pi-advisor-flow 0.8.3 → 0.9.0

package/CHANGELOG.md

@@ -4,6 +4,13 @@ All notable changes to this project are documented here.
 
 The format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0/).
 
+## 0.9.0 - 2026-09-28
+
+### Added
+
+- **Visual Advisor reviews** — image-capable Advisor models can now review real pixels, not just paths or descriptions. PNG, JPEG, GIF, and WebP images from selected conversation history and full-policy tool results are attached when the Advisor model accepts image input, bounded to 4 images, 8 MiB total, and 4 MiB each. Explicit `includeTrackedFiles` and `includeUntracked` handoffs accept image files under their existing consent rules. Images are format-validated before egress; unsupported formats, mismatched MIME types, oversized files, text-only models, and tool policies that withhold results are reported to the Advisor as pixels-not-reviewed so it cannot claim visual evidence it never saw. Scout and Jev never receive image bytes. Results show an `N images attached · size · withheld` label with human-readable sizes.
+- **Same-model Advisor suppression** — when the active Executor model matches the configured Advisor, `ask_advisor`, `/advisor-manual`, and automatic gates are skipped before any screening, budget, or tracked-file consent is spent, with a rendered skip result instead of an error. Switching either model (including via `/model`) re-enables consultations and shows a notice. Enabled by default; turn off **Disable same-model Advisor** in `/advisor-settings` (or `"advisorDisableSameModel": false`) for intentional higher-effort same-model reviews.
+
 ## 0.8.3 - 2026-09-27
 
 ### Changed

package/README.md

@@ -19,10 +19,11 @@ Keep implementation on a fast model and borrow frontier reasoning only when deci
 - **On-demand second opinions** through the `ask_advisor` tool or `/advisor-manual`.
 - **Configurable review gates** before plans, after repeated failures, and before declaring completion.
 - **Automatic loop detection** for repeated tool calls, with explicit proceed, revise, or blocked decisions.
-- **Separate model and reasoning controls** for the Executor and Advisor.
+- **Separate model and reasoning controls** for the Executor and Advisor, with same-model consultations skipped by default to avoid redundant calls.
 - **Model whitelist** that can restrict Advisor calls to exact `provider/model` Executor references.
 - **Advisor usage accounting** with per-response token and cost details, normalized usage in Pi's `/cost` totals, and an optional cumulative footer.
-- **Privacy controls** for conversation history, repository context, explicit file handoff, tool results, secret redaction, and outcome logging.
+- **Privacy controls** for conversation history, repository context, explicit file and image handoff, tool results, secret redaction, and outcome logging.
+- **Visual Advisor reviews** for supported PNG, JPEG, GIF, and WebP images in selected conversation or tool results when the Advisor model accepts images.
 - **Optional persistent activation, Simple mode, session summaries, and Herdr integration.**
 - **Compact searchable `/advisor-settings`** that matches Pi's settings list and saves changes immediately.
 - **Experimental Advisor Scout** that uses the configured Executor model to curate conversation evidence before every Advisor call.
@@ -70,7 +71,7 @@ Reload Pi after installing.
 /advisor-settings   # Configure behavior, modes, etc.
 ```
 
-On first use, or whenever a saved model is unavailable, `/advisor` opens the same available-model picker as `/advisor-models`; it never silently chooses an unconfigured model. You can also enable the flow and select both models at once:
+On first use, or whenever a saved model is unavailable, `/advisor` opens the same available-model picker as `/advisor-models`; it never silently chooses an unconfigured model. If the active Executor and Advisor use the same provider/model, calls and automatic gates are skipped with a notice; switching either model resumes consultations. Turn off **Disable same-model Advisor** in `/advisor-settings` (or set `"advisorDisableSameModel": false` globally) if you intentionally want a higher-effort review from that same model. You can also enable the flow and select both models at once:
 
 ```text
 /advisor executor=openai-codex/gpt-5.6-luna advisor=openai-codex/gpt-5.6-sol
@@ -108,7 +109,7 @@ Advisor Scout is off by default. When enabled in `/advisor-settings` or via `"ad
 
 ## Privacy
 
-Advisor requests can include user messages, tool calls, tool results, targeted questions, and repository information. Repository context is configurable from no access through changed-file summaries to a capped patch; when it is disabled, the Advisor is told so rather than shown an apparently clean tree. Explicit tracked and untracked file contents require separate global opt-ins and are sent as untrusted data. Secret redaction is off by default; when enabled, credential-shaped values in targeted questions are redacted before the provider request. Tools without an explicit policy use full context. Settings are global, so a project cannot silently change them.
+Advisor requests can include user messages, tool calls, tool results, targeted questions, and repository information. Repository context is configurable from no access through changed-file summaries to a capped patch; when it is disabled, the Advisor is told so rather than shown an apparently clean tree. Images from disclosed conversation and full-policy tool results can be sent as pixels only to image-capable Advisor models; Scout sees markers, not pixels. Exact tracked and untracked image files can be attached using `includeTrackedFiles` and `includeUntracked` under their existing separate global consent rules. Images are limited to four and 8 MiB total, with a 4 MiB per-image cap; unsupported, missing, or oversized images are reported as withheld, not reviewed. Explicit tracked and untracked file contents require separate global opt-ins and are sent as untrusted data. Secret redaction is off by default; when enabled, credential-shaped values in targeted questions are redacted before the provider request. Tools without an explicit policy use full context. Settings are global, so a project cannot silently change them.
 
 When Scout is enabled, the Executor model provider also receives bounded Advisor-eligible conversation history. Read [Privacy and data handling](https://github.com/philipbrembeck/pi-advisor/blob/main/docs/privacy.md) before using pi-advisor with sensitive work.