pr-agent

⚠️ Upgrade notes

• GitHub polling needs an allowlist. New host-only github.polling_repositories = ["owner/repo", ...]; empty (the default) ignores every notification, so polling deployments must set it (#4294).
• Jira project_keys = [] now disables lookups. The default is now ["*"]. If you set [] to mean "all projects", change it to ["*"] or remove it (#4049).
• litellm.extra_headers is host-only. Repository .pr_agent.toml and comment arguments can no longer set it (#4287).
• Default model is now gpt-6.1-sol with gpt-6-luna as fallback (was gpt-5.6 / gpt-5.6-terra). Pin config.model to keep the old one (#4088).
• Langfuse SDK 4.18.0 needs self-hosted Langfuse 3.63.0+; use the langfuse_otel callback (#4343).
• GitHub App token in Actions: set github.agent_login to <app-slug>[bot] so PR-Agent recognises its own comments (#4329).

✨ Highlights

• pr-agent doctor preflights a deployment (#4321).
• --json-output for /describe, /improve and /review (#4265, #4227, #4226).
• Jira scoped API tokens (#4250), /ask with the CI artifact as context (#4176), globs in repo_context_files (#4220), opt-in GitLab subgroup settings inheritance (#4059).
• Security hardening: DNS rebinding block, repo-context path escapes, host-only config output, loguru diagnose off, more credential redaction (#4252, #4279, #4293, #4280, #4328).

What's Changed

• Redact Stripe, Twilio and SendGrid credentials by @IlyaKhalafi in #4328
• ci(review): make self-hosted /improve post suggestions again by @Yoel-qin in #4174
• Enable automatic /review on new PRs (fixes #4162) by @kiranmutyala18 in #4170

🚀 Features

• feat(artifacts): support multiple artifact files by @utsab345 in #4338
• feat(cli): add a doctor command to preflight the deployment by @Yoel-qin in #4321
• feat(describe): write the description to --json-output by @Yoel-qin in #4265
• feat(jira): support scoped API tokens by @utsab345 in #4250
• feat(repo-context): glob entries and fallback alternatives in repo_context_files by @ira-at-work in #4220
• feat(ask): include organizational skills in line-level answers by @Yoel-qin in #4225
• feat(review): list the files left out in the JSON output by @Sajinthan in #4226
• feat(improve): write suggestions to --json-output by @Sajinthan in #4227
• feat(ask): answer questions with the CI artifact as context by @Yoel-qin in #4176
• feat(help): prioritize a matching document in constrained prompts by @oleksii-tumanov in #4101
• feat(gitlab): add opt-in closest-subgroup settings inheritance by @ashearin in #4059
• feat(improve): retain equal summaries across files by @oleksii-tumanov in #4128

🐛 Bug Fixes

• fix(ask_line): defer handler setup until a diff hunk matches by @oleksii-tumanov in #4354
• fix(mosaico): retain tasks with bounded owner history by @utsab345 in #4337
• fix(add-docs): retry malformed model responses with fallback models by @PeterDaveHello in #4346
• fix(routing): count diff hunks without splitting patch lines by @PeterDaveHello in #4344
• fix: retry whitespace-only model completions by @PeterDaveHello in #4348
• fix(azure): reject incomplete linked work item responses by @oleksii-tumanov in #4173
• fix(algo): prefer exact source lines before fuzzy matching by @PeterDaveHello in #4340
• fix(similar_issue): use the supported Qdrant query API by @PeterDaveHello in #4341
• fix(azure): handle created events without a web link by @oleksii-tumanov in #4171
• fix(ask): answer with a note when every changed file is filtered out by @sarankumar1325 in #4336
• fix(gitlab): paginate repository directory listings by @utsab345 in #4332
• fix(github): let the operator name the agent login an app token posts as by @sarankumar1325 in #4329
• fix: reject boolean routing limits by @utsab345 in #4302
• fix: parse GitLab boolean settings by @utsab345 in #4303
• fix: stop Gitea changed-files pagination at the advertised total by @fixfon in #4306
• fix(describe): retain valid YAML

⚠️ Breaking Changes

[!WARNING] This release changes several defaults and makes more settings host-only. Check each item below against your deployment before upgrading.

• pip install pr-agent no longer installs the git provider SDKs (#3679): PyGithub, python-gitlab, atlassian-python-api, the Azure DevOps SDKs, boto3, giteapy, the Google Cloud SDKs and a2a-sdk moved to the github, gitlab, bitbucket, azure, codecommit, gitea, google and mosaico extras. Install the ones you use, for example pip install "pr-agent[github]", or pr-agent[all] for the previous set. The Docker images already install all.
• Docker service images run as a non-root user (#3910): the webhook, app, polling and Mosaico targets run as UID/GID 10001 with HOME=/home/pragent. Mounted secrets must be readable by that user, output and database directories must be writable, and home-relative mounts should move to /home/pragent. The GitHub Action, CLI and Lambda images still run as root.
• Azure DevOps and Bitbucket Server webhooks require credentials (#3911): with azure_devops_server.webhook_username/webhook_password or bitbucket_server.webhook_secret unset, every webhook is rejected with HTTP 403. Set them on both the sender and PR-Agent before upgrading.
• GitLab webhooks backed by a secret provider need a new token format (#3913): when GitLab credentials come from AWS Secrets Manager or Google Cloud Storage, the webhook's Secret Token must be <secret-name>:<webhook-token>. Name-only tokens are rejected with HTTP 401, so update the webhooks when you deploy.
• Namespace-wide settings are off by default (#3908): PR-Agent no longer applies <namespace>/pr-agent-settings to every repository automatically. Set the host-only config.global_settings_repo = "pr-agent-settings" to keep the old behaviour.
• [tool.pr-agent] in pyproject.toml is no longer read (#3941): move those settings to .pr_agent.toml.
• More settings are host-only (#3840, #3904, #3950, #4001, #4054): a repository's .pr_agent.toml and comment arguments can no longer set provider endpoints (openai.api_base and the other api_base keys), keys ending in url, endpoint, org, key, token or secret, provider auth mode and TLS settings (github.deployment_type, gitlab.auth_type/ssl_verify, bitbucket.auth_type, gitea.skip_ssl_verification/ssl_ca_cert), jira.jira_site/jira_api_email/project_keys, or LiteLLM callbacks. They are ignored with a warning, so set them in host configuration instead. Comment arguments are restricted further as well (#3902).
• Gitea and Gerrit read repository settings from the target branch (#3906): Gitea uses the PR base and Gerrit the default branch, so a .pr_agent.toml change inside a PR no longer applies to that PR.
• Cross-repository reads need an allowlist (#3922, #3897): GitHub ticket context only reads linked issues and sub-issues from the PR's own repository or same-owner repositories listed in config.repo_context_sibling_repos. gitlab.expand_submodule_diffs only expands submodules listed there that sit in the MR's top-level group.
• **Ignore rules now cover

⚠️ Important Changes

• Gitea webhooks now require a secret (#3637): with gitea.webhook_secret unset, the Gitea app rejects every webhook with HTTP 403 instead of accepting it unsigned. Set the secret on the Gitea webhook and in your config before upgrading.
• The GitHub Action fails on tool errors (#3712): a tool failure that used to be logged and swallowed now fails the Action. Set github_action_config.fail_on_tool_errors = false to keep the old green behaviour.
• Comment commands need a leading / on GitLab, Bitbucket and Bitbucket Server (#3769): a comment only runs a tool when it starts with a slash command, so plain comments that mention one no longer trigger anything.
• extra_config_url is host-only (#3638): a repository's .pr_agent.toml can no longer set it. Configure it on the host instead, for example with PR_AGENT_EXTRA_CONFIG_URL.
• commitable_code_suggestions is now committable_code_suggestions (#3749): the old spelling still works with a deprecation warning until 1.0.
• The Bitbucket Server webhook image runs under gunicorn (#3667): the container now starts multiple workers through the shared gunicorn_config.py instead of a single Python process.

Known issue: Bedrock Kimi K3, new in this release (#3651), still receives a temperature by default, which Bedrock is reported to reject. Until #3752 lands, add the model ids to config.no_temperature_models.

What's Changed

• Run BitBucket Server webhook under gunicorn for worker isolation by @jkratz55 in #3667
• test: pin LiteLLM future Claude fallbacks by @PeterDaveHello in #3660
• test(ai_handlers): add tripwire for guarded private LiteLLM imports by @utsab345 in #3615
• test: use the bundled LiteLLM cost map in local unit runs by @akalongman in #3641
• fix(codecommit): align language classification with diff packing by @oleksii-tumanov in #3625
• refactor(cloud_auth): deduplicate repeated logic by @high-tail in #3607

🚀 Features

• feat(gitlab): distinguish applied and bot-resolved suggestion threads in the improve context by @JokeWaumans in #3774
• feat: add Bedrock Kimi K3 models by @ira-at-work in #3651
• feat: reply to GitLab trigger discussions by @deadpool-mjpatel4444 in #3426
• feat(gitlab): support inline key issues by @utsab345 in #3650
• feat(ai_handlers): surface prompt cache visibility by @utsab345 in #3618
• feat(models): add Claude Opus 5.5 support by @PeterDaveHello in #3629
• feat: validate /describe output against existing model by @taljeon in #3616

🐛 Bug Fixes

• fix(git_providers): log publish and settings errors at quiet verbosity by @PeterDaveHello in #3802
• fix(tickets): cap fetched sub-issues per ticket and order parent before children by @rishiiicreates in #3807
• fix(models): clamp OpenRouter Gemini reasoning disablement by @oleksii-tumanov in #3787
• fix(tools): re-raise any provider incomplete-files error, not just Bitbucket's by @PranavD2905 in #3800
• fix(bitbucket): encode filenames in line links by @kudala-bharani in #3796
• fix(servers): route inline comments to /ask_line only on a leading /ask by @utsab345 in #3797
• fix(review): link a TODO without a usable line number to its file by @Arta-shoja in #3788
• fix: stop caching repo context on a mutable ref by @utsab345 in #3776
• fix(typo): commitable to committable by @even-even in #3749
• fix(action): share GitHub helpers so the Action keeps the #3670 /ask fix by @Yoel-qin in #3785
• fix(github): fetch newest polling comments by @oleksii-tumanov in #3747
• fix(labels): build the label enum whe

What's Changed

🚀 Features

• feat(status): acknowledge a pull request the moment automatic commands start by @dwin-gharibi in #3091
• feat(gitlab): content-based auto-resolve for fixed inline threads by @Nie1iX in #3478
• feat(tickets): make description issue matching configurable by @utsab345 in #3385
• feat: register Grok 4.3 reasoning effort by @deadpool-mjpatel4444 in #3425
• feat(context): load repo-context files from sibling repos (same namespace/owner) by @utsab345 in #3362
• feat: add typed prompt output models by @utsab345 in #3371
• feat: per-directory .pr_agent.toml for monorepo settings by @utsab345 in #3282
• feat(improve): recover failed chunks with fallback models by @junnhwan in #3324
• feat(jira): add project-key allowlist to skip lookups for key-shaped noise by @Fayupable in #3304
• feat: add Jira ticket-context support by @dellch in #2420
• feat(telemetry): expose a native prometheus scrape endpoint by @utsab345 in #3238
• feat: generalize org-level settings to all providers via get_owning_namespace by @utsab345 in #3143
• feat(models): route small pull requests to a cheaper primary model by @ylcn91 in #3104
• feat(artifacts): read the CI artifact context from the CLI as well by @ylcn91 in #3105
• feat(models): add Claude Fable 5.1 support by @PeterDaveHello in #3101
• feat(models): add OpenAI GPT-6 Astra support by @PeterDaveHello in #3087
• feat(push_outputs): route describe and improve output to the configured sinks by @dwin-gharibi in #3098
• fix(tickets): bound the shorthand issue-number match at six digits, matching the branch pattern by @dwin-gharibi in #3082
• feat(reactions): configurable start and outcome reactions by @dwin-gharibi in #3089
• feat(review): publish sanitized failure reasons by @elijahchancey in #2996

🐛 Bug Fixes

• fix: check pinecone upsert response by @utsab345 in #3561
• fix: preserve quoting in CLI and configured Action commands by @oleksii-tumanov in #3585
• fix(github): send label names to the set-labels endpoint by @utsab345 in #3582
• fix: point shipped docs links at live targets by @utsab345 in #3556
• fix(diff): preserve source lines containing EOF marker text by @PeterDaveHello in #3572
• fix(codecommit): apply ignore rules and read .pr_agent.toml from the destination commit by @PeterDaveHello in #3576
• fix(polling): log worker capacity waits by @PeterDaveHello in #3574
• fix(similar_issue): treat a blank qdrant url as unset by @utsab345 in #3551
• fix(similar_issue): honor max issues scan limit by @mote24405-coder in #3565
• fix(similar_issue): let a second repository join an existing lancedb table by @utsab345 in #3563
• fix(similar_issue): make lancedb search and ingest match the other backends by @utsab345 in #3562
• fix: namespace pinecone similar issue by @utsab345 in #3560
• fix(similar_issue): stop lancedb duplicating the sentinel row by @utsab345 in #3558
• fix(similar_issue): keep the best hit per issue so the comment lists stay aligned by @utsab345 in #3557
• fix(similar_issue): list all lancedb tables instead of the deprecated 10-name call by @utsab345 in #3553
• fix: preserve generated changelog after write failures by @oleksii-tumanov in #3529
• fix(similar_issue): skip null comment bodies before splitting by @utsab345 in #3549
• fix: resolve the PR head SHA through a provider hook (#3431) by @sloemo01 in #3528
• fix: invert lancedb ingest table check by @utsab345 in #3531
• fix(gitlab): narrow broad exception handlers in the GitLab provider by @wfatih in #3525
• fix(gpt5): clamp reasoning_effort 'max' where xhigh is unsupported by @auroraxo in #3273
• fix: redact sensitive AWS Secrets Manager error logs by @PeterDaveHello in #3512
• fix(tests): validate Gitea E2E file writes and tool results by @PeterDaveHello in #3513
• fix: keep tags in reasoning support probe by @utsab345 in #3502
• fix(litellm): support adaptive thinking for profile ARNs by @rajin-khan in #3497
• fix(github-