Repository navigation
feat(ai): write the AI analysis in the session language, with a summary - #2118
Open
irvingouj@Devolutions (irvingoujAtDevolution) wants to merge 5 commits into
Open
irvingouj@Devolutions (irvingoujAtDevolution) wants to merge 5 commits into
irvingouj@Devolutions (irvingoujAtDevolution) wants to merge 5 commits into
Conversation
irvingouj@Devolutions (irvingoujAtDevolution)
added this pull request to stack #2120
October 19, 2026 18:07
irvingouj@Devolutions (irvingoujAtDevolution)
force-pushed
the
feat/ai-session-language-summary
branch
from
October 19, 2026 18:43
eef6b7b to
377239a
Compare
Copilot started reviewing on behalf of
irvingouj@Devolutions (irvingoujAtDevolution)
October 19, 2026 18:43
Contributor
There was a problem hiding this comment.
🟡 Changes recommended
Truncated summaries lose token usage, and malformed language tags pass validation.
2 open issues
What changed in this PR
Adds language-aware action descriptions and session summaries to the AI crate, supporting Gateway’s recording-log integration.
Changes:
- Passes session language into prompts and checkpoint compatibility checks.
- Generates summaries with retry and fallback handling.
- Marks every AI log entry with
source: "ai"and keeps model metadata in analysis results.
| File | Description |
|---|---|
crates/devolutions-gateway-ai/tests/providers.rs |
Tests language prompts and summary responses. |
crates/devolutions-gateway-ai/src/session_summary.rs |
Adds session-summary requests and serialization. |
crates/devolutions-gateway-ai/src/session_actions.rs |
Adds language-aware transcript descriptions. |
crates/devolutions-gateway-ai/src/screen_actions.rs |
Adds language-aware screenshot descriptions. |
crates/devolutions-gateway-ai/src/recording_analysis/tests.rs |
Tests summaries, fallback, and language-sensitive resume. |
crates/devolutions-gateway-ai/src/recording_analysis/slog.rs |
Writes summaries and AI source markers. |
crates/devolutions-gateway-ai/src/recording_analysis/mod.rs |
Integrates language, summaries, and usage accounting. |
crates/devolutions-gateway-ai/src/lib.rs |
Exposes and documents the summary module. |
crates/devolutions-gateway-ai/src/client.rs |
Adds shared language validation and prompt rules. |
Review effort: Balanced
Comment on lines
+242
to
+245
| tag.len() <= 64 | ||
| && tag | ||
| .split('-') | ||
| .all(|subtag| (1..=8).contains(&subtag.len()) && subtag.bytes().all(|byte| byte.is_ascii_alphanumeric())) |
Comment on lines
+590
to
+592
| Err(error) => { | ||
| warn!(%error, "Writing the session log without a summary"); | ||
| None |
irvingouj@Devolutions (irvingoujAtDevolution)
force-pushed
the
feat/ai-session-language-summary
branch
4 times, most recently
from
October 19, 2026 21:01
5873f2b to
91d1bea
Compare
The recorder's events log carries the locale of the session on its session.start. The AI log, a separate file next to it, must be written in that language, carry a summary, and mark its own lines. - AnalyzeRecording, DescribeSessionActions and DescribeScreenActions take an optional BCP 47 language; the prompts ask for descriptions in it. A value that is not a tag is ignored, so recorder data never adds instructions. Working files of another language are not reused. - New purpose: AiClient::summarize_session. The analysis calls it once when there are actions and writes the text as `summary` on session.end. - Every AI log line carries "source":"ai". - Prompt versions: session-actions-4, screen-actions-2, session-summary-1. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A summary request that fails for good (refused, cut short, empty) no longer throws away the actions already found: the log is written without `summary`, with a warning. A transient failure still fails the attempt, since a retry resumes from the described chunks. The language is checked once when the analysis is set up, so an invalid one is reported once, not on every request. The model and the prompt versions leave the log: they are part of the analysis result. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
- The reported prompt versions leave out the summary prompt when no summary was asked for or it could not be written. - `Writing` is reported once the summary is done, not before asking for it. - Docs no longer say the log names the model and prompt versions. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
irvingouj@Devolutions (irvingoujAtDevolution)
force-pushed
the
feat/ai-session-language-summary
branch
from
October 19, 2026 21:15
91d1bea to
ea28f41
Compare
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The AI analysis now writes in the session language and ends with a summary. This touches the AI crate only.
AnalyzeRecording::language("fr-CA"): action descriptions and the summary are written in that BCP 47 language. Without it, output stays in English as before. A value that isn't a valid tag is ignored, with a warning, because it comes from recorder data and goes into the prompt.session_summary(AiClient::summarize_session): one extra request after the actions, only when there are actions. Its tokens count in the usage. The summary goes insummaryon the AI log'ssession.end.summary, with a warning. A transient failure is retried, and the retry resumes from the described chunks.source: "ai", so a caller showing it next to the recorder's events log can tell the two apart.session-actions-4,screen-actions-2,session-summary-1. The analysis reportssession-actions-4+session-summary-1.Where to look hardest: the summary step at the end of
recording_analysis/mod.rs, and the resume check onWorkspaceSettings.languagein the same file.Known limit: the summary sends all actions in one request, so a very long session could go over the model's context window.
Tests:
cargo test -p devolutions-gateway-ai(the mock provider tests cover the language in the prompts, the summary call, no call when there are no actions, and a failed summary being transient).🤖 Generated with Claude Code