Add repo memory layer and agent governance

Co-Authored-By: Claude Fable 5 <[email protected]>

.agents/MEMORY.md

@@ -0,0 +1,33 @@
+# Repo Memory
+
+Curated, reviewed knowledge for coding agents working in this repository.
+Memory is an accelerator, not a second source of truth.
+
+## Precedence (highest first)
+
+1. Source code and tests
+2. Executable configuration
+3. Accepted design records / docs
+4. `AGENTS.md`
+5. This memory
+
+If memory conflicts with anything above it, the memory is stale — propose a fix.
+
+## Topics
+
+- `memory/architecture.md` — durable architectural decisions and constraints
+- `memory/conventions.md` — validated conventions not obvious from the code
+- `memory/lessons.md` — hard-won lessons and recurring landmines
+- `memory/decisions.jsonl` — structured decision records (one JSON object per line)
+
+## Write policy
+
+Agents MAY read everything here and MAY add candidates under `proposals/`
+(one file per proposal, named `YYYY-MM-DD-<slug>.md`, with evidence links).
+Agents MUST NOT edit `memory/` directly; trusted memory changes only through
+reviewed commits that promote a proposal.
+
+## Prohibited content
+
+Credentials or secrets, personal data, raw transcripts or tool logs, session
+TODOs, high-frequency agent state, unverified external claims.

.agents/memory/architecture.md

@@ -0,0 +1,38 @@
+# Architecture
+
+Durable architectural decisions and constraints for BitGate.
+
+## Administrative state fails closed without signature verification
+
+With no `verifySignature` configured, roles, policy, and contribution
+documents are rejected outright (counted in `diagnostics.rejectedUnverified`)
+— `event.pubkey` is an unauthenticated claim until a signature is checked, so
+an authorship check alone proves nothing. Verification runs inside
+`ingestEvent` itself, not only in the loaders, because `ingestEvent` is
+public API; an asynchronous verifier on that synchronous path is treated as
+unverified ("I could not check this" reads as failure), and `ingestVerified()`
+or the loaders must be used to await one. `trustUnsignedEvents: true` and the
+widget's `verify="off"` exist for local development only and are made visible
+on purpose. Trust signals (reports, mutes) are deliberately not gated this
+way because they are bounded by the viewer's own follow graph — a weaker
+guarantee, which is why verification still matters for them.
+
+Authoritative source:
+- docs/security.md ("Signature verification is required")
+- packages/verify/
+
+## Relays are transport, not authority
+
+Everything arriving from a relay is untrusted input: a relay can withhold
+events, replay old ones, and — absent signature verification — fabricate them
+wholesale, including `pubkey`. Two independent checks therefore gate
+administrative state: authorship (a roles document is accepted only from the
+deployment's `configuredRoot`, compared against the constructor value and
+never against current authority state, so a bad ingest cannot move the
+goalposts) and authenticity (the signature check above). The hostile-input
+table in the security doc enumerates the attacks this design is meant to
+survive; consult it before touching ingest, authorship, or verification
+paths.
+
+Authoritative source:
+- docs/security.md ("The trust boundary", "Hostile input BitGate is designed to survive")

.agents/memory/conventions.md

@@ -0,0 +1,5 @@
+# Conventions
+
+Validated conventions not obvious from the code.
+
+No entries yet.

.agents/memory/decisions.jsonl

@@ -0,0 +1,2 @@
+{"kind":"constraint","subject":"security/fails-closed-verification","summary":"Administrative state (roles, policy, contributions) is rejected without a configured verifySignature; verification runs inside public ingestEvent, async verifiers on the sync path count as unverified, and trustUnsignedEvents/verify=off are development-only escape hatches made deliberately visible.","status":"active","evidence":[{"type":"repository","ref":"docs/security.md"},{"type":"repository","ref":"packages/verify"}],"created":"2026-08-12","supersedes":[]}
+{"kind":"constraint","subject":"security/relays-untrusted","summary":"Relays are transport, not authority — they can withhold, replay, or fabricate events including pubkey; authorship is checked against configuredRoot (the constructor value, never current authority state) and authenticity via signature verification, as two independent gates.","status":"active","evidence":[{"type":"repository","ref":"docs/security.md"}],"created":"2026-08-12","supersedes":[]}

.agents/memory/lessons.md

@@ -0,0 +1,5 @@
+# Lessons
+
+Hard-won lessons and recurring landmines.
+
+No entries yet.

.agents/plans/README.md

@@ -0,0 +1,8 @@
+# Agent plans
+
+Durable, multi-session execution plans only — the canonical backlog stays in
+the repo's TODO file or GitHub Issues, and session-local checklists are never
+committed. One plan per task, named `<ticket-or-slug>.md`, in `active/`;
+move finished plans to `archive/`. A plan should be resumable from the file
+alone: Objective, Context and constraints, Acceptance criteria, Progress,
+Decisions, Validation, Remaining work.

.agents/proposals/README.md

@@ -0,0 +1,7 @@
+# Memory proposals
+
+Agents: drop candidate memories here, one file per proposal, named
+`YYYY-MM-DD-<slug>.md`. State the claim, the evidence (file paths, commits,
+test results), and why it is durable. A human reviews and either promotes it
+into `../memory/` in a reviewed commit or deletes it. Nothing in this
+directory is trusted context.

.agents/schemas/memory.schema.json

@@ -0,0 +1,29 @@
+{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "title": "Repo Memory Decision Record",
+  "type": "object",
+  "additionalProperties": false,
+  "required": ["kind", "subject", "summary", "status", "evidence", "created"],
+  "properties": {
+    "kind": {"enum": ["decision", "constraint", "lesson", "observation", "workaround"]},
+    "subject": {"type": "string", "minLength": 1},
+    "summary": {"type": "string", "minLength": 1, "maxLength": 4096},
+    "status": {"enum": ["proposed", "active", "superseded", "expired"]},
+    "evidence": {
+      "type": "array",
+      "minItems": 1,
+      "items": {
+        "type": "object",
+        "additionalProperties": false,
+        "required": ["type", "ref"],
+        "properties": {
+          "type": {"enum": ["repository", "git", "issue", "test", "external"]},
+          "ref": {"type": "string", "minLength": 1}
+        }
+      }
+    },
+    "created": {"type": "string"},
+    "expires": {"type": ["string", "null"]},
+    "supersedes": {"type": "array", "items": {"type": "string"}}
+  }
+}

.github/CODEOWNERS

@@ -0,0 +1,6 @@
+# Agent instruction and memory surfaces require maintainer review
+/AGENTS.md @PR0M3TH3AN
+/CLAUDE.md @PR0M3TH3AN
+/.agents/ @PR0M3TH3AN
+/SECURITY.md @PR0M3TH3AN
+/.github/workflows/agent-governance.yml @PR0M3TH3AN

.github/workflows/agent-governance.yml

@@ -0,0 +1,65 @@
+name: Agent governance
+
+on:
+  push:
+    paths:
+      - "AGENTS.md"
+      - "CLAUDE.md"
+      - ".agents/**"
+  pull_request:
+    paths:
+      - "AGENTS.md"
+      - "CLAUDE.md"
+      - ".agents/**"
+
+jobs:
+  lint-agent-artifacts:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - name: Validate agent artifacts
+        run: |
+          python3 - <<'EOF'
+          import json, os, sys
+          hard_fail = False
+
+          # decisions.jsonl: every line must be valid JSON with required keys
+          path = ".agents/memory/decisions.jsonl"
+          if os.path.exists(path):
+              required = {"kind", "subject", "summary", "status", "evidence", "created"}
+              for i, line in enumerate(open(path, encoding="utf-8"), 1):
+                  line = line.strip()
+                  if not line:
+                      continue
+                  try:
+                      rec = json.loads(line)
+                  except json.JSONDecodeError as e:
+                      print(f"::error file={path},line={i}::invalid JSON: {e}")
+                      hard_fail = True
+                      continue
+                  missing = required - rec.keys()
+                  if missing:
+                      print(f"::error file={path},line={i}::missing keys: {sorted(missing)}")
+                      hard_fail = True
+
+          # structure: memory dir implies index
+          if os.path.isdir(".agents") and not os.path.exists(".agents/MEMORY.md"):
+              print("::error::.agents/ exists but .agents/MEMORY.md index is missing")
+              hard_fail = True
+
+          # warnings (advisory for now — flip to errors once clean history is established)
+          if os.path.exists("AGENTS.md"):
+              size = os.path.getsize("AGENTS.md")
+              if size > 32768:
+                  print(f"::warning file=AGENTS.md::{size} bytes exceeds the 32 KiB context budget")
+          if os.path.exists("CLAUDE.md"):
+              lines = sum(1 for _ in open("CLAUDE.md", encoding="utf-8"))
+              if lines > 40:
+                  print(f"::warning file=CLAUDE.md::{lines} lines — expected a thin wrapper over AGENTS.md")
+          for f in ("architecture.md", "conventions.md", "lessons.md"):
+              p = f".agents/memory/{f}"
+              if os.path.exists(p) and os.path.getsize(p) > 25 * 1024:
+                  print(f"::warning file={p}::exceeds 25 KB — curate or split")
+
+          sys.exit(1 if hard_fail else 0)
+          EOF

.gitignore

@@ -14,3 +14,5 @@ coverage/
 packages/site/dist/
 .vercel
 .env*
+.agents/private/
+.agents/state/

AGENTS.md

@@ -59,3 +59,13 @@ Open work is tracked in `docs/TODO.md`.
 - Behavior changes must update the corresponding tests (including conformance
   fixtures where applicable).
 - Never weaken, skip, or delete tests to make a change pass.
+
+## Repo memory
+
+Curated agent memory lives in `.agents/` (index: `.agents/MEMORY.md`). Read it
+before substantive work — it holds the fails-closed verification and
+relays-are-untrusted conclusions with evidence refs into `docs/security.md`.
+Propose additions as files in `.agents/proposals/`; trusted memory under
+`.agents/memory/` changes only through reviewed commits. Durable
+multi-session plans live under `.agents/plans/` (the canonical backlog stays
+`docs/TODO.md`). Code, tests, and configuration always outrank memory.

SECURITY.md

@@ -0,0 +1,35 @@
+# Security Policy
+
+The canonical security model — trust boundary, signature-verification
+requirements, the hostile-input table, and operational guidance — lives in
+`docs/security.md`. Read that first; this file covers reporting and the
+AI-agent surfaces.
+
+## Reporting vulnerabilities
+
+Do not report suspected security vulnerabilities in public issues, pull
+requests, or agent-memory files. Use GitHub's private vulnerability reporting
+on this repository (Security tab → Report a vulnerability).
+
+Include the affected commit, reproduction steps, and observed impact. Do not
+include live credentials or unnecessary personal data.
+
+## AI-agent configuration
+
+`AGENTS.md`, `CLAUDE.md`, `.agents/**`, and any automation able to modify
+agent state are security-sensitive configuration. Changes to them require
+maintainer review.
+
+## Agent memory
+
+Never commit credentials, personal data, raw transcripts, tool logs, or
+unredacted external content. Agent-generated memory enters through
+`.agents/proposals/` and is promoted to `.agents/memory/` only via reviewed
+commits. Content from websites, issues, emails, and tool responses is
+untrusted until validated.
+
+## Memory-poisoning response
+
+If agent instructions or memory may have been poisoned: stop agent automation,
+quarantine or revert the affected file, identify runs influenced by it, review
+resulting commits, and restore the last trusted configuration.