Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
15 changes: 15 additions & 0 deletions CHANGELOG.md
Original file line number Diff line number Diff line change
Expand Up @@ -209,6 +209,21 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0

### Changed

- **Bundled Claude Code 2.1.325 and Antigravity CLI 1.5.9**: the
`propr/agent` image moves from Claude Code 2.1.323 and Antigravity CLI
1.5.8. Codex stays on 0.168.0 and Mistral Vibe on 2.25.8.
- Antigravity 1.5.9 is release `4515810895593472`, SHA-512 checked against
Google's linux_amd64 install manifest. Its model list is unchanged. It
gives every user the `run_workflow` tool, which lets the agent fan work
out to subagents from a Python script, so large headless tasks may use
more tokens and run longer. `-p`, `--print-timeout` and exit codes are
unchanged.
- Claude Code 2.1.325 asks for permission before sandboxed Bash commands
and before read-only git commands in other repositories. ProPR runs are
unaffected: tasks and goals pass `--dangerously-skip-permissions`,
analysis runs use `--tools ''` or the repository-scout MCP allow-list,
and ProPR never turns on Claude Code's sandbox. The bundled model catalog
is unchanged.
- **Faster pull request checks**: `Validate Changes` runs its lint, typecheck
and build gate before the browser smoke test and skips the browser run when
the gate fails, so a lint error is reported in minutes instead of after an
Expand Down
12 changes: 6 additions & 6 deletions Dockerfile.agent
Original file line number Diff line number Diff line change
Expand Up @@ -121,7 +121,7 @@ WORKDIR /home/node/workspace


FROM agent-base AS claude-cli
ARG CLAUDE_CLI_VERSION=2.1.323
ARG CLAUDE_CLI_VERSION=2.1.325
USER root
RUN npm install -g "@anthropic-ai/claude-code@${CLAUDE_CLI_VERSION}" \
&& npm cache clean --force \
Expand Down Expand Up @@ -196,9 +196,9 @@ RUN npm install -g "agent-tank@${AGENT_TANK_CLI_VERSION}" \


FROM agent-base AS antigravity-cli
ARG ANTIGRAVITY_CLI_VERSION=1.5.8
ARG ANTIGRAVITY_CLI_RELEASE_ID=4602374149505024
ARG ANTIGRAVITY_CLI_SHA512=f08faaee588327dba8722bdc60f26a14f6915a8f2d756ed7959cde0a706d4e62fd9f831ab1a077dace5f947caaa6e851457df5023180bb8d3f54386804a3787f
ARG ANTIGRAVITY_CLI_VERSION=1.5.9
ARG ANTIGRAVITY_CLI_RELEASE_ID=4515810895593472
ARG ANTIGRAVITY_CLI_SHA512=d01eb13ba7750eed18907f1ae8860864371e0ad9064415e8d033d22b4bd7b9b4e4c7f962058520b452d67db1c33163d9a889d55a7dffa159596ec6cd91a87b79
ENV PATH="/home/node/.local/bin:${PATH}" \
AGY_CLI_DISABLE_AUTO_UPDATE=true
# Both download URLs resolve the same pinned public GCS object. The JSON API
Expand Down Expand Up @@ -309,9 +309,9 @@ WORKDIR /home/node/workspace
FROM --platform=${AGENT_PLATFORM} scratch AS final
ARG AGENT_NODE_VERSION
ARG AGENT_YARN_VERSION
ARG CLAUDE_CLI_VERSION=2.1.323
ARG CLAUDE_CLI_VERSION=2.1.325
ARG CODEX_CLI_VERSION=0.168.0
ARG ANTIGRAVITY_CLI_VERSION=1.5.8
ARG ANTIGRAVITY_CLI_VERSION=1.5.9
ARG OPENCODE_CLI_VERSION=1.19.5
ARG VIBE_CLI_VERSION=2.25.8
ARG AGENT_TANK_CLI_VERSION=0.10.0
Expand Down
8 changes: 4 additions & 4 deletions THIRD_PARTY_LICENSES.md
Original file line number Diff line number Diff line change
Expand Up @@ -31,7 +31,7 @@ https://pkgs.alpinelinux.org/packages

---

## @anthropic-ai/claude-code@2.1.323
## @anthropic-ai/claude-code@2.1.325

```
© Anthropic PBC. All rights reserved. Use is subject to the Legal Agreements outlined here: https://code.claude.com/docs/en/legal-and-compliance.
Expand Down Expand Up @@ -306,9 +306,9 @@ below is bundled in the propr/app image under the stated license.

```
"module name","license","repository"
"@anthropic-ai/claude-code-linux-x64-musl@2.1.323","Custom: https://www.npmjs.com/package/",""
"@anthropic-ai/claude-code-linux-x64@2.1.323","Custom: https://www.npmjs.com/package/",""
"@anthropic-ai/claude-code@2.1.323","Custom: https://img.shields.io/badge/Node.js-22",""
"@anthropic-ai/claude-code-linux-x64-musl@2.1.325","Custom: https://www.npmjs.com/package/",""
"@anthropic-ai/claude-code-linux-x64@2.1.325","Custom: https://www.npmjs.com/package/",""
"@anthropic-ai/claude-code@2.1.325","Custom: https://img.shields.io/badge/Node.js-22",""
"@anthropic-ai/sdk@0.71.2","MIT","https://github.com/anthropics/anthropic-sdk-typescript"
"@babel/runtime@7.28.4","MIT","https://github.com/babel/babel"
"@dnd-kit/accessibility@3.1.1","MIT","https://github.com/clauderic/dnd-kit"
Expand Down
2 changes: 2 additions & 0 deletions docs/docs/features/agents-and-models.md
Original file line number Diff line number Diff line change
Expand Up @@ -177,6 +177,8 @@ Gemini 4 Hydrogen, the fast member of the Gemini 4 family, needs Antigravity CLI

Gemini 4 Carbon is not in the catalog. Antigravity CLI 1.5.8 offers it only to accounts that connect with a Gemini API key (`GEMINI_API_KEY`) or sign in with Application Default Credentials (`AGY_ADC_AUTH`), and the setup above signs in with `agy login`.

From Antigravity CLI 1.5.9, the agent has a `run_workflow` tool that runs a Python script to hand parts of a task to several subagents and combine their results. On large tasks this can use more tokens and take longer than a single agent working alone.

:::note[Gemini Enterprise review policy]
ProPR runs Antigravity tasks with `--dangerously-skip-permissions` and repository inspection with `--sandbox`, so the Auto Agent Mode that Antigravity CLI 1.5.7 made the default for headless runs does not change how ProPR runs approve actions. From Antigravity CLI 1.5.2, when a Gemini Enterprise admin requires review of agent actions, that flag no longer approves actions outside the workspace in headless runs: the CLI denies them. Work inside the task's worktree is unaffected, but a step that needs a path outside it fails instead of running. Ask the admin to relax the review requirement for the account ProPR uses if tasks need such actions.
:::
Expand Down
2 changes: 1 addition & 1 deletion docs/docs/features/pr-commands.md
Original file line number Diff line number Diff line change
Expand Up @@ -97,7 +97,7 @@ When the linked issue has an **Acceptance criteria** list, the review prompt lis

| Runtime | Context window used | Source |
|---|---|---|
| Claude Code | 1M for models the runtime marks native-1M (Opus 4.7+, Opus 5.x, Sonnet 5.x, Haiku 5.5, Fable, Mythos); 200K otherwise, including Opus/Sonnet 4.6 without the `[1m]` suffix or with `CLAUDE_CODE_DISABLE_1M_CONTEXT` | Model catalog bundled with Claude Code 2.1.323 |
| Claude Code | 1M for models the runtime marks native-1M (Opus 4.7+, Opus 5.x, Sonnet 5.x, Haiku 5.5, Fable, Mythos); 200K otherwise, including Opus/Sonnet 4.6 without the `[1m]` suffix or with `CLAUDE_CODE_DISABLE_1M_CONTEXT` | Model catalog bundled with Claude Code 2.1.325 |
| Codex | 272K for every listed model except GPT-6.1 Terra (500K), including the rest of the GPT-6 family | `context_window` in the models catalog bundled with Codex CLI 0.168.0; larger windows are opt-in through `model_context_window`, which ProPR does not set |
| Other runtimes | ProPR model catalog window, with a runtime reserve of 10% of the window (minimum 16K) | Catalog only; the runtime limit is not verified |
| Unknown model | 200K (Codex: its historical 272K) | Conservative fallback |
Expand Down
4 changes: 2 additions & 2 deletions docs/operations/managed-diagnostics.md
Original file line number Diff line number Diff line change
Expand Up @@ -47,7 +47,7 @@ Audited from `Dockerfile.agent`, `agents/version/types.ts`, `configManagerAgents
| --- | --- | --- | --- | --- |
| Application / Pino 9 | Application stdout; private spool under mounted data | Complete JSON records from Pino hook | Existing DB-backed task/LLM history stays under its existing policy | Application process; managed capture owns rotation. Fleet owns stdout/journald/Docker rotation. |
| Codex 0.168.0 | `M/<id>/.codex` or explicit `CODEX_CONFIG_PATH`/`HOST_CODEX_DIR` → `/home/node/.codex:rw` | `log/codex-tui.log`; `sessions/YYYY/MM/DD/rollout-*.jsonl`; same under `archived_sessions` | `auth.json`, config.toml, token/auth stores never exported; `state_<n>.sqlite` and `logs_<n>.sqlite` plus WAL/SHM measured and protected, not copied as files | Exec/app-server can keep descriptors open for the task/goal. No verified writer-reopen contract is assumed. Snapshot complete append records; rotate the managed copies. Native SQLite diagnostics require a future schema-aware reader and remain protected. |
| Claude 2.1.323 | `M/<id>/.claude` or explicit `CLAUDE_CONFIG_PATH`/`HOST_CLAUDE_DIR` → mounted Claude home | `debug/*.txt`; `projects/<slug>/*.jsonl`; `projects/<slug>/<session>/subagents/agent-*.jsonl` | `.credentials.json`, settings/config, caches excluded; `history.jsonl` measured/protected; transcripts stay available for native goals/resumption | Task/goal process; no signal/reopen assumed. Read-only measurements plus stdout/stderr capture; no rename/truncate of active transcript or debug files. |
| Claude 2.1.325 | `M/<id>/.claude` or explicit `CLAUDE_CONFIG_PATH`/`HOST_CLAUDE_DIR` → mounted Claude home | `debug/*.txt`; `projects/<slug>/*.jsonl`; `projects/<slug>/<session>/subagents/agent-*.jsonl` | `.credentials.json`, settings/config, caches excluded; `history.jsonl` measured/protected; transcripts stay available for native goals/resumption | Task/goal process; no signal/reopen assumed. Read-only measurements plus stdout/stderr capture; no rename/truncate of active transcript or debug files. |
| Antigravity 1.2.4 (release 6085322963025920) | `M/<id>/.gemini`, legacy explicit root → `.gemini-source`; entrypoint may copy selected auth into disposable `/tmp/propr-antigravity-home/.gemini` | In a mounted root: `antigravity-cli/log/*.log`, `antigravity-cli/brain/<conversation>/.system_generated/logs/transcript.jsonl`; managed stdout/stderr includes native-goal segments | `antigravity-cli/antigravity-oauth-token`, settings, arbitrary brain/project/cache contents excluded | Task/goal process. Disposable container-only files are not falsely reported as mounted history. No native rotation/reopen proven; mounted append history is measured but not exported without complete credential context. Existing transient transcript/UI handling remains unchanged. |
| OpenCode 1.18.31 | `M/<id>/.config/opencode`; sibling `M/<id>/.local/share/opencode` or explicit `HOST_OPENCODE_DATA_DIR` supplies credentials. Runtime uses disposable XDG data via `PROPR_EPHEMERAL_STATE=1` | Mounted data root `log/*.log` when present; managed stdout/stderr for every execution | `auth.json`, `account.json`, config excluded; mounted `opencode.db`/WAL/SHM measured/protected, never exported | Runtime native DB/logs can be container-only; no broad copy or SQLite live-file copy. Captured stream is the supported diagnostic history for disposable runs. |
| Vibe 2.25.8, Python 3.12 | Explicit `.vibe` source is read-only. Runtime: `VIBE_PROMPT_CACHE_DIR/propr-vibe-runtime/propr-vibe-home-*` assigned by worker → `/tmp/propr-vibe-home`, also `HOME`/`VIBE_HOME` | Assigned runtime `logs/session/<id>/messages.jsonl`, plus stdout/stderr; same allowlist in an explicit mounted source | config.toml, credentials.json, .env, prompt cache files, skills and secrets excluded | Task-local runtime; snapshot while active and on child close, before existing cleanup. Does not add or promise Vibe native-goal resumption. |
Expand Down Expand Up @@ -179,7 +179,7 @@ No live Hosted instance, installed provider-version behavior, encryption/readbac

Dependency PR #3014 was merged on 2026-10-12 before this implementation. No overlapping native maintenance writer is introduced. All supported native sources in the audit table remain protected: disposable diagnostic text files have an **unsupported writer lifecycle**, JSONL transcripts/history remain resume-dependent, and SQLite/WAL/SHM remain database state. Configuration, credentials and unrecognized entries are excluded from content collection. Completed status, file age, filenames and off-site export do not establish permission to delete any of these originals.

The current image defaults were checked against installed `claude --version` (2.1.323), `codex --version` (0.168.0) and `opencode --version` (1.18.31). This verifies installed versions, not their resume/rotation behavior. An isolated OpenCode 1.18.31 experiment ran `run --pure --model nonexistent/nonexistent test` with separate disposable HOME/XDG directories, model fetching and automatic updates disabled, and no customer credentials. Both default output and `--print-logs --log-level DEBUG` created `data/opencode/log/opencode.log`; stderr output did **not** suppress the original file. The invalid-model runs exited 1. The documented [print-logs option](https://opencode.ai/docs/cli/) is consequently not treated as a disk bound. Antigravity 1.2.4 and Vibe 2.25.8 layouts remain repository fixtures; their installed binaries were not tested here.
The current image defaults were checked against installed `claude --version` (2.1.325), `codex --version` (0.168.0) and `opencode --version` (1.18.31). This verifies installed versions, not their resume/rotation behavior. An isolated OpenCode 1.18.31 experiment ran `run --pure --model nonexistent/nonexistent test` with separate disposable HOME/XDG directories, model fetching and automatic updates disabled, and no customer credentials. Both default output and `--print-logs --log-level DEBUG` created `data/opencode/log/opencode.log`; stderr output did **not** suppress the original file. The invalid-model runs exited 1. The documented [print-logs option](https://opencode.ai/docs/cli/) is consequently not treated as a disk bound. Antigravity 1.2.4 and Vibe 2.25.8 layouts remain repository fixtures; their installed binaries were not tested here.

No provider-supported UTC daily plus size reopen/rollover has been verified for these persistent writers. No original is copytruncated, renamed, unlinked, archived or pruned, including after a task completes. Native daily/size rotation is **unsupported**, rather than simulated by rolling the copy spool. Container-only OpenCode/Antigravity state and disposable Vibe homes follow existing task cleanup; this change does not expand that cleanup or promise native-goal resumption. A real archive-retrieval/resume implementation and verified writer lifecycle are prerequisites for future native disposal, even with Fleet readback acknowledgement.

Expand Down
4 changes: 2 additions & 2 deletions docs/operations/native-provider-log-policy.md
Original file line number Diff line number Diff line change
Expand Up @@ -9,7 +9,7 @@ Image pins in `Dockerfile.agent` are the compatibility baseline. Producer source
| Provider/version | Original sources and writer lifetime | Treatment |
| --- | --- | --- |
| Codex 0.168.0 | `$CODEX_HOME/log/codex-tui.log` when enabled; task/goal tracing also writes `logs_<n>.sqlite` and may use WAL/SHM. `sessions/.../rollout-*.jsonl`, `archived_sessions`, and `state_<n>.sqlite` support session/thread state. App-server stderr has a separate subscriber. | Text diagnostic bytes separate; all SQLite/WAL/SHM protected as database, rollouts as session history. `RUST_LOG` controls stderr but not the independent SQL log layer. No file/DB cap claimed. |
| Claude Code 2.1.323 | `debug/<session>.txt` is diagnostic; project/session/subagent JSONL and `history.jsonl` are session history. Since 2.1.300 transcripts also record each API request's system prompt, tools and parameters (`api-request*` records), so they grow faster; paths are unchanged. Goal writers and UI readers outlive individual messages. | Diagnostic/session bytes separate. Existing task-mode `--no-session-persistence` is not applied to resumable goals. No tested safe debug reopen or daily/size cap; retained. |
| Claude Code 2.1.325 | `debug/<session>.txt` is diagnostic; project/session/subagent JSONL and `history.jsonl` are session history. Since 2.1.300 transcripts also record each API request's system prompt, tools and parameters (`api-request*` records), so they grow faster; paths are unchanged. Goal writers and UI readers outlive individual messages. | Diagnostic/session bytes separate. Existing task-mode `--no-session-persistence` is not applied to resumable goals. No tested safe debug reopen or daily/size cap; retained. |
| Antigravity 1.2.4 (release 6085322963025920) | Mounted `.gemini/antigravity-cli/log/*.log` diagnostics; `brain/<conversation>/.system_generated/logs/transcript.jsonl` session history. Entrypoint uses disposable container runtime too. Goals span invocations. | Mounted originals protected and counted by class. Container-only originals unmeasured. No rotation contract verified. |
| OpenCode 1.18.31 | Explicit source `.local/share/opencode`, runtime XDG data in `/tmp/propr-opencode-data/opencode`. Current writer appends `log/opencode.log`. DB/WAL/SHM hold session state. | **`--print-logs` still writes the file**, adding stderr. Older upstream `Log.init` fixtures that returned early on `print` do not match this binary. All originals protected. |
| Vibe 2.25.8, Python 3.12 | Assigned worker runtime under `VIBE_PROMPT_CACHE_DIR/propr-vibe-runtime/...`, mounted `/tmp/propr-vibe-home`; `logs/session/<id>/messages.jsonl` is conversation history. Source home is read-only. | Session bytes, not disposable debug bytes. Existing task runtime cleanup unchanged. No new native-goal resumption guarantee. Unregistered runtime roots unmeasured. |
Expand All @@ -18,7 +18,7 @@ Image pins in `Dockerfile.agent` are the compatibility baseline. Producer source

Credentials/configuration are metadata-only categories: this inventory reads no contents. Known auth/account/credential JSON, config/settings, `.env`, and the Antigravity OAuth store remain unchanged. Unknown files are counted without content reads; unknown directories are not traversed and make inventory incomplete. Existing producer redaction/private export restrictions still apply.

Codex evidence: [version-tagged app-server subscribers](https://github.com/openai/codex/blob/rust-v0.168.0/codex-rs/app-server/src/lib.rs), [independent SQL sink](https://github.com/openai/codex/blob/rust-v0.168.0/codex-rs/state/src/log_db.rs), [configuration schema](https://github.com/openai/codex/blob/rust-v0.168.0/codex-rs/core/config.schema.json). Installed version checks: Codex 0.168.0, Claude 2.1.323, OpenCode 1.18.31, Vibe 2.25.8. Antigravity image pin/layout audited; binary not exercised here.
Codex evidence: [version-tagged app-server subscribers](https://github.com/openai/codex/blob/rust-v0.168.0/codex-rs/app-server/src/lib.rs), [independent SQL sink](https://github.com/openai/codex/blob/rust-v0.168.0/codex-rs/state/src/log_db.rs), [configuration schema](https://github.com/openai/codex/blob/rust-v0.168.0/codex-rs/core/config.schema.json). Installed version checks: Codex 0.168.0, Claude 2.1.325, OpenCode 1.18.31, Vibe 2.25.8. Antigravity image pin/layout audited; binary not exercised here.

An isolated OpenCode 1.18.31 `serve --print-logs` check found an original 431-byte `opencode.log` with **two open writer descriptors**, alongside stderr output. The opt-in `nativeLogPolicyLive.test.ts` also verifies these descriptors and the original inode survive metadata sweeps while the provider continues serving authenticated localhost health/UI requests. Run with `PROPR_NATIVE_LOG_LIVE=1`; optionally set `PROPR_NATIVE_OPENCODE_BINARY` to the installed executable. Linux x64 executable SHA-256: `f9dab32248695e9ebd56b16a1921798fd85112cf5a69c7dfd0cabc1e17be4a11`. These checks use private temporary homes, no customer credentials and no model request. They prove unsupported active-file behavior, not resumability or production acceptance.

Expand Down
Loading
Loading