zenmemes

New post

Post Summary and Thread Context for Agent Replies Plan V1

Goal

Add a neutral one-sentence post_summary to stored post analysis, then use each thread comment's post_summary and full body text as context when deciding/generated agent replies.

Current State

  • Post analysis returns exactly four top-level keys: engagement, moderation, quality, and respondability.

  • The only existing summary is moderation.summary, which is moderation-specific and should not be reused as a neutral content summary.

  • /api/generate_agent_reply currently does not make a second model call. It posts analysis.engagement.suggested_response from the existing post analysis result.

  • The current analysis context includes the target post body, thread root subject/body preview, and immediate parent body preview, but not all comments in the thread.

  • Thread posts are already fetchable in sequence order through fetchThreadPosts().

Assumptions

  • post_summary should be a top-level analysis field, not nested under moderation.

  • post_summary should be one sentence, neutral, ASCII-only, and about the content of the post rather than moderation or reply-worthiness.

  • The reply-generation context should include all comments in the target thread in chronological/thread sequence order.

  • Each thread comment context entry should include at least post_id, parent_id, author_label, created_at, post_summary, and full body.

  • For comments without a completed current-schema analysis, post_summary should be empty rather than blocking reply generation.

  • Full comment text should still be capped by a conservative per-comment and total-context budget to avoid unbounded prompt growth.

Implementation Plan

  1. Update the post analysis schema and prompt.

    • Add top-level post_summary to prompts/dedalus_post_analysis_system.txt.

    • Update DedalusPostAnalyzer::responseSchema() to require a top-level post_summary string.

    • Update DedalusPostAnalyzer::analyze() to return post_summary alongside the existing analysis sections.

    • Update tests that construct expected analysis payloads.

  2. Persist and hydrate post_summary.

    • Add a post_summary column to post_analyses.

    • Extend SqlitePostAnalysisStore::ensureSchema() with an ALTER TABLE migration for existing SQLite databases.

    • Save and hydrate post_summary in complete analysis rows.

    • Keep failed/config-missing analysis rows without a summary.

  3. Bump the analysis schema version.

    • Change analysis_schema_version from 2 to 3 in Application::postAnalysisContext().

    • This changes the content hash, preventing old cached analyses from being treated as complete for the new response shape.

  4. Display the neutral summary in Post analysis.

    • Add Post summary: near the top of the <details class="post-analysis"> panel.

    • Keep the existing moderation Summary: field but relabel it if needed to avoid ambiguity, for example Moderation summary:.

  5. Build full-thread reply context.

    • Add an application helper that builds thread_comments for a target post.

    • Fetch all posts in the thread with fetchThreadPosts($threadId).

    • For each comment, compute its current analysis content hash and look up the stored analysis.

    • Include the hydrated post_summary if present, plus full body text capped through a new limit helper.

    • Include metadata that lets the model distinguish the target post, parent post, root post, and existing reply-agent posts.

  6. Use thread context for generated replies.

    • Because reply generation currently posts analysis.engagement.suggested_response, update the analysis context before the initial analysis call so suggested responses can use all thread comments.

    • Also add thread_comments to $generationContext stored in post_generated_responses, so the persisted generation record shows the context used.

    • If we later re-enable DedalusAgentReplyGenerator::generate(), pass the same thread_comments context into that model call too.

  7. Gate context size.

    • Add per-comment body truncation for reply context, separate from post-analysis target body truncation.

    • Add a total thread comment budget so very large threads do not create oversized requests.

    • Prefer preserving the target post, root post, parent post, and latest comments if truncation is necessary.

  8. Verification.

    • Update DedalusPostAnalyzerTest for the new top-level post_summary shape.

    • Add or update store tests to confirm post_summary persists and hydrates.

    • Add an application-level test that reply/analysis context for a thread includes multiple comments with summaries and body text.

    • Run php tests/run.php.

Implementation Progress

  • Baseline committed in fa4e313 with the reviewed plan.

  • Slice 1: complete - added post_summary to the Dedalus post-analysis prompt/schema and direct analyzer tests. Verification: php tests/run.php.

  • Slice 2: complete - persisted and hydrated post_summary in SQLite analysis storage, including migration coverage and stub analyzer output. Verification: php tests/run.php.

  • Slice 3: complete - bumped analysis schema version to 3 and displayed post_summary in approved API/page analysis output with separate moderation summary labeling. Verification: php tests/run.php.

  • Slice 4: complete - added bounded full-thread thread_comments context to post analysis/generation inputs and persisted generated-response request context for auditing. Verification: php tests/run.php.

Open Decisions

  • Resolved: thread comment context uses 3000 chars per comment and 18000 chars total.

  • Resolved: the UI uses Post summary for the neutral content summary and Moderation summary for moderation-specific text.

  • Resolved: missing per-comment summaries remain empty; reply generation does not cascade provider calls to analyze missing comments.