stack/docs/plans/ROADMAP.md

ROADMAP — agreed path forward

Status: planning aligned with owner 2026-09-03. Feature work follows this order unless the owner re-orders. Design notes below are commitments about intent, not implementation — each milestone still gets its own issue, tests, and BUILD-LOG phase.

Standing decisions (owner, 2026-09-03)

• Pi is the main TUI harness and the reference checkpoint. Multi-harness (Codex, Claude, OpenCode) comes after pi is stable as the primary seat.
• CI is deferred. Gitea runs on slow hardware; local suites + verify are the quality gate. Revisit when code volume justifies runners.
• The user never manually runs release.sh. The system determines what is installed and self-aligns (M16).
• Tooling documentation is agent-accessible via the ms-tools skill (M17), referenced from AGENTS.md.
• Capability restriction becomes progressive and role-based (M18), porting prior Mosaic Stack concepts where they fit the file-based runtime.

M16 — Release self-determination

Problem: after reset.sh the activation pointer is gone and the owner had to know about scripts/release.sh activate. The system must determine what is actually installed.

Design:

• New release.sh ensure (also invoked automatically by the launchers via common.sh): compares desired (RELEASE) vs packaged image vs active pointer.
• Fast path (drift = pointer missing/mismatched, image already packaged): restore the pointer without a health gate — the release was previously health-gated; reset is data loss, not a new release.
• Slow path (new version in RELEASE, or image missing): package beside → full health gate (exact-marker model run) → atomic pointer swap. This is the M3 gate-then-flip, now automatic.
• Acceptance: post-reset.sh, the next hello.sh/agent.sh self-restores activation with zero manual release commands; a bumped RELEASE auto- activates after its health gate; a failing health gate leaves the previous release active and records the refusal.

M17 — ms-tools skill (central tooling documentation)

Problem: tooling knowledge lives in docs/TOOLS.md + README + this file's history — agents must be told where to look every session.

Design:

• A pi skill: skills/ms-tools/SKILL.md (frontmatter: name, description; body: the tool surface — scripts, exit codes, safety notes, pointers into docs/TOOLS.md for depth). Skills load on demand, so the always-in-context cost stays in AGENTS.md (one pointer line).
• AGENTS.md gains: "tooling reference: invoke the ms-tools skill."
• Workers deliberately run --no-skills — the skill serves conductor-level and vanilla sessions, same boundary as AGENTS.md.
• docs/TOOLS.md remains the source of truth; the skill references it rather than duplicating content.

M18 — Seat-role progressive capability restriction

Problem: capability restriction exists but is per-task/per-mission only. Seats (agent identities) need governing ceilings that tasks cannot escalate past — the M9 intersection chain extended: role ceiling ∩ mission grant ∩ task grant = effective.

Design:

• roles/<role>.json (role contracts, tracked): roleVersion, name, tools ceiling (subset of pi built-ins), network: "none"|"api-only"| "open" (declared now, enforced when network policy lands), paths (future).
• Resolution chain in the runner: seat role contract ∩ mission capabilities ∩ task capabilities → effective tools (existing intersection logic, one more link). Empty at any stage = tool-free run, loudly recorded.
• agent.json keeps role (already shipped in M15) — it now binds to roles/<role>.json.
• Porting note: prior Mosaic Stack role/policy concepts inform this design, but implementation is fresh on the file-based runtime (the standing repository exclusion applies until the owner points at specific prior files to port).
• Adapters unchanged: MOSAIC_TOOLS already carries the effective set.

M19 — Harness auth tooling (pi checkpoint)

Problem: auth is manual (/login, auth.json); multi-account and multi-harness stacking have no tooling.

Current facts (observed): one auth.json holds multiple providers (anthropic, openai-codex, zai observed on the host) — provider stacking already works. Multiple accounts per provider is the open question (the owner runs two Claude accounts; openai-codex multi-account desired).

Design direction (pi-first checkpoint):

• scripts/auth.sh status — per-provider auth report (which providers have credentials, source: auth.json vs env), no secrets printed.
• Multi-account: investigate pi's auth schema for per-provider account entries; if unsupported natively, model accounts as named auth files (auth.<account>.json) with a conductor-level selector — runtime-injected per launch, same pattern as the read-only credential mount.
• Claude/Codex/OpenCode adapters stay parked; the auth tooling is harness-neutral so it serves them when they arrive.

Explicitly deferred

• CI runners (owner: Gitea hardware is slow; local suites are the gate) — revisit as code volume grows.
• Network egress policy enforcement (declared in role contracts at M18; enforcement is its own milestone).
• Push automation (push stays an explicit act).
• Second real adapter; fleet/multi-agent; knowledge storage; portal.

Completed reference (see docs/plans/CURRENT.md log)

M1 config-driven hello · M2 missions/tasks · M3 release model · M4 adapter seam · M5 workspaces/capabilities · M6 sessions · M7 ergonomics · M8 conductor loop · M9 mission capability policy · M10 retention · M11 session forking · M12 auto-apply policy · M13 interactive TUI agent · M14 live user context · M15 agent seats + roles/ convention.