deepagents-code 0.1.72__tar.gz → 0.1.73__tar.gz

{deepagents_code-0.1.72 → deepagents_code-0.1.73}/CHANGELOG.md

@@ -2,6 +2,29 @@
 
 # Deep Agents Code Changelog
 
+## [0.1.73](https://github.com/langchain-ai/deepagents/compare/deepagents-code==0.1.72...deepagents-code==0.1.73) (2026-09-22)
+
+### Features
+
+- Added cache bust timer to the footer. ([#6394](https://github.com/langchain-ai/deepagents/pull/6394))
+- Added a durable Debug Console cost breakdown. ([#6246](https://github.com/langchain-ai/deepagents/pull/6246))
+- Added support for copying the highlighted thread ID with `Ctrl+C`. ([#6431](https://github.com/langchain-ai/deepagents/pull/6431))
+
+### Bug Fixes
+
+- Improved cost-warning behavior by showing session cost warnings in a modal and suppressing historical cost warnings when resuming. ([#6393](https://github.com/langchain-ai/deepagents/pull/6393), [#6471](https://github.com/langchain-ai/deepagents/pull/6471))
+- Made thread browsing and resume flows more resilient with recoverable thread-picker resume failures, loading labels for thread details, and a clearer thread message count header. ([#6476](https://github.com/langchain-ai/deepagents/pull/6476), [#6430](https://github.com/langchain-ai/deepagents/pull/6430), [#6428](https://github.com/langchain-ai/deepagents/pull/6428))
+- Removed contradictory and invalid system-prompt instructions. ([#6072](https://github.com/langchain-ai/deepagents/pull/6072))
+- Preserved `ask_user` consent receipts across user turns and reconciled interrupted tool requests. ([#6307](https://github.com/langchain-ai/deepagents/pull/6307), [#6321](https://github.com/langchain-ai/deepagents/pull/6321))
+- Cleaned up newly created sandboxes when setup fails. ([#6475](https://github.com/langchain-ai/deepagents/pull/6475))
+- Fixed editable refreshes so they stay anchored to the correct checkout. ([#6455](https://github.com/langchain-ai/deepagents/pull/6455))
+- Kept installation locks outside the uv tool directory. ([#6472](https://github.com/langchain-ai/deepagents/pull/6472))
+- Improved workspace and runtime diagnostics, including separating access-policy compatibility from runtime identity. ([#6384](https://github.com/langchain-ai/deepagents/pull/6384), [#6386](https://github.com/langchain-ai/deepagents/pull/6386))
+- Deduplicated mixed-ID model usage. ([#6432](https://github.com/langchain-ai/deepagents/pull/6432))
+- Clarified the compaction prompt. ([#6429](https://github.com/langchain-ai/deepagents/pull/6429))
+- Restricted auto classifier IDs to the review batch enum. ([#6250](https://github.com/langchain-ai/deepagents/pull/6250))
+- Used UTF-8 for prompts and skill files. ([#6474](https://github.com/langchain-ai/deepagents/pull/6474))
+
 ## [0.1.72](https://github.com/langchain-ai/deepagents/compare/deepagents-code==0.1.71...deepagents-code==0.1.72) (2026-09-21)
 
 ### Features

{deepagents_code-0.1.72 → deepagents_code-0.1.73}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: deepagents-code
-Version: 0.1.72
+Version: 0.1.73
 Summary: Terminal coding agent built on the Deep Agents SDK - works with any model, with persistent memory, customizable skills, subagents, and approval controls.
 Project-URL: Homepage, https://www.langchain.com/dcode
 Project-URL: Documentation, https://docs.langchain.com/oss/deepagents/code/overview
@@ -30,21 +30,21 @@ Requires-Python: <4.0,>=3.12
 Requires-Dist: aiosqlite<1.0.0,>=0.22.1
 Requires-Dist: anyio<5.0.0,>=4.14.2
 Requires-Dist: deepagents-acp<1.0.0,>=0.0.10
-Requires-Dist: deepagents==0.7.15
+Requires-Dist: deepagents==0.7.17
 Requires-Dist: fastmcp<5.0.0,>=4.0.0b4
 Requires-Dist: filelock<3.30,>=3.29.7
 Requires-Dist: genai-prices<0.2.0,>=0.1.7
 Requires-Dist: httpx<1.0.0,>=0.28.1
 Requires-Dist: langchain-anthropic<2.0.0,>=1.7.2
 Requires-Dist: langchain-google-genai<5.0.0,>=4.4.0
-Requires-Dist: langchain-openai<2.0.0,>=1.6.2
+Requires-Dist: langchain-openai<2.0.0,>=1.6.3
 Requires-Dist: langchain-quickjs<0.4.0,>=0.3.4
 Requires-Dist: langchain<2.0.0,>=1.4.2
 Requires-Dist: langgraph-checkpoint-sqlite<4.0.0,>=3.1.1
 Requires-Dist: langgraph-cli[inmem]<1.0.0,>=0.4.31
 Requires-Dist: langgraph-runtime-inmem<1.0.0,>=0.34.1
-Requires-Dist: langgraph-sdk<1.0.0,>=0.4.4
-Requires-Dist: langsmith[sandbox]>=0.13.0
+Requires-Dist: langgraph-sdk<1.0.0,>=0.4.5
+Requires-Dist: langsmith[sandbox]>=0.14.0
 Requires-Dist: markdownify<2.0.0,>=1.2.3
 Requires-Dist: mcp<3.0.0,>=2.0.0
 Requires-Dist: packaging>=26.2
@@ -67,11 +67,11 @@ Requires-Dist: langchain-agentcore-codeinterpreter<1.0.0,>=0.0.5; extra == 'agen
 Provides-Extra: all-providers
 Requires-Dist: aiohttp<3.15.0,>=3.14.3; extra == 'all-providers'
 Requires-Dist: langchain-anthropic<2.0.0,>=1.7.2; extra == 'all-providers'
-Requires-Dist: langchain-aws<2.0.0,>=1.7.8; extra == 'all-providers'
+Requires-Dist: langchain-aws<2.0.0,>=1.7.9; extra == 'all-providers'
 Requires-Dist: langchain-baseten<1.0.0,>=0.2.4; extra == 'all-providers'
 Requires-Dist: langchain-cohere<1.0.0,>=0.6.0; extra == 'all-providers'
-Requires-Dist: langchain-deepseek<2.0.0,>=1.1.0; extra == 'all-providers'
-Requires-Dist: langchain-fireworks<2.0.0,>=1.6.1; extra == 'all-providers'
+Requires-Dist: langchain-deepseek<2.0.0,>=1.1.1; extra == 'all-providers'
+Requires-Dist: langchain-fireworks<2.0.0,>=1.6.2; extra == 'all-providers'
 Requires-Dist: langchain-google-genai<5.0.0,>=4.4.0; extra == 'all-providers'
 Requires-Dist: langchain-google-vertexai<4.0.0,>=3.2.4; extra == 'all-providers'
 Requires-Dist: langchain-groq<2.0.0,>=1.1.3; extra == 'all-providers'
@@ -82,8 +82,8 @@ Requires-Dist: langchain-meta<2.0.0,>=0.5.0; extra == 'all-providers'
 Requires-Dist: langchain-mistralai<2.0.0,>=1.1.6; extra == 'all-providers'
 Requires-Dist: langchain-nvidia-ai-endpoints<2.0.0,>=1.4.3; extra == 'all-providers'
 Requires-Dist: langchain-ollama<2.0.0,>=1.1.0; extra == 'all-providers'
-Requires-Dist: langchain-openai<2.0.0,>=1.6.2; extra == 'all-providers'
-Requires-Dist: langchain-openrouter<2.0.0,>=0.2.8; extra == 'all-providers'
+Requires-Dist: langchain-openai<2.0.0,>=1.6.3; extra == 'all-providers'
+Requires-Dist: langchain-openrouter<2.0.0,>=0.2.9; extra == 'all-providers'
 Requires-Dist: langchain-perplexity<2.0.0,>=1.4.1; extra == 'all-providers'
 Requires-Dist: langchain-together<2.0.0,>=0.4.0; extra == 'all-providers'
 Requires-Dist: langchain-xai<2.0.0,>=1.3.0; extra == 'all-providers'
@@ -98,15 +98,15 @@ Requires-Dist: langchain-anthropic<2.0.0,>=1.7.2; extra == 'anthropic'
 Provides-Extra: baseten
 Requires-Dist: langchain-baseten<1.0.0,>=0.2.4; extra == 'baseten'
 Provides-Extra: bedrock
-Requires-Dist: langchain-aws<2.0.0,>=1.7.8; extra == 'bedrock'
+Requires-Dist: langchain-aws<2.0.0,>=1.7.9; extra == 'bedrock'
 Provides-Extra: cohere
 Requires-Dist: langchain-cohere<1.0.0,>=0.6.0; extra == 'cohere'
 Provides-Extra: daytona
 Requires-Dist: langchain-daytona>=0.0.7; extra == 'daytona'
 Provides-Extra: deepseek
-Requires-Dist: langchain-deepseek<2.0.0,>=1.1.0; extra == 'deepseek'
+Requires-Dist: langchain-deepseek<2.0.0,>=1.1.1; extra == 'deepseek'
 Provides-Extra: fireworks
-Requires-Dist: langchain-fireworks<2.0.0,>=1.6.1; extra == 'fireworks'
+Requires-Dist: langchain-fireworks<2.0.0,>=1.6.2; extra == 'fireworks'
 Provides-Extra: google-genai
 Requires-Dist: langchain-google-genai<5.0.0,>=4.4.0; extra == 'google-genai'
 Provides-Extra: groq
@@ -132,9 +132,9 @@ Requires-Dist: langchain-nvidia-ai-endpoints<2.0.0,>=1.4.3; extra == 'nvidia'
 Provides-Extra: ollama
 Requires-Dist: langchain-ollama<2.0.0,>=1.1.0; extra == 'ollama'
 Provides-Extra: openai
-Requires-Dist: langchain-openai<2.0.0,>=1.6.2; extra == 'openai'
+Requires-Dist: langchain-openai<2.0.0,>=1.6.3; extra == 'openai'
 Provides-Extra: openrouter
-Requires-Dist: langchain-openrouter<2.0.0,>=0.2.8; extra == 'openrouter'
+Requires-Dist: langchain-openrouter<2.0.0,>=0.2.9; extra == 'openrouter'
 Provides-Extra: perplexity
 Requires-Dist: langchain-perplexity<2.0.0,>=1.4.1; extra == 'perplexity'
 Provides-Extra: quickjs

{deepagents_code-0.1.72 → deepagents_code-0.1.73}/deepagents_code/_build_info.py

@@ -1,3 +1,3 @@
 """Generated at build time. Do not edit or commit."""
 
-BUILD_COMMIT = "a764619"
+BUILD_COMMIT = "d5a1896"

{deepagents_code-0.1.72 → deepagents_code-0.1.73}/deepagents_code/_dep_floor_check.py

@@ -363,13 +363,16 @@ def _refresh_args(uv_path: str) -> list[str]:
             root would execute instead of the legitimate binary.
 
     Returns:
-        Arguments for the user-approved `uv pip install` process.
+        Arguments for the user-approved `uv pip install` process, with uv's
+        configuration discovery anchored to the editable checkout.
     """
     checkout = _checkout_root()
     args = [
         uv_path,
         "pip",
         "install",
+        "--directory",
+        str(checkout),
         "--python",
         sys.executable,
         "-e",

{deepagents_code-0.1.72 → deepagents_code-0.1.73}/deepagents_code/_paths.py

@@ -861,7 +861,12 @@ def _installation_paths() -> InstallationPaths:
     """
     root = _normalize_absolute(Path(sys.prefix), what="sys.prefix")
     resources = root / "share" / "deepagents-code"
-    locks = root.parent / f".{root.name}.deepagents-code-locks"
+    # uv enumerates every directory beside the tool environment as a package.
+    # Keep persistent locks beside that container instead, outside the environment
+    # uv replaces on upgrade. Mirror this in install.sh's acquire_install_lock.
+    locks = (
+        root.parent.parent / f".{root.parent.name}.{root.name}.deepagents-code-locks"
+    )
     return InstallationPaths(
         root=root,
         managed_bin_dir=resources / "bin",

{deepagents_code-0.1.72 → deepagents_code-0.1.73}/deepagents_code/_server_config.py

@@ -79,6 +79,62 @@ execute one directory's configuration against another directory's trust
 decision.
 """
 
+MODEL_COMPATIBLE_FIELDS = frozenset(
+    {
+        "auto_classifier_model",
+        "cli_max_retries",
+        "model",
+        "rubric_model",
+        "summarization_model",
+    }
+)
+"""Cosmetic model settings that do not affect durable access-policy compat.
+
+Changing which model serves a bound thread — or the retry budget or auxiliary
+grader/classifier models — does not change what the thread is *allowed* to do
+(trust, tool, sandbox, and approval policy are unchanged). These fields are
+therefore excluded from the policy fingerprint, so a model switch does not
+invalidate a durable binding. They remain part of the runtime fingerprint so
+the runtime rebuilds. `None` and an empty/unset value are normalized as
+equivalent here.
+"""
+
+RUNTIME_ONLY_FIELDS = frozenset(
+    {
+        "assistant_id",
+        "model_params",
+        "profile_overrides",
+        "recursion_limit",
+        "rubric_max_iterations",
+        "system_prompt",
+    }
+)
+"""Runtime-identity fields that are not durable access policy.
+
+These shape *how* the runtime executes (model kwargs, prompt, step budget,
+graph id) but are not the trust/tool/sandbox/approval policy a binding guards.
+They are fingerprinted into the runtime identity (so changing them rebuilds the
+runtime) yet never persisted or reported — `model_params`, `profile_overrides`,
+and `system_prompt` can carry secrets and prompt material.
+"""
+
+WORKSPACE_IDENTITY_FIELDS = frozenset({"cwd", "project_root"})
+"""Fields that name *which* workspace a binding is for.
+
+Workspace identity is part of durable access-policy compatibility: a binding
+is meaningless if it can be replayed against a different directory. These are
+compared exactly (canonical paths) and are never normalized away.
+"""
+
+
+def _classified_fields() -> frozenset[str]:
+    """Return every field the policy/runtime split accounts for.
+
+    Returns:
+        The union of all classification sets plus workspace identity.
+    """
+    return MODEL_COMPATIBLE_FIELDS | RUNTIME_ONLY_FIELDS | WORKSPACE_IDENTITY_FIELDS
+
 
 def _same_workspace_project(first: str | None, second: str) -> bool:
     """Whether two paths name the same project directory.
@@ -636,8 +692,34 @@ class ServerConfig:
             return replace(self, trust_project_extensions=False)
         return self
 
-    def workspace_fingerprint(self) -> str:
-        """Fingerprint the resolved runtime config except workspace identity.
+    def policy_fingerprint(self) -> str:
+        """Fingerprint durable access policy plus workspace identity.
+
+        Covers the trust/tool/sandbox/approval payload (the
+        `to_workspace_payload()` keys) and the workspace's `cwd`/`project_root`.
+        Cosmetic model settings (`MODEL_COMPATIBLE_FIELDS`) are excluded so a
+        model switch does not invalidate a durable binding; runtime-only fields
+        (`RUNTIME_ONLY_FIELDS`) are excluded because they are not persisted.
+
+        Returns:
+            The canonical SHA-256 fingerprint.
+        """
+        from deepagents_code.workspace import canonical_fingerprint
+
+        return canonical_fingerprint(
+            {
+                "cwd": self.cwd,
+                "policy": self.to_workspace_payload(),
+                "project_root": self.project_root,
+            }
+        )
+
+    def runtime_fingerprint(self) -> str:
+        """Fingerprint the full runtime identity, including model settings.
+
+        Any change to model, model params, prompt, policy, or runtime fields
+        yields a different value, so the runtime cache can rebuild on any of
+        them while the binding only guards the policy fingerprint.
 
         Returns:
             The canonical SHA-256 fingerprint.
@@ -649,6 +731,16 @@ class ServerConfig:
 
         return canonical_fingerprint(values)
 
+    def workspace_fingerprint(self) -> str:
+        """Fingerprint the resolved runtime config except workspace identity.
+
+        Retained for compatibility; equivalent to `runtime_fingerprint()`.
+
+        Returns:
+            The canonical SHA-256 fingerprint.
+        """
+        return self.runtime_fingerprint()
+
     def __post_init__(self) -> None:
         """Normalize fields and validate invariants.
 
@@ -1015,3 +1107,38 @@ def _normalize_path(
             "Ensure the path exists and is accessible."
         )
         raise ValueError(msg) from exc
+
+
+def classified_config_fields() -> dict[str, str]:
+    """Classify every `ServerConfig` field for the policy/runtime split.
+
+    Every dataclass field must land in exactly one bucket so the fingerprint
+    split cannot silently drop a setting. A field missing from a bucket is a
+    programming error and fails closed (raises) rather than being treated as
+    either policy or runtime by default.
+
+    Returns:
+        A mapping of field name to one of `"model"`, `"runtime"`,
+        `"identity"`, or `"policy"`.
+
+    Raises:
+        RuntimeError: If a `ServerConfig` field is not classified.
+    """
+    import dataclasses
+
+    policy_fields = set(ServerConfig().to_workspace_payload())
+    classification: dict[str, str] = {}
+    for field_info in dataclasses.fields(ServerConfig):
+        name = field_info.name
+        if name in MODEL_COMPATIBLE_FIELDS:
+            classification[name] = "model"
+        elif name in RUNTIME_ONLY_FIELDS:
+            classification[name] = "runtime"
+        elif name in WORKSPACE_IDENTITY_FIELDS:
+            classification[name] = "identity"
+        elif name in policy_fields:
+            classification[name] = "policy"
+        else:
+            msg = f"ServerConfig field {name!r} is not classified"
+            raise RuntimeError(msg)
+    return classification

{deepagents_code-0.1.72 → deepagents_code-0.1.73}/deepagents_code/_session_stats.py

@@ -14,6 +14,7 @@ from __future__ import annotations
 
 import logging
 import math
+import re
 import sys
 from collections.abc import Hashable, Mapping
 from dataclasses import dataclass, field, replace
@@ -176,22 +177,24 @@ class RecordedUsage:
     """
 
 
-UsageLedgerKey = str | tuple[Hashable, str]
-"""Key of the recorded-request ledger: a message ID, optionally scoped.
+@dataclass(frozen=True, slots=True)
+class ModelInvocationKey:
+    """Stable identity of one model invocation in the usage ledger."""
 
-A bare message-ID string keys requests recorded without an attempt scope
-(the legacy behavior). When a caller passes `attempt_scope` to
-`record_message_usage`/`record_model_usage_event`, the key is
-`(attempt_scope, message_id)` instead, so a retry that reuses the provider's
-message ID is a separate request while chunks of one attempt still merge.
-Callers that retain this ledger across retries should use this widened key type.
+    invocation_id: str
+
+
+@dataclass(frozen=True, slots=True)
+class MessageUsageKey:
+    """Message identity, optionally scoped to one model attempt."""
+
+    message_id: str
+    attempt_scope: Hashable | None = None
+
+
+UsageLedgerKey = ModelInvocationKey | MessageUsageKey
+"""Model run identities and message aliases for recorded requests."""
 
-The two shapes coexist in one ledger, so de-duplication cannot rely on the
-key alone: an attempt scope lives for one model call, while a HITL resume pass
-replays messages with no scope open. `finalize_recorded_requests` closes that
-gap by projecting every scoped key down to its bare message ID at each round
-boundary -- see its docstring.
-"""
 
 ModelStatsKey = tuple[str, str]
 """Per-model dict key: the `(provider, model_name)` pair.
@@ -453,67 +456,67 @@ class RecordedRequest:
     round. A later chunk for a finalized request is a replay, not a revision.
     """
 
+    message_ids: frozenset[str] = frozenset()
+    """Provider and framework message IDs observed for this invocation."""
 
-def finalize_recorded_requests(
-    recorded_requests: dict[UsageLedgerKey, RecordedRequest],
-) -> None:
-    """Close every request in a ledger that outlives its stream round.
-
-    A chunked request is left open between chunks so later ones can revise it.
-    That is only correct *within* one round: when a consumer reuses its ledger
-    across HITL resume passes, the replayed chunks of an already-recorded
-    request would otherwise merge a second time and double its tokens and cost.
-    Closing the ledger at each round boundary makes the replay indistinguishable
-    from the stray-chunk case `record_message_usage` already rejects.
-
-    Attempt-scoped keys need one extra step. A scope identifies one model
-    attempt and is closed when that attempt ends, so the replay on the next
-    resume pass arrives with no scope and keys by the bare message ID -- which
-    would miss the `(attempt_scope, message_id)` entry entirely and count the
-    whole request a second time. Project each scoped entry down to its bare
-    message ID as well, so the replay finds a finalized row whichever shape it
-    keys by. Later attempts overwrite earlier ones, leaving the values from the
-    attempt that actually succeeded; the projected row exists only to reject
-    replays, so its counts are never added to `stats` again.
+    ledger_key: UsageLedgerKey | None = None
+    """Canonical ledger entry, promoted when a model run identity is learned."""
 
-    Args:
-        recorded_requests: Ledger to close. Mutated in place.
-    """
-    for request_id, recorded in list(recorded_requests.items()):
-        closed = recorded if recorded.finalized else replace(recorded, finalized=True)
-        recorded_requests[request_id] = closed
-        if isinstance(request_id, tuple):
-            recorded_requests[request_id[1]] = closed
+    request_bucket: str | None = None
+    """Stable provisional bucket even when a legacy record gains a run identity."""
 
+    invocation_id: str | None = None
+    """Strong run identity, when supplied by either delivery path."""
 
-def _provisional_bucket_key(request_id: UsageLedgerKey | None) -> str | None:
-    """Return a string key naming the request a provisional delta belongs to.
+    aliases: frozenset[UsageLedgerKey] = frozenset()
+    """Scoped identities pointing to this canonical record."""
 
-    The consumer keys a pool of provisional dollars by this value, so it must
-    separate exactly what the ledger separates. An attempt-scoped key therefore
-    keeps its scope: two attempts that reuse one provider message ID are
-    distinct requests, and collapsing them to the bare ID would let one
-    attempt's late retraction draw on the other attempt's deposit.
 
-    Scope cannot change under a request while it still reports deltas. A chunk
-    or completion whose record is already finalized is rejected before any
-    delta is built, and `finalize_recorded_requests` closes every entry at the
-    end of a stream round -- so one request's deltas all carry one scope, and
-    the key stays stable for as long as the pool holds its money.
+def finalize_recorded_requests(
+    recorded_requests: dict[UsageLedgerKey, RecordedRequest],
+) -> None:
+    """Finalize canonical records and expose message aliases for unscoped replay."""
+    for request_id, recorded in list(recorded_requests.items()):
+        if request_id != recorded.ledger_key:
+            continue
+        closed = replace(recorded, finalized=True)
+        for alias in closed.aliases:
+            if recorded_requests.get(alias) is recorded:
+                recorded_requests[alias] = closed
+        for message_id in closed.message_ids:
+            recorded_requests[MessageUsageKey(message_id)] = closed
 
-    Args:
-        request_id: Ledger key for the request, or `None` when the message
-            carried no usable ID.
 
-    Returns:
-        An opaque bucket key, or `None` when there is no ledger key to name.
-    """
+def _provisional_bucket_key(request_id: UsageLedgerKey | None) -> str | None:
+    """Return the initial provisional bucket, retained when run identity improves."""
     if request_id is None:
         return None
-    if isinstance(request_id, tuple):
-        attempt_scope, message_id = request_id
-        return f"{attempt_scope!r}\x00{message_id}"
-    return request_id
+    if isinstance(request_id, ModelInvocationKey):
+        return repr(request_id.invocation_id)
+    if request_id.attempt_scope is not None:
+        return repr((request_id.attempt_scope, request_id.message_id))
+    return request_id.message_id
+
+
+def _usage_ledger_key(
+    recorded_requests: dict[UsageLedgerKey, RecordedRequest],
+    message_id: str | None,
+    attempt_scope: Hashable | None,
+    invocation_id: str | None,
+) -> tuple[UsageLedgerKey | None, frozenset[UsageLedgerKey]]:
+    """Return a canonical key and compatible scoped aliases, never across runs."""
+    message_key = MessageUsageKey(message_id, attempt_scope) if message_id else None
+    run_key = ModelInvocationKey(invocation_id) if invocation_id else None
+    aliases = frozenset(key for key in (run_key, message_key) if key is not None)
+    for key in (run_key, message_key):
+        previous = recorded_requests.get(key) if key is not None else None
+        if previous is not None and (
+            not invocation_id
+            or not previous.invocation_id
+            or previous.invocation_id == invocation_id
+        ):
+            return previous.ledger_key, aliases
+    return run_key or message_key, aliases
 
 
 def _names_a_model(message: object) -> bool:
@@ -783,6 +786,7 @@ def _move_request_to_named_model(
         cost_usd=cost_usd,
         usage_metadata=previous.usage_metadata,
         finalized=previous.finalized,
+        message_ids=previous.message_ids,
     )
     return _cost_delta(
         previous.cost_usd,
@@ -887,6 +891,12 @@ def _finalize_from_completed(
         cost_usd=cost_usd,
         usage_metadata=usage,
         finalized=True,
+        message_ids=previous.message_ids
+        | (
+            frozenset({message_id})
+            if (message_id := getattr(message, "id", None))
+            else frozenset()
+        ),
     )
     return RecordedUsage(
         input_tokens=input_count - previous.input_tokens,
@@ -913,6 +923,7 @@ def record_message_usage(
     kind: UsageKind = "assistant",
     recorded_requests: dict[UsageLedgerKey, RecordedRequest] | None = None,
     attempt_scope: Hashable | None = None,
+    invocation_id: str | None = None,
 ) -> RecordedUsage | None:
     """Record usage attached to one streamed model message.
 
@@ -947,13 +958,12 @@ def record_message_usage(
             this specific request, when available.
         kind: Request class used by the type breakdown.
         recorded_requests: Ledger of requests this stream consumer has already
-            recorded, keyed by message ID -- or by `(attempt_scope, message_id)`
-            when `attempt_scope` is set. Mutated in place.
-        attempt_scope: Identity of the attempt that produced this message, or
-            `None` for legacy unscoped recording. Retries of one logical request
-            can reuse the provider's message ID; scoping keeps each attempt's
-            usage separate, while chunks and corrections of one attempt (same
-            scope, same ID) still merge into a single request.
+            recorded, keyed by stable invocation identity when available and by
+            message ID otherwise. Mutated in place.
+        attempt_scope: Lifecycle identity of the attempt that produced this
+            message. Keeps retries distinct when no model run ID is available.
+        invocation_id: LangChain model run ID shared by streaming and callback
+            deliveries for this invocation. Preferred over provider message IDs.
 
     Returns:
         The tokens and cost *this message* contributed, or `None` when it has no
@@ -966,18 +976,89 @@ def record_message_usage(
     if not isinstance(usage, Mapping) or not usage:
         return None
 
-    # Imported here, not at module scope: this module is deliberately free of
-    # heavy top-level dependencies (see the module docstring).
-    from langchain_core.messages import AIMessageChunk
+    from deepagents_code.cost_tracking import _MODEL_INVOCATION_METADATA_KEY
 
     if recorded_requests is None:
         recorded_requests = {}
-    message_id = getattr(message, "id", None)
-    request_id: UsageLedgerKey | None = (
-        message_id if isinstance(message_id, str) and message_id else None
+    raw_message_id = getattr(message, "id", None)
+    message_id = (
+        raw_message_id if isinstance(raw_message_id, str) and raw_message_id else None
+    )
+    metadata_id = (request_metadata or {}).get(_MODEL_INVOCATION_METADATA_KEY)
+    invocation_id = invocation_id or (
+        metadata_id if isinstance(metadata_id, str) and metadata_id else None
+    )
+    if not invocation_id and message_id and message_id.startswith("lc_run--"):
+        # Non-streaming generations append an index to the model run UUID.
+        # Preserve the existing inference for IDs without a generation suffix.
+        match = re.fullmatch(
+            r"lc_run--([0-9a-fA-F]{8}-(?:[0-9a-fA-F]{4}-){3}[0-9a-fA-F]{12})"
+            r"(?:-\d+)?",
+            message_id,
+        )
+        invocation_id = match[1] if match else message_id.removeprefix("lc_run--")
+    request_id, aliases = _usage_ledger_key(
+        recorded_requests, message_id, attempt_scope, invocation_id
     )
-    if request_id is not None and attempt_scope is not None:
-        request_id = (attempt_scope, request_id)
+    previous = recorded_requests.get(request_id) if request_id is not None else None
+    request_bucket = (
+        previous.request_bucket if previous else _provisional_bucket_key(request_id)
+    )
+    if previous is not None and invocation_id:
+        request_id = ModelInvocationKey(invocation_id)
+        recorded_requests[request_id] = previous
+    result = _record_message_usage(
+        stats,
+        message,
+        fallback_model=fallback_model,
+        fallback_provider=fallback_provider,
+        request_metadata=request_metadata,
+        kind=kind,
+        recorded_requests=recorded_requests,
+        request_id=request_id,
+    )
+    if request_id is not None and (updated := recorded_requests.get(request_id)):
+        aliases |= (
+            frozenset(
+                alias
+                for alias in previous.aliases
+                if recorded_requests.get(alias) is previous
+            )
+            if previous
+            else frozenset()
+        )
+        updated = replace(
+            updated,
+            ledger_key=request_id,
+            request_bucket=request_bucket,
+            invocation_id=invocation_id
+            or (previous.invocation_id if previous else None),
+            aliases=aliases | {request_id},
+            message_ids=updated.message_ids | ({message_id} if message_id else set()),
+        )
+        for alias in updated.aliases:
+            recorded_requests[alias] = updated
+    return replace(result, request_id=request_bucket) if result else None
+
+
+def _record_message_usage(
+    stats: SessionStats,
+    message: object,
+    *,
+    fallback_model: str,
+    fallback_provider: str,
+    request_metadata: Mapping[str, Any] | None,
+    kind: UsageKind,
+    recorded_requests: dict[UsageLedgerKey, RecordedRequest],
+    request_id: UsageLedgerKey | None,
+) -> RecordedUsage | None:
+    """Return the usage delta applied to the resolved canonical request."""
+    usage = getattr(message, "usage_metadata", None)
+    if not isinstance(usage, Mapping) or not usage:
+        return None
+
+    from langchain_core.messages import AIMessageChunk
+
     is_chunk = isinstance(message, AIMessageChunk)
     if request_id is not None and request_id in recorded_requests and not is_chunk:
         previous = recorded_requests[request_id]
@@ -1102,6 +1183,7 @@ def record_message_usage(
             cost_usd=cost_usd,
             usage_metadata=accumulated_usage,
             finalized=not is_chunk,
+            message_ids=previous.message_ids if previous else frozenset(),
         )
     # Provisional pricing needs this message's delta, while the context display
     # needs the request's running token total after folding the message in.
@@ -1176,6 +1258,7 @@ def record_model_usage_event(
         )
         return None
     request_id = data.get("request_id")
+    invocation_id = data.get("invocation_id")
     thread_id = data.get("thread_id")
     scope = data.get("scope")
     usage = data.get("usage_metadata")
@@ -1183,6 +1266,7 @@ def record_model_usage_event(
     if (
         not isinstance(request_id, str)
         or not request_id
+        or (invocation_id is not None and not isinstance(invocation_id, str))
         or not isinstance(thread_id, str)
         or not thread_id
         or not isinstance(scope, str)
@@ -1220,6 +1304,7 @@ def record_model_usage_event(
         kind="subagent",
         recorded_requests=recorded_requests,
         attempt_scope=attempt_scope,
+        invocation_id=invocation_id or None,
     )