argus-code-review 0.2.7__tar.gz → 0.2.9__tar.gz

{argus_code_review-0.2.7 → argus_code_review-0.2.9}/CHANGELOG.md

@@ -7,6 +7,30 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [0.2.9] - 2026-10-02
+
+### Changed
+
+- Upgraded frontier and default model aliases in `argus.llm.models` (TECH-7124, #31):
+  - `claude-frontier` and `claude-opus` upgraded to `claude-opus-5-5` (previously `claude-fable-5` and `claude-opus-5`).
+  - `claude-default` upgraded to `claude-sonnet-5-5` (previously `claude-sonnet-4-6`).
+  - `gpt-frontier` upgraded to `gpt-6.1-sol` (previously `gpt-5.6-sol`).
+- Updated model pricing rates in `argus.llm.models` for `claude-sonnet-5-5`, `claude-opus-5-5`, and `gpt-6.1-sol` (TECH-7124, #31).
+- Re-verified the Anthropic 1M context beta (`context-1m-2025-08-07`) against `claude-sonnet-5-5` via the Redesign Health Anthropic proxy (TECH-7124, #31).
+- Updated approved model policies in `pr-review-specialist-llm-patterns.md`, runner warning logs, bench defaults, and CLI help text for the new model aliases (TECH-7124, #31).
+
+## [0.2.8] - 2026-10-02
+
+### Changed
+
+- Raised the Claude-path turn budget from 30 to 50 and the default reviewer session
+  timeout from 900 to 1500 seconds globally across all runner platforms through
+  `ARGUS_SESSION_TIMEOUT` (TECH-7093), preserving headroom for longer sessions.
+- Added Claude's budget disclosure and supported `PostToolUse`/`PostToolUseFailure`
+  hook-based convergence warnings (derived from budget constants; 37 and 47 tool calls
+  by default), using truthful top-level tool-call counts, including failed calls.
+- Decoupled the Claude and OpenAI turn-budget constants; OpenAI remains at 30 turns.
+
 ## [0.2.7] - 2026-09-26
 
 ### Fixed
@@ -356,7 +380,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   packaged set.
 - `argus --version`, `argus prompts list`, and `argus prompts export`.
 
-[Unreleased]: https://github.com/redesignhealth/argus-review/compare/v0.2.7...HEAD
+[Unreleased]: https://github.com/redesignhealth/argus-review/compare/v0.2.9...HEAD
+[0.2.9]: https://github.com/redesignhealth/argus-review/compare/v0.2.8...v0.2.9
+[0.2.8]: https://github.com/redesignhealth/argus-review/compare/v0.2.7...v0.2.8
 [0.2.7]: https://github.com/redesignhealth/argus-review/compare/v0.2.6...v0.2.7
 [0.2.6]: https://github.com/redesignhealth/argus-review/compare/v0.2.5...v0.2.6
 [0.2.5]: https://github.com/redesignhealth/argus-review/compare/v0.2.4...v0.2.5

{argus_code_review-0.2.7 → argus_code_review-0.2.9}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: argus-code-review
-Version: 0.2.7
+Version: 0.2.9
 Summary: Self-orchestrated PR review agent using LangGraph + Claude Agent SDK
 Project-URL: Repository, https://github.com/redesignhealth/argus-review
 Project-URL: Issues, https://github.com/redesignhealth/argus-review/issues
@@ -204,8 +204,7 @@ argus review owner/repo --pr 123 --dismiss "B2 -- pre-existing, not from this PR
 
 # Override the reviewer models (or set ARGUS_SPECIALIST_MODEL/ARGUS_FRONTIER_MODEL instead).
 # --frontier-model controls both the planner/coverage tier AND the cross-cutting
-# reviewer -- claude-fable-5 here is already the planner/coverage default, but
-# it also moves cross-cutting OFF its cheaper claude-opus-5 default onto fable-5.
+# reviewer (defaults to claude-opus-5-5).
 # Cost note: --specialist-model here overrides the bulk reviewer path (system
 # reviewer, specialists, tests-and-docs) from its low-cost Gemini default onto
 # claude-sdk with the specified model, and also updates the writer and lite-review
@@ -214,7 +213,7 @@ argus review owner/repo --pr 123 --dismiss "B2 -- pre-existing, not from this PR
 # the 1M-context beta from that same highest-volume path, since the beta is
 # only verified against the unoverridden default (autocompact may thrash on
 # long reviews under this override -- see argus/runners.py for the tradeoff).
-argus review owner/repo --pr 123 --specialist-model claude-opus-5 --frontier-model claude-fable-5
+argus review owner/repo --pr 123 --specialist-model claude-opus-5-5 --frontier-model claude-opus-5-5
 
 # Clear an already-set ARGUS_SPECIALIST_MODEL/ARGUS_FRONTIER_MODEL for just this
 # run by passing an empty string -- useful when the env var is set globally

{argus_code_review-0.2.7 → argus_code_review-0.2.9}/README.md

@@ -153,8 +153,7 @@ argus review owner/repo --pr 123 --dismiss "B2 -- pre-existing, not from this PR
 
 # Override the reviewer models (or set ARGUS_SPECIALIST_MODEL/ARGUS_FRONTIER_MODEL instead).
 # --frontier-model controls both the planner/coverage tier AND the cross-cutting
-# reviewer -- claude-fable-5 here is already the planner/coverage default, but
-# it also moves cross-cutting OFF its cheaper claude-opus-5 default onto fable-5.
+# reviewer (defaults to claude-opus-5-5).
 # Cost note: --specialist-model here overrides the bulk reviewer path (system
 # reviewer, specialists, tests-and-docs) from its low-cost Gemini default onto
 # claude-sdk with the specified model, and also updates the writer and lite-review
@@ -163,7 +162,7 @@ argus review owner/repo --pr 123 --dismiss "B2 -- pre-existing, not from this PR
 # the 1M-context beta from that same highest-volume path, since the beta is
 # only verified against the unoverridden default (autocompact may thrash on
 # long reviews under this override -- see argus/runners.py for the tradeoff).
-argus review owner/repo --pr 123 --specialist-model claude-opus-5 --frontier-model claude-fable-5
+argus review owner/repo --pr 123 --specialist-model claude-opus-5-5 --frontier-model claude-opus-5-5
 
 # Clear an already-set ARGUS_SPECIALIST_MODEL/ARGUS_FRONTIER_MODEL for just this
 # run by passing an empty string -- useful when the env var is set globally

{argus_code_review-0.2.7 → argus_code_review-0.2.9}/argus/bench_default.toml

@@ -33,9 +33,7 @@ caching  = "auto"
 
 [roles.cross-cutting]
 platform    = "claude-sdk"
-# Opus, not frontier (Fable): matches argus.runners._CROSS_CUTTING_MODEL --
-# evals showed no measurable quality gain from frontier on this stage, at
-# ~2x the per-token cost.
+# Matches argus.runners._CROSS_CUTTING_MODEL (resolves to claude-opus-5-5).
 model       = "claude-opus"
 prompt_name = "pr-review-cross-cutting"
 caching     = "auto"

{argus_code_review-0.2.7 → argus_code_review-0.2.9}/argus/cli.py

@@ -368,10 +368,7 @@ def _add_review_args(parser: argparse.ArgumentParser) -> None:
         default=_MODEL_OVERRIDE_UNSET,
         help=(
             "Override the model used by the planner, coverage check, and "
-            "cross-cutting reviewer (defaults: claude-fable-5 for the first "
-            "two, claude-opus-5 for cross-cutting -- note this one flag "
-            "collapses both onto the SAME model when set, moving "
-            "cross-cutting off its cheaper default; or ARGUS_FRONTIER_MODEL "
+            "cross-cutting reviewer (default: claude-opus-5-5, or ARGUS_FRONTIER_MODEL "
             "if already set in the environment). Same effect as setting "
             "ARGUS_FRONTIER_MODEL. Pass an empty string to clear an "
             "already-set ARGUS_FRONTIER_MODEL for this run."

{argus_code_review-0.2.7 → argus_code_review-0.2.9}/argus/config.py

@@ -2,7 +2,7 @@
 
 A slim ``pydantic-settings`` implementation. All configuration comes from
 environment variables (or a local ``.env`` file loaded via
-``argus.dotenv_utils``) — no AWS, no SSM.
+``argus.dotenv_utils``) -- no AWS, no SSM.
 """
 
 from __future__ import annotations
@@ -15,7 +15,7 @@ from pydantic_settings import BaseSettings, SettingsConfigDict
 # ``argus.runners._SUBPROCESS_TIMEOUT_S`` (the fallback used when no Settings
 # instance is available, mostly tests) imports this same constant rather than
 # hardcoding its own copy, so the two can never drift out of sync.
-DEFAULT_ARGUS_SESSION_TIMEOUT_S = 900
+DEFAULT_ARGUS_SESSION_TIMEOUT_S = 1500
 
 
 class Settings(BaseSettings):
@@ -62,7 +62,7 @@ class Settings(BaseSettings):
             ``argus.precheck``.
         ARGUS_STOCK_SEMGREP_PACKS: Comma-separated semgrep registry pack IDs
             (e.g. ``"p/secrets"``) to run alongside (or instead of) a custom
-            ``ARGUS_RULES_DIR`` — unlike a local rules directory, each pack
+            ``ARGUS_RULES_DIR`` -- unlike a local rules directory, each pack
             is fetched over the network by semgrep itself on first use
             (cached locally after). Unset by default: this is an opt-in
             addition of vetted, community-maintained rules, not a silent
@@ -74,8 +74,8 @@ class Settings(BaseSettings):
             (``PrecheckResult.failed_scanners``/``missing_scanners``
             non-empty) instead of the default fail-open
             behavior (surface it in the review comment's degraded-coverage
-            section — see ``argus.helpers.build_degraded_coverage_labels``
-            — but let the review's own verdict stand on its own merits).
+            section -- see ``argus.helpers.build_degraded_coverage_labels``
+            -- but let the review's own verdict stand on its own merits).
             False by default: every other part of this module's design is
             deliberately fail-open (a broken scanner should never be the
             reason a review can't complete), and this flag exists for
@@ -91,7 +91,7 @@ class Settings(BaseSettings):
         ARGUS_SPECIALIST_MODEL: Override the model used by the system
             reviewer, specialist reviewers, the writer, and the lite-review
             path (``argus.llm.models.CLAUDE_DEFAULT``, default
-            ``claude-sonnet-4-6``). Set via ``--specialist-model``; read
+            ``claude-sonnet-5-5``). Set via ``--specialist-model``; read
             directly from ``os.environ`` by ``argus.llm.models`` at import
             time and by ``argus.bench`` (where setting it forces bulk reviewers
             to claude-sdk with claude-default). Also read off a ``Settings``
@@ -117,14 +117,21 @@ class Settings(BaseSettings):
             Context7 call.
         ARGUS_SESSION_TIMEOUT: Wall-clock seconds a reviewer subprocess is
             allowed to run before it is killed and reported as a failure.
-            Defaults to 900 (15 minutes) — first raised from 300 to 420 after
+            Defaults to 1500 (25 minutes) -- first raised from 300 to 420 after
             production logs showed legitimate (non-runaway) specialist
             reviewers finishing as late as 294s, right at the old timeout's
             edge; raised again to 600 to match rh-data-platform's
             production-proven value ahead of this package taking over as the
             actual production reviewer (rh-data-platform's review_service is
-            being retired in its favor); raised once more to 900 for
-            additional headroom across all three reviewer platforms.
+            being retired in its favor); raised to 900 for additional headroom
+            across all three reviewer platforms; raised to 1500 in TECH-7093.
+            This timeout is deliberately global to all runner platforms through
+            ARGUS_SESSION_TIMEOUT, rather than scoped only to Claude, to preserve
+            a single operator control across runners. In production, an observed
+            cross-cutting session ran 533.8s under the 30-turn cap; linear
+            scaling 30 to 50 projects ~890s, meaning 900s leaves virtually zero
+            headroom. 1500s preserves roughly 1.69x margin, critical because a
+            timeout kill discards all reviewer output.
         GOOGLE_API_KEY: Gemini platform credential, consumed via the
             ``google_credential`` property by ``argus.gemini_runner``
             whenever a role's bench entry resolves to
@@ -206,7 +213,7 @@ class Settings(BaseSettings):
 
         Callers that need to hand this credential to something that itself
         reads an env var (the spawned ``claude`` CLI subprocess) must set
-        the SAME variable name the caller configured — forcing everything
+        the SAME variable name the caller configured -- forcing everything
         to ``ANTHROPIC_API_KEY`` would send a proxy/gateway bearer token as
         an ``x-api-key``, which not every gateway accepts. ``ANTHROPIC_API_KEY``
         wins when both are set, matching the Anthropic SDK's own precedence.

{argus_code_review-0.2.7 → argus_code_review-0.2.9}/argus/gemini_runner.py

@@ -41,8 +41,9 @@ function call the model requests each turn (there can be more than
 one), feeding all of their results back as a single follow-up turn --
 until the model stops requesting function calls, calls
 ``finish_review``, or the turn budget (``_MAX_TURNS_GEMINI``, this
-module's own budget, independent of the Claude/OpenAI paths' shared
-``argus.runners._MAX_TURNS``) is exhausted. Findings arrive via
+module's own budget, independent of the Claude path's
+``argus.runners._MAX_TURNS_CLAUDE`` and the OpenAI path's
+``argus.openai_runner._MAX_TURNS_OPENAI``) is exhausted. Findings arrive via
 ``report_finding`` tool calls into ``review_tools``' per-session sink,
 not as a JSON blob embedded in the model's own text -- so, to keep this
 task's blast radius contained to this file plus ``argus.bench``'s

{argus_code_review-0.2.7 → argus_code_review-0.2.9}/argus/llm/models.py

@@ -13,9 +13,9 @@ gpt-5.5 was NOT on the approved model list (see
 gpt-5.4, with a regression-guard test
 (``test_gpt_frontier_pinned_to_approved_model`` in
 ``tests/test_llm_models.py``) added specifically to catch a repeat. The
-alias has since been bumped again, this time to ``gpt-5.6-sol`` -- unlike
-the gpt-5.5 attempt, the ``gpt-5.6`` family genuinely IS on the approved
-model list (the entire family is approved in
+alias has since been bumped again, this time to ``gpt-6.1-sol`` (TECH-7124) -- unlike
+the gpt-5.5 attempt, the ``gpt-6.1-sol`` model genuinely IS on the approved
+model list (the model is approved in
 ``pr-review-specialist-llm-patterns.md`` alongside this change), so this is
 not a repeat of that mistake. Any future bump of this alias must likewise
 confirm the target model is on the approved list -- and update the policy
@@ -48,11 +48,15 @@ Per-token model pricing is sourced centrally from ``argus.llm.pricing``
 
 Tier semantics:
     *frontier* -- best reasoning available in the family; slow / expensive.
-    *opus*     -- next tier down from frontier -- strong reasoning at roughly
-                  half frontier's per-token cost. Introduced for call sites
-                  (e.g. the cross-cutting reviewer) where evals showed no
-                  measurable quality gain from frontier, so the cost isn't
-                  justified.
+                  Under Claude 5.5, ``claude-frontier`` and ``claude-opus``
+                  intentionally share ``claude-opus-5-5``.
+    *opus*     -- historically the next tier down from frontier -- strong reasoning
+                  at roughly half frontier's per-token cost (e.g. Opus 5 vs Fable 5).
+                  Under Claude 5.5, ``claude-frontier`` and ``claude-opus``
+                  intentionally share ``claude-opus-5-5``, so the prior cost
+                  differentiation rationale no longer applies while preserving
+                  the distinct alias keys for call sites (e.g. the cross-cutting
+                  reviewer) and override compatibility.
     *default*  -- workhorse balance of cost and capability.
     *mini*     -- fast and cheap; suitable for high-volume, low-stakes calls.
 
@@ -63,9 +67,10 @@ Runtime overrides:
 
     ``ARGUS_FRONTIER_MODEL`` (``--frontier-model``) overrides both
     ``CLAUDE_FRONTIER`` (planner, coverage) and ``CLAUDE_OPUS``
-    (cross-cutting) -- there is a single frontier-tier knob at the CLI, even
-    though the two aliases keep independent defaults so cross-cutting still
-    runs on the cheaper Opus tier when no override is given.
+    (cross-cutting) -- there is a single frontier-tier knob at the CLI.
+    Under Claude 5.5, both aliases intentionally share the same default
+    (``claude-opus-5-5``), and setting ``ARGUS_FRONTIER_MODEL`` overrides both
+    onto the specified model.
 
     Both env vars must be set before this module is first imported --
     ``cli.py`` sets them from the CLI flags ahead of any deferred import of
@@ -82,21 +87,18 @@ from argus.llm.pricing import estimate_cost_usd
 logger = logging.getLogger(__name__)
 
 ALIAS_MAP: Final[dict[str, str]] = {
-    # OpenAI -- gpt-5.4 / gpt-5.6 families
-    # gpt-frontier is bumped to gpt-5.6-sol, and gpt-mini is bumped to
-    # gpt-5.6-luna, as the gpt-5.6 family is now on the approved model list
-    # (see pr-review-specialist-llm-patterns.md), whose Default column was
-    # updated to gpt-5.6-luna in the same change. gpt-5.5/gpt-5.5-mini
-    # remain NOT on the approved list -- do not bump either alias to them
-    # until the policy table is updated (see the module docstring above).
+    # OpenAI -- gpt-5.4 / gpt-5.6 / gpt-6 families
+    # gpt-frontier is bumped to gpt-6.1-sol (TECH-7124), and gpt-mini is bumped to
+    # gpt-5.6-luna, as both are on the approved model list
+    # (see pr-review-specialist-llm-patterns.md).
     # test_llm_models.py pins both of these exact values as a regression
     # guard.
-    "gpt-frontier": "gpt-5.6-sol",
+    "gpt-frontier": "gpt-6.1-sol",
     "gpt-mini": "gpt-5.6-luna",
     # Anthropic
-    "claude-frontier": "claude-fable-5",
-    "claude-opus": "claude-opus-5",
-    "claude-default": "claude-sonnet-4-6",
+    "claude-frontier": "claude-opus-5-5",
+    "claude-opus": "claude-opus-5-5",
+    "claude-default": "claude-sonnet-5-5",
     "claude-mini": "claude-haiku-4-5",
     # Google -- gemini-3 family. Real call site: argus.gemini_runner
     # (Track 3), dispatched via argus.bench's "gemini" platform.

{argus_code_review-0.2.7 → argus_code_review-0.2.9}/argus/openai_runner.py

@@ -40,7 +40,7 @@ the model requests each turn (there can be more than one), chaining the
 conversation state forward via ``previous_response_id`` and feeding back
 all tool outputs as ``function_call_output`` items -- until the model stops
 requesting function calls, calls ``finish_review``, or the turn budget
-(``argus.runners._MAX_TURNS``) is exhausted. Findings arrive via
+(``_MAX_TURNS_OPENAI``) is exhausted. Findings arrive via
 ``report_finding`` tool calls into ``review_tools``' per-session sink,
 not as a JSON blob embedded in the model's own text -- so, to keep this
 task's blast radius contained, the final ``SessionResult.result_text`` is
@@ -110,7 +110,6 @@ from argus.bench import BenchEntry
 from argus.llm.models import estimate_cost_usd
 from argus.llm.models import resolve as resolve_model_alias
 from argus.runners import (
-    _MAX_TURNS,
     _MID_BUDGET_NUDGE,
     _NUDGE_TURNS_BEFORE_BUDGET,
     _SUBPROCESS_TIMEOUT_S,
@@ -123,6 +122,8 @@ from argus.runners import (
 
 logger = logging.getLogger(__name__)
 
+_MAX_TURNS_OPENAI = 30
+
 _DEFAULT_READ_LIMIT = 2000  # mirrors argus.review_tools._DEFAULT_READ_LIMIT
 
 # Matches the fenced-json extraction argus.helpers.parse_review_result
@@ -417,7 +418,7 @@ async def _run_turns(
     whatever those completed turns actually produced and billed.
     """
     system_prompt = (
-        system_prompt + "\n\n" + _TURN_BUDGET_SYSTEM_PROMPT_LINE.format(max_turns=_MAX_TURNS)
+        system_prompt + "\n\n" + _TURN_BUDGET_SYSTEM_PROMPT_LINE.format(max_turns=_MAX_TURNS_OPENAI)
     )
     api_key = getattr(settings, "OPENAI_API_KEY", None)
     base_url = getattr(settings, "OPENAI_BASE_URL", None)
@@ -504,13 +505,13 @@ async def _run_turns(
         )
 
     exhausted = False
-    _mid_budget_nudge_turn = _compute_mid_budget_nudge_turn(_MAX_TURNS)
+    _mid_budget_nudge_turn = _compute_mid_budget_nudge_turn(_MAX_TURNS_OPENAI)
     try:
         async with asyncio.timeout(timeout_s):
             with review_tools.review_session(repo_root) as findings_sink:
                 previous_response_id: str | None = None
                 tool_outputs: list[dict[str, Any]] = []
-                for _turn in range(_MAX_TURNS):
+                for _turn in range(_MAX_TURNS_OPENAI):
                     create_kwargs: dict[str, Any] = {
                         "model": model,
                         "instructions": system_prompt,
@@ -625,7 +626,7 @@ async def _run_turns(
                     if finished:
                         break
 
-                    if _turn + 1 == _MAX_TURNS - _NUDGE_TURNS_BEFORE_BUDGET:
+                    if _turn + 1 == _MAX_TURNS_OPENAI - _NUDGE_TURNS_BEFORE_BUDGET:
                         tool_outputs.append({"role": "user", "content": _TURN_BUDGET_NUDGE})
 
                     # Not an `elif` -- these are two independent checkpoints
@@ -643,7 +644,7 @@ async def _run_turns(
                     logger.warning(
                         "OpenAI session [%s] exhausted its %d-turn budget without a finish_review call",
                         label or "unlabeled",
-                        _MAX_TURNS,
+                        _MAX_TURNS_OPENAI,
                     )
                     exhausted = True
 

{argus_code_review-0.2.7 → argus_code_review-0.2.9}/argus/prompts/pr-review-specialist-llm-patterns.md

@@ -7,8 +7,8 @@ Only approved model families. Check model strings in code against this table:
 
 | Provider | Approved | Default |
 |----------|----------|---------|
-| Anthropic | claude-opus-4, claude-sonnet-4, claude-haiku-4 (use `CLAUDE_MINI`), claude-fable-5 (use `CLAUDE_FRONTIER`), claude-opus-5 (use `CLAUDE_OPUS`), claude-sonnet-4-6 (use `CLAUDE_DEFAULT`) families; claude-sonnet-5 (the previous `CLAUDE_DEFAULT` pin) remains valid as an explicit `ARGUS_SPECIALIST_MODEL`/`--specialist-model` override value, not as a new hardcoded default | claude-sonnet-4-6 |
-| OpenAI | gpt-5.4 family (gpt-5.4, gpt-5.4-mini) and gpt-5.6 family | gpt-5.6-luna |
+| Anthropic | claude-opus-4, claude-sonnet-4, claude-haiku-4 (use `CLAUDE_MINI`), claude-opus-5-5 (use `CLAUDE_FRONTIER` / `CLAUDE_OPUS`), claude-sonnet-5-5 (use `CLAUDE_DEFAULT`) families; legacy Claude models (claude-fable-5, claude-opus-5, claude-sonnet-4-6, claude-sonnet-5) remain valid only as explicit CLI/env runtime overrides (`ARGUS_SPECIALIST_MODEL`/`ARGUS_FRONTIER_MODEL` or `--specialist-model`/`--frontier-model`) for backward compatibility, not as new hardcoded defaults | claude-sonnet-5-5 |
+| OpenAI | gpt-5.4 family (gpt-5.4, gpt-5.4-mini), gpt-5.6 family, and gpt-6.1-sol (use `GPT_FRONTIER`) | gpt-5.6-luna |
 | Google | gemini-3 family | gemini-3.8-flash |
 
 Flag any use of: o3, o1, gpt-4 family, gpt-5-mini, claude-3/3.5 family, gemini-1.5/2.5 family.