to the current installer DoD. The current installer still requires all previously agreed install/update/preserve/restore/purge, exact multi-instance, provider/pairing, GUI/CLI, signing and live acceptance; it is not declared complete by shipping a download.

## Resource-constrained throughput

Four disjoint source writers may progress concurrently. One repository still has at most one mutating assignment; subagents do not create additional writers. Worktrees separate source changes but do not isolate shared OS state.

Use ordinary unprivileged/hosted CI for normal PR checks. Reserve the exact signing identity, privileged host target, test namespace or reboot window only for its real use. At most one privileged mutation per shared host and one user of a signing identity at a time. CPU-heavy local build/test slots and shared model/token limits must be measured and configured; four lanes are not evidence of enough capacity. Source editing and independent tests need not hold the whole Mac.

Avoid hold-and-wait: do not retain an unrelated source or signer reservation while awaiting an external gate. Uncertain effects retain their exact hold until reconciled. A source boundary may be released only with an explicit clean checkpoint and two-sided transfer where another lane owned it; no timeout or empty PR list grants ownership. All affected lanes acknowledge a shared host window, not merely two named legacy lanes.

Required Quality/Security reviewers may run independently and read-only on the same immutable candidate when capacity and owning policy allow it. A changed SHA invalidates applicable evidence. Use deterministic tools for eligible deterministic validation instead of another LLM reasoning turn; preserve full mandatory role contracts, policy and coverage. Optional specialists need a concrete consumer and expected benefit. Do not lower model/assurance quality without qualified owning policy.

Measure accepted vertical outcomes and closed acceptance gates per elapsed time, blocked time by owner, first-pass acceptance, rework and cost per accepted outcome. PRs/LOC/tokens are secondary diagnostics. This policy is a practical heuristic, not a mathematically proven maximum-throughput schedule. It supplies no new completion-hour forecast.

## Workspace: bounded first vertical slice

[LANE_4 #208](https://github.com/pcvantol/forge/issues/208) selects `L4-WORKSPACE-SERVER-READONLY-V1-20261001`. Reconcile the existing Workspace foundation-only/parking statements in its **own protected entry documentation** before implementation. Only this slice is unparked; the owner has not ordered all future features at once.

Deliver stable own identity/state, explicit safe data roots, authenticated versioned own HTTP services, a thin CLI and a minimal read-only browser surface against the real server. Empty/partial/stale/unconfigured states stay honest. Build a wheel, install it non-editably outside the checkout, and exercise real API/CLI/UI, normal process restart and two-instance isolation. Test/demo data remains explicitly labelled. No new system account, launchd deployment, production credential or shared installer test resource is needed for this first slice.

Map completion narrowly to `WH-CONTRACT/SERVICES/HTTP/CLI` and applicable `WPK-IDENTITY/PACKAGE` subsets, including required Server-identity and role-entrypoint gates. Full `WH-Q`, live peers, public `WPK-PUBLISH/READBACK/CONSUMER`, rich DAG editing, role decisions and installer integration remain open until their separate required proof exists. A package cannot be publicly published before owning its names and configuring the approved publisher.

## Transition and reporting

Before effects, each new executor reads all active registers, verifies local writers/worktrees/locks and acknowledges its exact product scope. The new registries do not prove a local Work session started. LANE_1/2 adopt only the new routing; their active assignments continue unchanged. No simultaneous transition mutates their repositories or test accounts.

Every checkpoint gives the concrete accepted result, exact source/artifacts, next executable acceptance test, blocker and owner, required human action, active resources and same-assignment resume reference. Track `DESIGNED`, `SOURCE_IMPLEMENTED`, `PROTECTED_MERGED`, `IN_RELEASE` and `LIVE_QUALIFIED` separately. An issue comment is coordination, not an atomic lease or automatic event dispatcher.

No Mission-3/reset/T0, production CENTRAL mutation, security-policy weakening, private key export, implied reboot or fabricated independent approval is added. Supported qualified Forge execution should eventually replace direct Work prompts for eligible owner-selected scope, but only when its real installed capability and authority support that scope. This document does not start that cutover or implement a scheduler.

Validation: the new offline guard checks the four disjoint owners/registers, all inherited historical IDs/source references, unchanged active assignments, acyclic **illustrative subset** and no false Workspace-to-installer dependency. Existing full repository validation remains required on the actual PR head. These tests qualify planning structure, not runtime concurrency or installer acceptance.
Loading
Loading