agent-bios 0.15.0 → 0.17.0

package/package.json

@@ -1,8 +1,8 @@
 {
   "name": "agent-bios",
-  "version": "0.15.0",
-  "releaseDate": "2026-09-01",
-  "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`.",
+  "version": "0.17.0",
+  "releaseDate": "2026-09-11",
+  "description": "A thin, low-level instruction layer for LLM CLI agents: one set of principles and behavior whichever model you run. Stores a private, editable corpus for explicitly activated sessions.",
   "bin": {
     "agent-bios": "install.sh"
   },
@@ -24,6 +24,7 @@
     "learn/promotions.json",
     "launch/agent-launch.zsh",
     "launch/agent-launch.py",
+    "launch/shell_integration.py",
     "compose/pkgid.py",
     "compose/register-hooks.py",
     "compose/assemble.py",
@@ -32,6 +33,15 @@
     "compose/corpus-state.py",
     "compose/write-update-cache.py",
     "compose/canary.sh",
+    "compose/corpus.py",
+    "compose/corpus_ui.py",
+    "compose/corpus_store.py",
+    "compose/corpus_catalog.py",
+    "compose/corpus_session.py",
+    "compose/corpus_install.py",
+    "compose/corpus_transaction.py",
+    "compose/corpus_understand.py",
+    "compose/bootstrap/",
     "launch/check-prompting-targets.sh",
     "learn/check-learning.py",
     "learn/collect-learning.py",

package/provenance.json

@@ -1 +1 @@
-{"commit":"ff09dd50ec264eee660513f8a6768d3481a4e9c8","committedAt":"2026-09-01T14:37:00+09:00","dirty":false}
+{"commit":"7b616f4ae5366b1395560b3a88556287ac36122d","committedAt":"2026-09-11T15:52:49+09:00","dirty":false}

package/session-cost.py

@@ -18,12 +18,14 @@ the bill: measured over two real sessions here, cache read + cache write were
 92-94% of cost and output 6-8%, so context size is the cost, and the only lever
 on it is how long a session is allowed to grow before it is reset.
 """
-import json, sys, glob, os, itertools, tempfile
+import json, sys, glob, os, itertools, tempfile, subprocess
 from datetime import datetime, timezone
 
 # $/MTok: input, output, cache_read, cache_write_5m, cache_write_1h
 PRICES = {
     "claude-fable-5":  (10.0, 50.0, 1.00, 12.50, 20.0),
+    # Fable 5.1 model overview, verified 2026-09-08.
+    "claude-fable-5-1": (10.0, 50.0, 0.25, 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),
@@ -35,7 +37,8 @@ PRICES = {
 }
 
 def price_for(model):
-    for k, v in PRICES.items():
+    # Version-specific rates take precedence over their family prefix.
+    for k, v in sorted(PRICES.items(), key=lambda item: len(item[0]), reverse=True):
         if model.startswith(k):
             return v
     return None
@@ -393,6 +396,23 @@ def _self_test():
         return p
 
     with tempfile.TemporaryDirectory() as td:
+        # A version-specific cache rate must reach the public report without
+        # changing the older model's rate. Run the real CLI/scan/cost path.
+        for model, expected in (("claude-fable-5-1", "$0.25"),
+                                ("claude-fable-5", "$1.00")):
+            transcript = write(td, model + ".jsonl", [
+                {"type": "assistant", "message": {
+                    "id": "cache-price", "model": model,
+                    "usage": {"cache_read_input_tokens": 1_000_000}}},
+            ])
+            result = subprocess.run(
+                [sys.executable, os.path.abspath(__file__), transcript],
+                capture_output=True, text=True)
+            check(model + " cache price reaches the report",
+                  (result.returncode,
+                   "modeled total cost: " + expected in result.stdout),
+                  (0, True))
+
         # Claude: two requests. The first is written twice as it streams, the
         # second is a sidechain that must not be counted.
         claude = write(td, "claude.jsonl", [

package/wrappers/codex-helm.sh

@@ -233,9 +233,9 @@ EOF
   if [ "$fanout" -eq 1 ]; then
     cat <<EOF
 - Fan-out is authorized when delegation gates fire; do de-minimis work directly. Native multi-agent is off unless the main is explicit Ultra.
-- Internal roles via codex-run --profile inherit and bounded stdin packet: WORKHORSE=gpt-5.6-terra/high/read-only-or-workspace-write; SWEEP=gpt-5.6-luna/low/read-only; REVIEWER=gpt-5.6-terra/high/read-only.
-- Native spawn cannot pin role/effort. Do not use native spawn_agent for FRONTIER. Use codex-run with gpt-5.6-sol/read-only and task-fit effort: max default, Ultra for divisible work, lower when cost/latency dominates.
-- Internal FRONTIER command base: $codex_run --profile inherit --model gpt-5.6-sol --effort <effort> --multi-agent auto --sandbox read-only --cd "\$PWD" -c agents.max_threads=$max_threads -c agents.max_depth=$max_depth. The adapter enables native multi-agent only for ultra. Send a self-contained packet on stdin; never full-history resume.
+- Internal roles via codex-run --profile inherit and bounded stdin packet: WORKHORSE=gpt-5.6-terra/xhigh/read-only-or-workspace-write; SWEEP=gpt-5.6-luna/max/read-only; REVIEWER=gpt-5.6-terra/xhigh/read-only.
+- Native spawn cannot pin role/effort. Do not use native spawn_agent for FRONTIER. Use codex-run with gpt-6-astra/read-only and task-fit effort: max default, Ultra for divisible work, lower when cost/latency dominates.
+- Internal FRONTIER command base: $codex_run --profile inherit --model gpt-6-astra --effort <effort> --multi-agent auto --sandbox read-only --cd "\$PWD" -c agents.max_threads=$max_threads -c agents.max_depth=$max_depth. The adapter enables native multi-agent only for ultra. Send a self-contained packet on stdin; never full-history resume.
 - Run independent units in parallel; respect depth/thread limits; wait and synthesize bounded reports.
 EOF
   else