fix(aider): CONVENTIONS.md is a roster index, not 3.8 million characters of agents (#871)

Aider loads a conventions file into context and keeps it there for the whole
session — that is what the file is for, and the docs say to load it with
--read so prompt caching can hold it. This integration concatenated every
agent body into it. At 279 agents that is 3,816,372 characters, roughly a
million tokens. No model takes that. Anyone who ran

    ./scripts/install.sh --tool aider

got a CONVENTIONS.md that either blows the context window on the first turn
or bills for a million tokens trying.

CONVENTIONS.md is now the roster index it was described as: one entry per
agent with the name, the description, the division, and the path to the
agent file. 96,823 characters, down from 3.8 million. The header explains
how to pull a single agent's full instructions into the session:

    /read-only /path/to/agency-agents/engineering/engineering-frontend-developer.md

Naming an agent in a prompt still works the way it did — the description is
what the model needed for that, and it is still there.

test-convert-outputs.sh now holds the index to being an index: it fails if
CONVENTIONS.md grows past 250,000 characters, if it does not list exactly
one path per roster agent, or if any path it prints does not resolve. The
existing round-trip check on the accumulated file still covers the names and
descriptions.

scripts/convert.sh

@@ -14,7 +14,7 @@
 #   gemini-cli   — Gemini CLI subagent files (~/.gemini/agents/*.md)
 #   opencode     — OpenCode agent files (.opencode/agents/*.md)
 #   cursor       — Cursor rule files (.cursor/rules/*.mdc)
-#   aider        — Single CONVENTIONS.md for Aider
+#   aider        — Single CONVENTIONS.md roster index for Aider
 #   windsurf     — Single .windsurfrules for Windsurf
 #   openclaw     — OpenClaw workspaces (integrations/openclaw/<agent>/SOUL.md)
 #   qwen         — Qwen Code SubAgent files (~/.qwen/agents/*.md)
@@ -568,11 +568,21 @@ trap 'rm -f "$AIDER_TMP" "$WINDSURF_TMP"' EXIT
 cat > "$AIDER_TMP" <<'HEREDOC'
 # The Agency — AI Agent Conventions
 #
-# This file provides Aider with the full roster of specialized AI agents from
-# The Agency (https://github.com/msitarzewski/agency-agents).
+# The roster of specialized AI agents from The Agency
+# (https://github.com/msitarzewski/agency-agents): each one's name, what it is
+# for, and where its full instructions live.
 #
-# To activate an agent, reference it by name in your Aider session prompt, e.g.:
-#   "Use the Frontend Developer agent to review this component."
+# Aider keeps a conventions file in context for the whole session, so this is an
+# index and not the agents themselves. Inlining every body would make this file
+# about 3.8 million characters, which no model will take.
+#
+# To use an agent:
+#   1. Name it in your prompt — "Use the Frontend Developer agent to review
+#      this component." The description below is usually enough for that.
+#   2. For its full instructions, pull the agent file into the session:
+#        /read-only /path/to/agency-agents/engineering/engineering-frontend-developer.md
+#
+# Paths below are relative to an agency-agents checkout.
 #
 # Generated by scripts/convert.sh — do not edit manually.
 
@@ -590,12 +600,16 @@ HEREDOC
 
 accumulate_aider() {
   local file="$1"
-  local name description body
+  local name description source division
 
   name="$(get_field "name" "$file")"
   description="$(get_field "description" "$file")"
-  body="$(get_body "$file")"
+  source="${file#"$REPO_ROOT"/}"
+  division="${source%%/*}"
 
+  # One index entry per agent, not the agent. A conventions file is read into
+  # every request; the bodies together are 3.8 million characters and this
+  # index is about 90,000.
   cat >> "$AIDER_TMP" <<HEREDOC
 
 ---
@@ -604,7 +618,8 @@ accumulate_aider() {
 
 > ${description}
 
-${body}
+Division: ${division}
+Full instructions: ${source}
 HEREDOC
 }
 

scripts/install.sh

@@ -18,7 +18,7 @@
 #   gemini-cli   -- Install agents to ~/.gemini/agents/
 #   opencode     -- Copy agents to .opencode/agents/ in current directory
 #   cursor       -- Copy rules to .cursor/rules/ in current directory
-#   aider        -- Copy CONVENTIONS.md to current directory
+#   aider        -- Copy the CONVENTIONS.md roster index to current directory
 #   windsurf     -- Copy .windsurfrules to current directory
 #   openclaw     -- Copy workspaces to ~/.openclaw/agency-agents/
 #   qwen         -- Copy SubAgents to ~/.qwen/agents/ (user-wide) or .qwen/agents/ (project)
@@ -926,6 +926,8 @@ install_aider() {
   fi
   install_file "$src" "$dest"
   ok "Aider: installed -> $dest"
+  dim  "       CONVENTIONS.md is the roster index. Load one agent's full instructions with"
+  dim  "       /read-only $REPO_ROOT/<path shown in the index>"
   $SELECTION_ACTIVE && warn "Aider: single-file format — team/agent filtering N/A (installs the full roster)."
   warn "Aider: project-scoped. Run from your project root to install there."
 }

scripts/test-convert-outputs.sh

@@ -353,6 +353,31 @@ if split_bad:
 else:
     ok(f"openclaw: all {N} agents keep every source fenced block whole in one output file")
 
+# --- Layer A (context budget): the Aider index has to stay an index ----------
+# Aider keeps a conventions file in context for the whole session. Inlining the
+# agent bodies made CONVENTIONS.md 3.8 million characters, which no model will
+# take, so it carries one index entry per agent instead: description plus the
+# path to the real file. Two things have to hold for that to be worth anything —
+# the file stays small enough to load, and every path it prints resolves.
+AIDER_INDEX_CEILING = 250_000
+aider_index = os.path.join(OUT, "aider", "CONVENTIONS.md")
+if os.path.isfile(aider_index):
+    text = open(aider_index, encoding="utf-8").read()
+    if len(text) > AIDER_INDEX_CEILING:
+        bad(f"aider: CONVENTIONS.md is {len(text):,} characters — it is loaded into "
+            f"every request, so it has to stay an index, not the agents themselves")
+    paths = re.findall(r"^Full instructions: (.+)$", text, re.M)
+    dangling = sorted({p for p in paths if not os.path.isfile(os.path.join(R, p))})
+    if len(paths) != N:
+        bad(f"aider: CONVENTIONS.md points at {len(paths)} agent files, roster has {N}")
+    elif dangling:
+        for d in dangling[:3]:
+            bad(f"aider: CONVENTIONS.md points at a file that does not exist: {d}")
+        if len(dangling) > 3:
+            bad(f"aider: ...and {len(dangling)-3} more dangling paths")
+    elif len(text) <= AIDER_INDEX_CEILING:
+        ok(f"aider: index is {len(text):,} characters and all {N} agent paths resolve")
+
 # --- Layer A (app-facing): every SOURCE frontmatter strict-parsed above -------
 for m in src_bad[:5]: bad(m)
 if len(src_bad) > 5: bad(f"...and {len(src_bad)-5} more source frontmatter problems")