docs: reset authority to current Career OS product model

[avabbbb]

Goal

Remove documentation authority drift and make the repository understandable to both humans and Coding Agents.

The repository had accumulated several generations of mutually conflicting design:

• built-in Main Agent / Pi;
• DSH/Codex-first Harness shell;
• external-Harness-only reasoning;
• current OfferU Desktop + external-local-Agent-first + fallback + Guided Career OS.

The problem was not that the old documents were historically false. The problem was that many of them were still presented as current/accepted authority.

New authority model

Default reading path is now intentionally small:

1. GOAL.md
2. docs/product/current-product.md
3. CONTEXT.md
4. ARCHITECTURE.md
5. STATUS.md
6. HANDOFF.md

Topic docs are loaded only when needed.

Current product model

The current documents now consistently define:

• OfferU Desktop as the primary Career OS experience;
• external local Agents as the preferred reasoning layer;
• OfferU built-in Agent as an allowed fallback;
• exactly one reasoning authority per Run;
• composable third-party Career Skills;
• Operation Registry as execution/permission authority;
• Career Runtime as truth authority;
• Guided Today / Next Best Action as beginner interaction;
• Zero-Setup + Zero-Figure-It-Out as product goals;
• user-triggered browser job capture;
• email/signal → candidate → reviewed stage event;
• no silent final application submit.

ADR simplification

No new ADR number is introduced.

The old 58-entry ADR ledger is preserved under:

docs/archive/architecture/adr-history-pre-2026-09-23.md

The live docs/adr/README.md is now a short current decision summary. Future work should not create numbered ADRs for every implementation adjustment.

Historical preservation

Old DSH/Harness/Pi-era documents are preserved under docs/archive/, including:

• historical ADR ledger;
• previous CONTEXT glossary;
• DSH-era agent system / workbench / integrations;
• old Bridge v1 target;
• old control-plane target;
• Role Intelligence v0 design;
• Sep 2026 product critique;
• previous STATUS / HANDOFF snapshots.

Dated operation/eval documents that remain in place are explicitly marked HISTORICAL SNAPSHOT.

Current docs refreshed

• README / README_ZH
• GOAL
• CONTEXT
• ARCHITECTURE
• STATUS
• HANDOFF
• docs/README
• agent-system
• workbench-interaction
• harness-integrations
• agent-control-plane
• agent-bridge-protocol
• agent-runtime README
• Tool Surface V2 status
• conformance stale link
• old Real-Agent-E2E guide status

Why this is intentionally smaller

Agent-facing documentation benefits from progressive disclosure: current high-value context should be easy to discover, while historical rationale remains available only when needed.

References:

• https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents
• https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills

Scope

Docs only. No production code, schema, runtime, Operation, permission or Eval behavior changes.

Follow-up

PR #18's Zero-Setup product proposal has been incorporated into docs/product/current-product.md. After this merges, #18 should be closed to avoid duplicate product authority.