codejury 1.1.0__tar.gz → 1.2.0__tar.gz

{codejury-1.1.0 → codejury-1.2.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: codejury
-Version: 1.1.0
+Version: 1.2.0
 Summary: AI-assisted security review for code diffs and whole repositories.
 Author: AISecLabs
 License-Expression: MIT
@@ -92,26 +92,41 @@ directory. Pass `--dir` to install it somewhere else.
 Set a provider key through flags or environment variables:
 
 ```bash
-export CODEJURY_API_KEY=...
 export CODEJURY_MODEL=claude-opus-4-8
+export CODEJURY_API_KEY=...
 export CODEJURY_API_BASE=...   # optional gateway or proxy
 ```
 
-Repo Review uses a second, deliberately different model for two roles at once. On the recall
-side it finds alongside the main model, so the union takes whatever either catches and a
-single model's blind spot no longer caps recall. On the precision side it must agree before a
-candidate is refuted, so a deletion needs two models with uncorrelated blind spots and no lone
-skeptic drops a real finding. Point it at a cross-vendor model. With none set, both models
-collapse to one, the finder is single-model and the verify stage keeps every candidate, the
-recall-safe default.
+Both review paths name three model roles, finder, challenger, and judge. The finder finds, the
+challenger refutes, the judge confirms before a deletion. Each role defaults to the base
+`--model`, so a single-model run sets only `--model`. Put a different vendor in any seat for
+cross-model review, where uncorrelated blind spots catch what one model misses, and a deletion
+needs the judge to be a distinct model from the challenger so no lone skeptic drops a real
+finding. With the judge not distinct, nothing is refuted, the recall-safe default.
+
+Each role takes a full backend, the same five fields as the base, every one unset by default:
+`CODEJURY_<ROLE>_PROVIDER`, `_MODEL`, `_API_KEY`, `_API_BASE`, `_WIRE_API`, with `<ROLE>` one of
+`FINDER`, `CHALLENGER`, `JUDGE`, and the matching `--<role>-provider`, `--<role>-model`,
+`--<role>-api-key`, `--<role>-api-base`, `--<role>-wire-api` flags. An unset field inherits the
+base, so you set none of these for a single-model run and override only the seat you want to
+change. The key and the api-base inherit the base only when the role keeps the base provider. A
+role that switches vendor brings its own key, since the base key belongs to the base vendor. For
+example a Claude base finder challenged by GPT and confirmed by Claude:
 
 ```bash
-export CODEJURY_CHECKER_PROVIDER=openai   # default
-export CODEJURY_CHECKER_MODEL=...         # a different model from CODEJURY_MODEL
-export CODEJURY_CHECKER_API_KEY=...
-export CODEJURY_CHECKER_API_BASE=...      # optional
+export CODEJURY_CHALLENGER_PROVIDER=openai
+export CODEJURY_CHALLENGER_MODEL=...         # a GPT model, the skeptic
+export CODEJURY_CHALLENGER_API_KEY=...
+export CODEJURY_CHALLENGER_WIRE_API=responses   # the gpt-5 reasoning models speak Responses
+export CODEJURY_JUDGE_MODEL=...              # a Claude model, the confirmer, distinct from the challenger
 ```
 
+The same `CODEJURY_FINDER_*` / `CODEJURY_CHALLENGER_*` / `CODEJURY_JUDGE_*` and the matching
+`--finder-* / --challenger-* / --judge-*` flags work on both `review diff` and `review repo`.
+Note that `review repo --run` finds with one model, the finder, it no longer adds a second
+co-finder, and a seat that runs on the subscription supplies its own review, so it ignores
+that seat's backend flags while the others still apply.
+
 The tool does not auto-load `.env`.
 
 Useful flags:
@@ -121,21 +136,28 @@ Useful flags:
 - `--api-key <key>`
 - `--api-base <url>`
 - `--retries <n>`
+- `--timeout <seconds>`
 
 ## Data Boundary
 
 The tool sends code-derived content to the model provider you configure, so know what
 leaves the machine before reviewing a proprietary repository:
 
-- Diff Review sends the unified diff under review.
-- Repo Review with `--reviewer model` sends bounded source snippets, the detected stack
+- Diff Review on the `api` row sends the unified diff under review. On the `subscription`
+  row it sends the diff in the `claude -p` prompt through your Claude Code account, and the
+  diff agent uses no file tools, so only the diff text leaves the machine, not local files.
+- Under the default `--executor auto`, each seat follows the `api` row when it has a key
+  and the `subscription` row when it falls back to your Claude Code subscription, so what
+  leaves the machine is decided per seat by whether that seat has a key.
+- Repo Review with `--executor api` sends bounded source snippets, the detected stack
   notes, the vulnerability guidance, and the findings.
-- Verification with `--reviewer model` sends the cited source file and the finding
-  details. With `--reviewer claude-cli`, Claude Code receives the finding details and
-  reads the code itself through its read-only tools.
-- `--reviewer claude-cli` does not use the configured provider key. It runs Claude Code
-  with read-only file tools, and Claude Code may send prompts and the code it reads
-  through your Claude Code account, so the code does not stay local.
+- Verification with `--executor api` sends the cited source file and the finding
+  details. On the `subscription` row, Claude Code receives the finding details and reads
+  the code itself through its read-only tools.
+- A repo seat on the `subscription` row does not use the configured provider key. It runs
+  Claude Code with read-only file tools, and Claude Code may send prompts and the code it
+  reads through your Claude Code account, so the code does not stay local. The diff agent is
+  narrower, it reads no files and sees only the diff in the prompt.
 
 A custom `--api-base` or a LiteLLM proxy becomes part of the trust boundary, so the data
 above also reaches that gateway. Prefer the `CODEJURY_API_KEY` environment variable over
@@ -163,8 +185,20 @@ codejury review diff --file changes.diff --mode adversarial
 
 # emit SARIF and fail on HIGH or CRITICAL findings
 codejury review diff --file changes.diff --format sarif --fail-on high
+
+# review with no provider key, riding your Claude Code subscription
+codejury review diff --file changes.diff --executor subscription
+
+# adversarial with a keyless Claude finder and judge plus an OpenAI challenger on its own key
+codejury review diff --file changes.diff --mode adversarial \
+  --challenger-provider openai --challenger-api-key "$OPENAI_API_KEY"
 ```
 
+Diff Review takes the same `--executor auto|api|subscription` as Repo Review, see Review
+Strategy. The default `auto` calls the provider when a seat has a key and falls back to your
+Claude Code subscription for a keyless Anthropic seat. Unlike the repo agent, the diff agent
+answers from the diff in the prompt and reads no files.
+
 `codejury review diff --dry-run` uses a mock provider and a built-in demo diff, so it needs
 no API key.
 
@@ -227,20 +261,26 @@ codejury review repo /path/to/repo --run
 
 A `--run` chooses how each unit is reviewed:
 
-- `--reviewer model` is the default. It makes one grounded model call per unit. Add
-  `--facts` to ground that call in a tool-extracted call graph, storage layout, and read
-  and write sets when the domain binds a facts backend, such as the EVM Slither backend.
-  This is what gives a smart contract review its call relationships, so prefer
-  `--run --facts` on Solidity.
-- `--reviewer claude-cli` runs each unit and its verification as a headless `claude -p`
-  agent that reads and traces the files itself with read-only tools, using your Claude Code
-  access and no provider key. Use it when you want a tool-using agent rather than a single
-  grounded call.
-
-Add `--checker-model`, a different model from `--model`, to enable cross-model
-verification. It finds alongside the main model on the recall side and must agree before a
-finding is dropped on the precision side. With none set, the verify stage refutes nothing,
-the recall-safe default.
+- `--executor auto` is the default. Each seat, the finder and the skeptic, calls the
+  provider when it has a reachable key and falls back to a headless `claude -p`
+  subscription agent for a keyless Anthropic seat, so a keyless run works with no provider
+  key. A keyless non-Anthropic seat, such as an OpenAI finder with no key, is a loud error,
+  it has no subscription to fall back to. This is what lets a Claude finder ride your
+  subscription while an OpenAI challenger uses its own key.
+- `--executor api` makes one grounded model call per unit and requires a key, a missing key
+  is a loud startup error, the same point as auto. Add `--facts` to ground that call in a
+  tool-extracted call graph, storage layout, and read and write sets when the domain binds a
+  facts backend, such as the EVM Slither backend. This is what gives a smart contract review
+  its call relationships, so prefer `--run --facts` on Solidity.
+- `--executor subscription` always runs each unit and its verification as a headless
+  `claude -p` agent that reads and traces the files itself with read-only tools, using your
+  Claude Code access and no provider key. Use it when you want a tool-using agent rather
+  than a single grounded call even where a key is present.
+
+Set a distinct `--judge-model`, the confirmer, from the challenger to enable cross-model
+verification. The challenger refutes a finding and the judge must agree before it is dropped,
+so a deletion needs two models. With the judge not distinct from the challenger, the verify
+stage refutes nothing, the recall-safe default.
 
 ## Supported Knowledge
 

{codejury-1.1.0 → codejury-1.2.0}/README.md

@@ -50,26 +50,41 @@ directory. Pass `--dir` to install it somewhere else.
 Set a provider key through flags or environment variables:
 
 ```bash
-export CODEJURY_API_KEY=...
 export CODEJURY_MODEL=claude-opus-4-8
+export CODEJURY_API_KEY=...
 export CODEJURY_API_BASE=...   # optional gateway or proxy
 ```
 
-Repo Review uses a second, deliberately different model for two roles at once. On the recall
-side it finds alongside the main model, so the union takes whatever either catches and a
-single model's blind spot no longer caps recall. On the precision side it must agree before a
-candidate is refuted, so a deletion needs two models with uncorrelated blind spots and no lone
-skeptic drops a real finding. Point it at a cross-vendor model. With none set, both models
-collapse to one, the finder is single-model and the verify stage keeps every candidate, the
-recall-safe default.
+Both review paths name three model roles, finder, challenger, and judge. The finder finds, the
+challenger refutes, the judge confirms before a deletion. Each role defaults to the base
+`--model`, so a single-model run sets only `--model`. Put a different vendor in any seat for
+cross-model review, where uncorrelated blind spots catch what one model misses, and a deletion
+needs the judge to be a distinct model from the challenger so no lone skeptic drops a real
+finding. With the judge not distinct, nothing is refuted, the recall-safe default.
+
+Each role takes a full backend, the same five fields as the base, every one unset by default:
+`CODEJURY_<ROLE>_PROVIDER`, `_MODEL`, `_API_KEY`, `_API_BASE`, `_WIRE_API`, with `<ROLE>` one of
+`FINDER`, `CHALLENGER`, `JUDGE`, and the matching `--<role>-provider`, `--<role>-model`,
+`--<role>-api-key`, `--<role>-api-base`, `--<role>-wire-api` flags. An unset field inherits the
+base, so you set none of these for a single-model run and override only the seat you want to
+change. The key and the api-base inherit the base only when the role keeps the base provider. A
+role that switches vendor brings its own key, since the base key belongs to the base vendor. For
+example a Claude base finder challenged by GPT and confirmed by Claude:
 
 ```bash
-export CODEJURY_CHECKER_PROVIDER=openai   # default
-export CODEJURY_CHECKER_MODEL=...         # a different model from CODEJURY_MODEL
-export CODEJURY_CHECKER_API_KEY=...
-export CODEJURY_CHECKER_API_BASE=...      # optional
+export CODEJURY_CHALLENGER_PROVIDER=openai
+export CODEJURY_CHALLENGER_MODEL=...         # a GPT model, the skeptic
+export CODEJURY_CHALLENGER_API_KEY=...
+export CODEJURY_CHALLENGER_WIRE_API=responses   # the gpt-5 reasoning models speak Responses
+export CODEJURY_JUDGE_MODEL=...              # a Claude model, the confirmer, distinct from the challenger
 ```
 
+The same `CODEJURY_FINDER_*` / `CODEJURY_CHALLENGER_*` / `CODEJURY_JUDGE_*` and the matching
+`--finder-* / --challenger-* / --judge-*` flags work on both `review diff` and `review repo`.
+Note that `review repo --run` finds with one model, the finder, it no longer adds a second
+co-finder, and a seat that runs on the subscription supplies its own review, so it ignores
+that seat's backend flags while the others still apply.
+
 The tool does not auto-load `.env`.
 
 Useful flags:
@@ -79,21 +94,28 @@ Useful flags:
 - `--api-key <key>`
 - `--api-base <url>`
 - `--retries <n>`
+- `--timeout <seconds>`
 
 ## Data Boundary
 
 The tool sends code-derived content to the model provider you configure, so know what
 leaves the machine before reviewing a proprietary repository:
 
-- Diff Review sends the unified diff under review.
-- Repo Review with `--reviewer model` sends bounded source snippets, the detected stack
+- Diff Review on the `api` row sends the unified diff under review. On the `subscription`
+  row it sends the diff in the `claude -p` prompt through your Claude Code account, and the
+  diff agent uses no file tools, so only the diff text leaves the machine, not local files.
+- Under the default `--executor auto`, each seat follows the `api` row when it has a key
+  and the `subscription` row when it falls back to your Claude Code subscription, so what
+  leaves the machine is decided per seat by whether that seat has a key.
+- Repo Review with `--executor api` sends bounded source snippets, the detected stack
   notes, the vulnerability guidance, and the findings.
-- Verification with `--reviewer model` sends the cited source file and the finding
-  details. With `--reviewer claude-cli`, Claude Code receives the finding details and
-  reads the code itself through its read-only tools.
-- `--reviewer claude-cli` does not use the configured provider key. It runs Claude Code
-  with read-only file tools, and Claude Code may send prompts and the code it reads
-  through your Claude Code account, so the code does not stay local.
+- Verification with `--executor api` sends the cited source file and the finding
+  details. On the `subscription` row, Claude Code receives the finding details and reads
+  the code itself through its read-only tools.
+- A repo seat on the `subscription` row does not use the configured provider key. It runs
+  Claude Code with read-only file tools, and Claude Code may send prompts and the code it
+  reads through your Claude Code account, so the code does not stay local. The diff agent is
+  narrower, it reads no files and sees only the diff in the prompt.
 
 A custom `--api-base` or a LiteLLM proxy becomes part of the trust boundary, so the data
 above also reaches that gateway. Prefer the `CODEJURY_API_KEY` environment variable over
@@ -121,8 +143,20 @@ codejury review diff --file changes.diff --mode adversarial
 
 # emit SARIF and fail on HIGH or CRITICAL findings
 codejury review diff --file changes.diff --format sarif --fail-on high
+
+# review with no provider key, riding your Claude Code subscription
+codejury review diff --file changes.diff --executor subscription
+
+# adversarial with a keyless Claude finder and judge plus an OpenAI challenger on its own key
+codejury review diff --file changes.diff --mode adversarial \
+  --challenger-provider openai --challenger-api-key "$OPENAI_API_KEY"
 ```
 
+Diff Review takes the same `--executor auto|api|subscription` as Repo Review, see Review
+Strategy. The default `auto` calls the provider when a seat has a key and falls back to your
+Claude Code subscription for a keyless Anthropic seat. Unlike the repo agent, the diff agent
+answers from the diff in the prompt and reads no files.
+
 `codejury review diff --dry-run` uses a mock provider and a built-in demo diff, so it needs
 no API key.
 
@@ -185,20 +219,26 @@ codejury review repo /path/to/repo --run
 
 A `--run` chooses how each unit is reviewed:
 
-- `--reviewer model` is the default. It makes one grounded model call per unit. Add
-  `--facts` to ground that call in a tool-extracted call graph, storage layout, and read
-  and write sets when the domain binds a facts backend, such as the EVM Slither backend.
-  This is what gives a smart contract review its call relationships, so prefer
-  `--run --facts` on Solidity.
-- `--reviewer claude-cli` runs each unit and its verification as a headless `claude -p`
-  agent that reads and traces the files itself with read-only tools, using your Claude Code
-  access and no provider key. Use it when you want a tool-using agent rather than a single
-  grounded call.
-
-Add `--checker-model`, a different model from `--model`, to enable cross-model
-verification. It finds alongside the main model on the recall side and must agree before a
-finding is dropped on the precision side. With none set, the verify stage refutes nothing,
-the recall-safe default.
+- `--executor auto` is the default. Each seat, the finder and the skeptic, calls the
+  provider when it has a reachable key and falls back to a headless `claude -p`
+  subscription agent for a keyless Anthropic seat, so a keyless run works with no provider
+  key. A keyless non-Anthropic seat, such as an OpenAI finder with no key, is a loud error,
+  it has no subscription to fall back to. This is what lets a Claude finder ride your
+  subscription while an OpenAI challenger uses its own key.
+- `--executor api` makes one grounded model call per unit and requires a key, a missing key
+  is a loud startup error, the same point as auto. Add `--facts` to ground that call in a
+  tool-extracted call graph, storage layout, and read and write sets when the domain binds a
+  facts backend, such as the EVM Slither backend. This is what gives a smart contract review
+  its call relationships, so prefer `--run --facts` on Solidity.
+- `--executor subscription` always runs each unit and its verification as a headless
+  `claude -p` agent that reads and traces the files itself with read-only tools, using your
+  Claude Code access and no provider key. Use it when you want a tool-using agent rather
+  than a single grounded call even where a key is present.
+
+Set a distinct `--judge-model`, the confirmer, from the challenger to enable cross-model
+verification. The challenger refutes a finding and the judge must agree before it is dropped,
+so a deletion needs two models. With the judge not distinct from the challenger, the verify
+stage refutes nothing, the recall-safe default.
 
 ## Supported Knowledge