docs(research): preserve agent intelligence and learning archive

[TheHalfMoon]

Purpose

Preserve Winds strategic research in GitHub so high-value donor/research work is not lost while keeping it explicitly outside active Spec 003 implementation scope.

Adds

• docs/research/007-agent-intelligence-memory-learning-archive-2026-08-16.md
• docs/research/008-nicobailon-pi-extension-ecosystem-audit-2026-08-17.md
• founder product direction for Winds as an evidence-first control plane above heterogeneous agent CLIs
• Verified Experience / memory / predictive-execution research
• donor/reference notes for Goose, Kilo Code, Zed, Kiro, delegate-skills, Pi, and related agent systems
• comprehensive Nico Bailon Pi-extension ecosystem audit covering delegation, interactive CLI execution, coordination, memory, model/skill policy, context lifecycle, human approval UX, observability, and provenance/license cautions

Explicit non-authorization

This is a research archive only. It does not amend the Winds Constitution, Spec 003, plan, or tasks; does not authorize Agent Fleet/MCP/ACP/public protocol/daemon/memory/model-training code; and does not change the active Spec 003 implementation scope.

Kept as Draft so the research is safely persisted on GitHub without pretending it has passed the Winds implementation acceptance/review gates.

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: f92ddf8d-9b93-4197-b4ab-05b0603d6bbe

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

_{Comment @coderabbitai help to get the list of available commands.}

[TheHalfMoon]

HISTORICAL_RESEARCH_ARCHIVE_CLASSIFICATION_2026_09_01

This Draft PR remains intentionally preserved as historical research input, not a landing candidate.

Current repository truth has advanced substantially since this archive was created:

• Spec 003 is canonically closed;
• Spec 006 is active and currently blocked at T079's mandatory real bounded Codex App Server proof;
• the Herdr parity-and-beyond roadmap is now canonical research-only guidance via merged PR docs(research): land Herdr parity and beyond roadmap #85;
• the newer Verified Learning Loop roadmap is undergoing current-base landing qualification in PR docs(research): land verified learning loop roadmap #87 and explicitly requires post-Spec-006 reconciliation of this historical archive before any formal learning program.

PR #21 still contains unique historical donor/research material, including the agent-intelligence archive and Nico Bailon Pi-extension ecosystem audit. Those files are not on canonical main, so closing/deleting this archive merely for backlog tidiness would reduce discoverability of preserved research input.

Current disposition:

PR_21=OPEN_DRAFT_HISTORICAL_RESEARCH_ARCHIVE
MERGE_UNCHANGED=NO
IMPLEMENTATION_AUTHORITY=NO
SPEC_006_AUTHORITY=NO
T079_T080_AUTHORITY_CHANGE=NO
FUTURE_ACTION=RECONCILE_AFTER_SPEC_006_BEFORE_ANY_FORMAL_LEARNING_PROGRAM

Do not treat the old Spec 003/T049 status text inside the archived documents as current repository-state truth; it is historical checkpoint context from the archive date. Live canonical governance remains authoritative.

Leaving this PR open as Draft is intentional archival preservation, not unfinished implementation work.

[TheHalfMoon]

MUNDER_DIFFLIN_DONOR_RESEARCH_NOTE

Source reviewed:

repository=chaitanyagiri/munder-difflin
revision=d27303e8a4cc86a98cfef408bb7e7b1fadb5ccad
tree=1a716287ebc57da16e012e5b674ec507f391d698
observed_version=0.4.6
source_code_license=MIT
classification=RESEARCH_DONOR_ONLY
implementation_authority=NO
T079_PROOF_SUBSTITUTE=NO

Useful patterns for future Winds reconciliation:

1. Terminal/event-plane separation. Munder Difflin treats raw PTY bytes and structured lifecycle events as separate data planes. This is directionally aligned with Winds' requirement to avoid inferring authoritative runtime state from terminal prose alone.
2. Concrete cross-platform PTY/process hardening. src/main/pty.ts, src/main/shellEnv.ts, src/main/ptyEnv.ts, and src/main/procKill.ts contain useful implementation lessons around login-shell PATH discovery, command-name validation, Windows npm .cmd shim handling, environment identity stripping, and process-tree cleanup.
3. Provider-specific integration before generic abstraction. src/shared/agentProvider.ts carries concrete provider behavior and capability differences instead of forcing one universal protocol. That is compatible with Winds' current Codex/Claude-first, no-generic-plugin-framework direction.
4. Single-committer coordination. The hive design lets agents write scoped files while one trusted main process performs Git commits, reducing concurrent Git/index corruption. This is a useful donor pattern for future multi-agent coordination research, not current Spec 006 authority.
5. Mailbox/blackboard/append-only event patterns. Per-agent inbox/outbox, idempotent cursors, hop caps, and an append-only event log are useful research inputs for post-Spec-006 coordination/verified-experience design.
6. Preserve sandbox rather than reflexively disable it. The Codex provider moved away from --dangerously-bypass-approvals-and-sandbox toward an explicit sandbox posture plus --add-dir for a narrowly required writable path. The general lesson is useful even though Winds must independently revalidate current Codex semantics and its own stronger authority contract.

Important non-donor / incompatible behavior:

• Munder Difflin's current Codex path is primarily terminal/PTY CLI integration with positional prompt injection and hook/config bridging, not the official Codex App Server structured-control path selected by Winds Spec 006.
• It may auto-install missing agent CLIs. Winds T079 explicitly forbids manufacturing prerequisites, credential/terms escalation, or installing Codex merely to create proof evidence.
• Its Codex hook bridge creates a per-agent CODEX_HOME and links/copies existing Codex auth/config state. That is NOT acceptable as a way to satisfy T079 in the current governed environment and must not be copied into the proof path.
• The provider file itself contains at least one TODO-verify Codex model-id note, so repository claims must be treated as donor observations, not vendor authority.
• Munder's GOD-agent autonomy, native approval assumptions, and terminal-driven orchestration are materially broader than the currently authorized Winds T079 slice.

T079 disposition:

MUNDER_DIFFLIN_UNBLOCKS_T079=NO
MUNDER_DIFFLIN_CHANGES_T079_CANDIDATE=NO
MUNDER_DIFFLIN_AUTHORIZES_T080=NO

The required real T079 evidence remains a bounded, legitimate, pre-existing local Codex App Server round-trip against the exact Winds candidate under canonical Spec 006. This donor can inform future implementation/reconciliation after the required evidence gate, but cannot replace that gate.

Keep this note research-only. Do not amend active Spec 006 Tasks or PR #80 solely because this donor exists.