agent-bios 0.9.2 → 0.9.4

package/claude/agents/workhorse.md

@@ -1,8 +1,8 @@
 ---
 name: workhorse
 description: WORKHORSE tier — bounded implementation, fixes, tests, and per-item judgment. Spawn decision-complete work with frozen scope/interfaces and a machine-checkable done-when; two or more independent items spawn in parallel.
-model: claude-sonnet-5
-effort: high
+model: claude-opus-5
+effort: medium
 ---
 
 Complete one bounded implementation, fix, or per-item judgment from a packet naming objective, frozen scope and inputs, allowed actions, output, done-when, and verification. Preserve out-of-scope behavior; batch independent reads; escalate missing decisions or authority instead of resolving them. Stage output for the main's acceptance — no external irreversible actions (push, install, credential, or remote mutation). Run the narrowest reliable changed-path check. Report: status, files_or_items_touched, evidence, verification or gap, risks_or_escalations.

package/claude/guides/claude-prompting.md

@@ -17,7 +17,7 @@ core_rules:
   - name the boundary explicitly — what to do without asking, and what to stop and ask about
 targets:
   - claude-fable-5
-  - claude-opus-4-8
+  - claude-opus-5
   - claude-sonnet-5
   - claude-haiku-4-5
 verification_focus:
@@ -55,15 +55,21 @@ pick the path.
 ## When to add blocks
 
 - Long-horizon or autonomous work: give the full spec up front in one
-  well-specified turn and run at a high effort. Add an explicit self-check
-  cadence; separate fresh-context verifier subagents beat self-critique.
+  well-specified turn and run at a high effort. Do **not** add a self-check
+  cadence or a dedicated verifier subagent: the helm binding verifies its own
+  work unasked, so an instruction to verify buys over-verification instead.
+  Deleting inherited verification scaffolding costs no capability — this inverts
+  the usual self-check advice, so carve this tier out of a prompt library that
+  applies that advice uniformly.
 - Review: state the evidence bar and the verdict shape. This tier follows
   severity filters literally, so "only report high-severity" depresses measured
   recall even as bug-finding improves — ask for every finding with confidence
   and severity attached, and filter downstream.
-- Delegation: say when to delegate. Left unprompted this tier under-reaches for
-  subagents, file-based memory, and custom tools — it will not spend an
-  expensive capability unless it is reasonably sure the capability is needed.
+- Delegation: say when *not* to delegate, and cap the spawn count. The helm
+  binding reaches for subagents readily — the reverse of the binding it replaced
+  — and every spawn rebuilds context, reports back, and is then re-read, so
+  unbounded delegation multiplies cost and latency. File-based memory and custom
+  tools are the opposite case: they still need an explicit when-to-use trigger.
 - Autonomous runs with no human watching: say so. Otherwise it asks permission
   it does not need and blocks. Grant autonomy on minor choices (naming,
   defaults, equivalent approaches) while keeping the ask for scope changes and
@@ -80,13 +86,17 @@ pick the path.
   per subtask, and the orchestrator is not pinned to the slowest one.
 - Effort ladder — `low`, `medium`, `high`, `xhigh`, `max` on the frontier, helm,
   and workhorse bindings; `xhigh` is the best setting for most coding and
-  agentic work, `high` a sound default, `low` for cheap subagent scans. Sweep the
-  levels on a real eval set rather than reaching for the top: higher effort up
-  front often *reduces* turn count and total cost on agentic work, while some
-  tasks land equally well a level down.
+  agentic work, `high` a sound default, `low` for cheap subagent scans. Start
+  there and sweep *down* on a real eval set rather than reaching for the top: the
+  lower rungs are unusually strong on the helm binding, higher effort up front
+  often *reduces* turn count and total cost on agentic work, and some tasks land
+  equally well a level down. Effort does not control response length — see
+  Working rules.
 - Per-model constraints differ across the `targets` bindings — thinking
   configuration, sampling parameters, and effort support are not uniform, and
-  the sweep binding is the most restricted. Confirm the constraint against the
+  the sweep binding is the most restricted. On the helm binding, for instance,
+  thinking is on by default and turning it off is accepted only at `high` effort
+  or below. Confirm the constraint against the
   live surface before relying on it in a dispatch; do not assume the frontier
   binding's rules apply to the sweep one.
 
@@ -98,6 +108,13 @@ pick the path.
 - Do not add "summarize every N tool calls" scaffolding — this tier narrates on
   its own. If it narrates too much for a coding agent, set a silence default
   instead: text only on a finding, a direction change, or a blocker.
+- Length is a prompting lever, not an effort lever. This tier writes longer
+  answers and longer files than its predecessors, and lowering `effort` does not
+  reliably shorten visible output — only an explicit instruction does. Calibrate
+  the deliverable's length separately from the conversation's.
+- Scope self-correction. Left alone this tier narrates its own earlier mistakes
+  at length, which reads as thrash. Ask it to correct only what would change the
+  reader's decisions, say it plainly, and carry on.
 - Give it somewhere to write learnings, tell it to consult that place later, and
   give the file a format. It performs notably better with a memory surface.
 - Keep the deliverable readable: the final message is the reader's first look at
@@ -111,7 +128,8 @@ pick the path.
 2. Name the boundaries — what to do freely, what to stop and ask about.
 3. Give each tool a when-to-call description.
 4. Say how progress claims must be grounded, and how the deliverable should read.
-5. Delete the step-by-step scaffolding inherited from older models, then A/B the
+5. Delete the step-by-step scaffolding inherited from older models — including
+   any verify-your-work step and any encouragement to delegate — then A/B the
    prompt with it removed before keeping either version.
 
 ## Sources

package/claude/guides/cli-multi-model-workflow.md

@@ -156,13 +156,13 @@ Write for the next agent and re-verification, not narrative. Required content:
 
 This is the human-readable projection of concrete models/tools; `config/agent-launch.toml` is the machine launch authority and parity checks keep them aligned. Re-probe when the binding is older than ~8 weeks or a newer observable model/tool changes the surface. `agent-bios install` overwrites deployed bindings, so edit the repo copy.
 
-Binding (2026-07-13):
+Binding (2026-07-25):
 
 | Slot | Binding | Notes |
 |---|---|---|
 | FRONTIER | Claude Fable 5 · GPT-5.6 Sol (read-only; max default, task-fit effort including Ultra) | bounded hardest decisions and verdicts |
-| HELM | Claude Opus 4.8 (xhigh) · GPT-5.6 Sol (xhigh main; main Ultra requires explicit selection; bounded FRONTIER Ultra allowed) | standing main; Codex defaults bypass, explicit sandbox narrows |
-| WORKHORSE | Claude Sonnet 5 · GPT-5.6 Terra (high) | implementation and per-item judgment |
+| HELM | Claude Opus 5 (xhigh) · GPT-5.6 Sol (xhigh main; main Ultra requires explicit selection; bounded FRONTIER Ultra allowed) | standing main; Codex defaults bypass, explicit sandbox narrows |
+| WORKHORSE | Claude Opus 5 (medium) · GPT-5.6 Terra (high) | implementation and per-item judgment |
 | SWEEP | Claude Haiku 4.5 · GPT-5.6 Luna (low) | clear repeatable scans and mechanical work |
 | VERIFIER-A | onto-mcp review | logic/authority kind; tier units per funnel |
 | VERIFIER-B | `ultracode-for-codex` CLI v0.4.5 (Codex-backed; callable from Claude or Codex) | code/execution kind |
@@ -186,8 +186,8 @@ Dispatch packets:
 | GPT-5.6 Sol FRONTIER | outcome, evidence, decision boundary, stop/verification; task-fit effort; read-only |
 | GPT-5.6 Terra WORKHORSE | outcome, frozen scope/inputs, authority, done-when, evidence/report, escalation; high |
 | GPT-5.6 Luna SWEEP | exact search space, rules, ambiguity behavior, stop, output; low; no architecture/debugging |
-| Claude Opus 4.8 HELM | xhigh for agentic work; high minimum for sensitive judgment; lower only when bounded/cost-led |
-| Claude Sonnet 5 WORKHORSE | exact scope, apply-to-all rules, tools, verification, report; high default |
+| Claude Opus 5 HELM | xhigh for agentic work; high minimum for sensitive judgment; lower only when bounded/cost-led |
+| Claude Opus 5 WORKHORSE | exact scope, apply-to-all rules, tools, verification, report; high default |
 | Claude Haiku 4.5 SWEEP | closed, high-volume, latency-sensitive work with exact schema/examples |
 
 Use only task-relevant tools; parallelize independent calls. Worker report: `status`, `files_or_items_touched`, `evidence`, `verification`, `risks_or_escalations`. Official basis: OpenAI [model](https://developers.openai.com/api/docs/guides/latest-model?model=gpt-5.6), [migration](https://developers.openai.com/api/docs/guides/upgrading-to-gpt-5p6-sol), [prompting](https://developers.openai.com/api/docs/guides/prompt-guidance-gpt-5p6), [Codex models](https://learn.chatgpt.com/docs/models); Anthropic [subagents](https://code.claude.com/docs/en/sub-agents), [model effort](https://code.claude.com/docs/en/model-config).

package/codex/guides/claude-prompting.md

@@ -17,7 +17,7 @@ core_rules:
   - name the boundary explicitly — what to do without asking, and what to stop and ask about
 targets:
   - claude-fable-5
-  - claude-opus-4-8
+  - claude-opus-5
   - claude-sonnet-5
   - claude-haiku-4-5
 verification_focus:
@@ -55,15 +55,21 @@ pick the path.
 ## When to add blocks
 
 - Long-horizon or autonomous work: give the full spec up front in one
-  well-specified turn and run at a high effort. Add an explicit self-check
-  cadence; separate fresh-context verifier subagents beat self-critique.
+  well-specified turn and run at a high effort. Do **not** add a self-check
+  cadence or a dedicated verifier subagent: the helm binding verifies its own
+  work unasked, so an instruction to verify buys over-verification instead.
+  Deleting inherited verification scaffolding costs no capability — this inverts
+  the usual self-check advice, so carve this tier out of a prompt library that
+  applies that advice uniformly.
 - Review: state the evidence bar and the verdict shape. This tier follows
   severity filters literally, so "only report high-severity" depresses measured
   recall even as bug-finding improves — ask for every finding with confidence
   and severity attached, and filter downstream.
-- Delegation: say when to delegate. Left unprompted this tier under-reaches for
-  subagents, file-based memory, and custom tools — it will not spend an
-  expensive capability unless it is reasonably sure the capability is needed.
+- Delegation: say when *not* to delegate, and cap the spawn count. The helm
+  binding reaches for subagents readily — the reverse of the binding it replaced
+  — and every spawn rebuilds context, reports back, and is then re-read, so
+  unbounded delegation multiplies cost and latency. File-based memory and custom
+  tools are the opposite case: they still need an explicit when-to-use trigger.
 - Autonomous runs with no human watching: say so. Otherwise it asks permission
   it does not need and blocks. Grant autonomy on minor choices (naming,
   defaults, equivalent approaches) while keeping the ask for scope changes and
@@ -80,13 +86,17 @@ pick the path.
   per subtask, and the orchestrator is not pinned to the slowest one.
 - Effort ladder — `low`, `medium`, `high`, `xhigh`, `max` on the frontier, helm,
   and workhorse bindings; `xhigh` is the best setting for most coding and
-  agentic work, `high` a sound default, `low` for cheap subagent scans. Sweep the
-  levels on a real eval set rather than reaching for the top: higher effort up
-  front often *reduces* turn count and total cost on agentic work, while some
-  tasks land equally well a level down.
+  agentic work, `high` a sound default, `low` for cheap subagent scans. Start
+  there and sweep *down* on a real eval set rather than reaching for the top: the
+  lower rungs are unusually strong on the helm binding, higher effort up front
+  often *reduces* turn count and total cost on agentic work, and some tasks land
+  equally well a level down. Effort does not control response length — see
+  Working rules.
 - Per-model constraints differ across the `targets` bindings — thinking
   configuration, sampling parameters, and effort support are not uniform, and
-  the sweep binding is the most restricted. Confirm the constraint against the
+  the sweep binding is the most restricted. On the helm binding, for instance,
+  thinking is on by default and turning it off is accepted only at `high` effort
+  or below. Confirm the constraint against the
   live surface before relying on it in a dispatch; do not assume the frontier
   binding's rules apply to the sweep one.
 
@@ -98,6 +108,13 @@ pick the path.
 - Do not add "summarize every N tool calls" scaffolding — this tier narrates on
   its own. If it narrates too much for a coding agent, set a silence default
   instead: text only on a finding, a direction change, or a blocker.
+- Length is a prompting lever, not an effort lever. This tier writes longer
+  answers and longer files than its predecessors, and lowering `effort` does not
+  reliably shorten visible output — only an explicit instruction does. Calibrate
+  the deliverable's length separately from the conversation's.
+- Scope self-correction. Left alone this tier narrates its own earlier mistakes
+  at length, which reads as thrash. Ask it to correct only what would change the
+  reader's decisions, say it plainly, and carry on.
 - Give it somewhere to write learnings, tell it to consult that place later, and
   give the file a format. It performs notably better with a memory surface.
 - Keep the deliverable readable: the final message is the reader's first look at
@@ -111,7 +128,8 @@ pick the path.
 2. Name the boundaries — what to do freely, what to stop and ask about.
 3. Give each tool a when-to-call description.
 4. Say how progress claims must be grounded, and how the deliverable should read.
-5. Delete the step-by-step scaffolding inherited from older models, then A/B the
+5. Delete the step-by-step scaffolding inherited from older models — including
+   any verify-your-work step and any encouragement to delegate — then A/B the
    prompt with it removed before keeping either version.
 
 ## Sources

package/codex/guides/cli-multi-model-workflow.md

@@ -156,13 +156,13 @@ Write for the next agent and re-verification, not narrative. Required content:
 
 This is the human-readable projection of concrete models/tools; `config/agent-launch.toml` is the machine launch authority and parity checks keep them aligned. Re-probe when the binding is older than ~8 weeks or a newer observable model/tool changes the surface. `agent-bios install` overwrites deployed bindings, so edit the repo copy.
 
-Binding (2026-07-13):
+Binding (2026-07-25):
 
 | Slot | Binding | Notes |
 |---|---|---|
 | FRONTIER | Claude Fable 5 · GPT-5.6 Sol (read-only; max default, task-fit effort including Ultra) | bounded hardest decisions and verdicts |
-| HELM | Claude Opus 4.8 (xhigh) · GPT-5.6 Sol (xhigh main; main Ultra requires explicit selection; bounded FRONTIER Ultra allowed) | standing main; Codex defaults bypass, explicit sandbox narrows |
-| WORKHORSE | Claude Sonnet 5 · GPT-5.6 Terra (high) | implementation and per-item judgment |
+| HELM | Claude Opus 5 (xhigh) · GPT-5.6 Sol (xhigh main; main Ultra requires explicit selection; bounded FRONTIER Ultra allowed) | standing main; Codex defaults bypass, explicit sandbox narrows |
+| WORKHORSE | Claude Opus 5 (medium) · GPT-5.6 Terra (high) | implementation and per-item judgment |
 | SWEEP | Claude Haiku 4.5 · GPT-5.6 Luna (low) | clear repeatable scans and mechanical work |
 | VERIFIER-A | onto-mcp review | logic/authority kind; tier units per funnel |
 | VERIFIER-B | `ultracode-for-codex` CLI v0.4.5 (Codex-backed; callable from Claude or Codex) | code/execution kind |
@@ -186,8 +186,8 @@ Dispatch packets:
 | GPT-5.6 Sol FRONTIER | outcome, evidence, decision boundary, stop/verification; task-fit effort; read-only |
 | GPT-5.6 Terra WORKHORSE | outcome, frozen scope/inputs, authority, done-when, evidence/report, escalation; high |
 | GPT-5.6 Luna SWEEP | exact search space, rules, ambiguity behavior, stop, output; low; no architecture/debugging |
-| Claude Opus 4.8 HELM | xhigh for agentic work; high minimum for sensitive judgment; lower only when bounded/cost-led |
-| Claude Sonnet 5 WORKHORSE | exact scope, apply-to-all rules, tools, verification, report; high default |
+| Claude Opus 5 HELM | xhigh for agentic work; high minimum for sensitive judgment; lower only when bounded/cost-led |
+| Claude Opus 5 WORKHORSE | exact scope, apply-to-all rules, tools, verification, report; high default |
 | Claude Haiku 4.5 SWEEP | closed, high-volume, latency-sensitive work with exact schema/examples |
 
 Use only task-relevant tools; parallelize independent calls. Worker report: `status`, `files_or_items_touched`, `evidence`, `verification`, `risks_or_escalations`. Official basis: OpenAI [model](https://developers.openai.com/api/docs/guides/latest-model?model=gpt-5.6), [migration](https://developers.openai.com/api/docs/guides/upgrading-to-gpt-5p6-sol), [prompting](https://developers.openai.com/api/docs/guides/prompt-guidance-gpt-5p6), [Codex models](https://learn.chatgpt.com/docs/models); Anthropic [subagents](https://code.claude.com/docs/en/sub-agents), [model effort](https://code.claude.com/docs/en/model-config).

package/config/agent-launch.toml

@@ -48,7 +48,7 @@ workhorse = "${CODEX_HOME}/agents/workhorse.toml"
 sweep = "${CODEX_HOME}/agents/sweep.toml"
 
 [hosts.claude]
-models = ["claude-fable-5", "claude-opus-4-8", "claude-sonnet-5", "claude-haiku-4-5"]
+models = ["claude-fable-5", "claude-opus-5", "claude-sonnet-5", "claude-haiku-4-5"]
 # onto review seat used when a CROSS-family main (Codex) routes onto review to
 # this (anthropic/claude) family; must be an onto review-role registered (provider, model).
 onto_review = { provider = "anthropic", model = "claude-fable-5" }
@@ -58,12 +58,12 @@ model = "claude-fable-5"
 effort = "max"
 
 [hosts.claude.tiers.helm]
-model = "claude-opus-4-8"
+model = "claude-opus-5"
 effort = "xhigh"
 
 [hosts.claude.tiers.workhorse]
-model = "claude-sonnet-5"
-effort = "high"
+model = "claude-opus-5"
+effort = "medium"
 
 [hosts.claude.tiers.sweep]
 model = "claude-haiku-4-5"

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "agent-bios",
-  "version": "0.9.2",
-  "releaseDate": "2026-07-23",
+  "version": "0.9.4",
+  "releaseDate": "2026-07-25",
   "description": "A thin, low-level instruction layer for LLM CLI agents: one set of principles and behavior whichever model you run. Deploys into $HOME by copy via an explicit `agent-bios install`.",
   "bin": {
     "agent-bios": "scripts/install.sh"

package/scripts/check-parity.sh

@@ -273,8 +273,8 @@ for tier, (model, effort) in expected_launch_tiers.items():
 
 expected_claude_tiers = {
     "frontier": ("claude-fable-5", "max"),
-    "helm": ("claude-opus-4-8", "xhigh"),
-    "workhorse": ("claude-sonnet-5", "high"),
+    "helm": ("claude-opus-5", "xhigh"),
+    "workhorse": ("claude-opus-5", "medium"),
     "sweep": ("claude-haiku-4-5", "low"),
 }
 claude_launch_tiers = launch_profile.get("hosts", {}).get("claude", {}).get("tiers", {})
@@ -409,8 +409,8 @@ for guide_name in [
         mark_fail(f"{guide} SWEEP must bind exact GPT-5.6 Luna")
     for slot, claude_model in (
         ("FRONTIER", "Claude Fable 5"),
-        ("HELM", "Claude Opus 4.8"),
-        ("WORKHORSE", "Claude Sonnet 5"),
+        ("HELM", "Claude Opus 5"),
+        ("WORKHORSE", "Claude Opus 5"),
         ("SWEEP", "Claude Haiku 4.5"),
     ):
         if rows[slot] and claude_model not in rows[slot][1]:

package/scripts/session-cost.py

@@ -2,7 +2,8 @@
 """Aggregate token usage & cost for a Claude Code session (main + subagents).
 
 Usage: session-cost.py <session>.jsonl [...]
-Reads the session transcript and, if present, <session-dir>/subagents/agent-*.jsonl.
+Reads the session transcript, splitting main-loop from subagent (sidechain)
+usage, and also picks up <session-dir>/subagents/agent-*.jsonl when present.
 Prints per-source, per-model token sums, modeled cost, and wall-clock span.
 """
 import json, sys, glob, os
@@ -12,6 +13,7 @@ from datetime import datetime
 PRICES = {
     "claude-fable-5":  (10.0, 50.0, 1.00, 12.50, 20.0),
     "claude-mythos-5": (10.0, 50.0, 1.00, 12.50, 20.0),
+    "claude-opus-5":   (5.0, 25.0, 0.50, 6.25, 10.0),
     "claude-opus-4-8": (5.0, 25.0, 0.50, 6.25, 10.0),
     "claude-opus-4-7": (5.0, 25.0, 0.50, 6.25, 10.0),
     "claude-opus-4-6": (5.0, 25.0, 0.50, 6.25, 10.0),
@@ -27,9 +29,17 @@ def price_for(model):
     return None
 
 def scan(path):
-    """-> {model: {in,out,cr,cw5,cw1,turns}}, (t_min, t_max)"""
-    agg, tmin, tmax, seen = {}, None, None, set()
-    for line in open(path):
+    """-> {(scope, model): {in,out,cr,cw5,cw1,turns}}, (t_min, t_max), {scope: agent_ids}
+
+    A single API response is written to the transcript several times as it
+    streams, each line carrying output_tokens *so far* (e.g. 2, 2, 2, 540).
+    Keep the record with the LARGEST output_tokens per message id: keeping the
+    first one instead under-reports subagent output by ~95%, because sidechain
+    messages get snapshotted far more often than main-loop ones do.
+    """
+    best, tmin, tmax = {}, None, None
+    agents = {"main": set(), "sub": set()}
+    for line in open(path, errors="replace"):
         try:
             d = json.loads(line)
         except json.JSONDecodeError:
@@ -41,11 +51,17 @@ def scan(path):
         u, model = m.get("usage"), m.get("model")
         if not u or not model or model == "<synthetic>":
             continue
-        rid = m.get("id") or d.get("requestId")
-        if rid in seen:
-            continue  # one usage record per API request
-        seen.add(rid)
-        a = agg.setdefault(model, dict(inp=0, out=0, cr=0, cw5=0, cw1=0, turns=0))
+        scope = "sub" if d.get("isSidechain") else "main"
+        if d.get("agentId"):
+            agents[scope].add(d["agentId"])
+        key = (scope, m.get("id") or d.get("requestId"))
+        prev = best.get(key)
+        if prev is None or u.get("output_tokens", 0) > prev[1].get("output_tokens", 0):
+            best[key] = (model, u)
+
+    agg = {}
+    for (scope, _), (model, u) in best.items():
+        a = agg.setdefault((scope, model), dict(inp=0, out=0, cr=0, cw5=0, cw1=0, turns=0))
         a["inp"] += u.get("input_tokens", 0)
         a["out"] += u.get("output_tokens", 0)
         a["cr"]  += u.get("cache_read_input_tokens", 0)
@@ -56,7 +72,7 @@ def scan(path):
         else:
             a["cw5"] += u.get("cache_creation_input_tokens", 0)
         a["turns"] += 1
-    return agg, (tmin, tmax)
+    return agg, (tmin, tmax), agents
 
 def cost(model, a):
     p = price_for(model)
@@ -70,19 +86,27 @@ def fmt(n):
 
 def report(session_path):
     base = session_path[:-6]  # strip .jsonl
-    sources = [("main", session_path)]
+    sources = [(None, session_path)]
     sources += [(os.path.basename(f)[:-6], f)
                 for f in sorted(glob.glob(os.path.join(base, "subagents", "*.jsonl")))]
     print(f"\n=== {os.path.basename(session_path)} ===")
     grand, unpriced = 0.0, []
-    hdr = f"{'source':<38}{'model':<22}{'turns':>6}{'input':>9}{'output':>9}{'cache_rd':>10}{'cache_wr':>10}{'cost$':>9}"
+    hdr = (f"{'source':<38}{'model':<22}{'turns':>6}{'input':>9}{'output':>9}"
+           f"{'cache_rd':>10}{'cache_wr':>10}{'cost$':>9}")
     print(hdr); print("-" * len(hdr))
     span_min = span_max = None
-    for name, path in sources:
-        agg, (tmin, tmax) = scan(path)
+    for label, path in sources:
+        agg, (tmin, tmax), agents = scan(path)
         if tmin:
             span_min = min(span_min or tmin, tmin); span_max = max(span_max or tmax, tmax)
-        for model, a in agg.items():
+        for (scope, model), a in sorted(agg.items()):
+            if label is not None:
+                name = label
+            elif scope == "sub":
+                n = len(agents["sub"])
+                name = f"subagents (n={n})" if n else "subagents"
+            else:
+                name = "main"
             c = cost(model, a)
             cs = f"{c:9.2f}" if c is not None else "        ?"
             if c is None: