zenmemes

New post

Feature plan: Step 1 · Step 2 · Step 3 · Step 4

Agent Response Modes Step 4 Implementation Summary

Stage 1 - Response-task catalog and contracts

  • Changes:

    • Added the five approved selectable response modes with stable types, labels, and descriptions.

    • Added per-mode generation instructions, including the initial supplied-context-only analysis mode.

    • Preserved default and legacy task compatibility while allowing selected-task storage and bounded generation input.

  • Verification:

    • php -l src/ForumRewrite/Agent/AgentResponseTask.php

    • php -l src/ForumRewrite/Agent/AgentResponseGenerator.php

    • ./v3 test AgentResponseTaskTest AgentResponseGeneratorTest — 9 passed.

  • Notes:

    • No request endpoint or presentation behavior changed in this stage; a selectable mode cannot reach generation until Stage 2 wiring is complete.

Stage 2 - Selected-mode request and fulfillment lifecycle

  • Changes:

    • Extended the existing request API with validated response_mode input while retaining the legacy default for callers that omit it.

    • Persisted the selected task through request completion and failure, surfaced its safe label in lifecycle responses, and fulfilled every supported task through mode-specific generation input.

    • Added end-to-end coverage for selected-mode request, invalid-mode rejection, durable task retention, and worker fulfillment; no schema migration was required.

  • Verification:

    • php -l src/ForumRewrite/Http/PostWorkflowApiController.php

    • php -l src/ForumRewrite/Agent/PostWorkflowService.php

    • php -l src/ForumRewrite/Agent/AgentReplyFulfillmentService.php

    • php -l src/ForumRewrite/Agent/SqliteAgentReplyGenerationStore.php

    • ./v3 test AgentReplyGenerationTest AgentResponseTaskTest AgentResponseGeneratorTest WriteApiSmokeTest — passed.

  • Notes:

    • Historical unmarked requests remain legacy analysis replies; request rows now retain a selected task after completion or failure so subsequent lifecycle feedback remains mode-aware.

Stage 3 - Shared response-mode chooser

  • Changes:

    • Replaced the direct request action on both canonical post-card surfaces with “Choose agent response,” backed by one server-rendered mode catalog and one shared client dialog.

    • Added named choices, descriptions, a supplied-context analysis notice, keyboard/escape-capable native dialog behavior, selected-mode feedback, and a narrow-screen bottom-sheet presentation.

    • Retained generic status text for historical default requests while selected-mode requests display their label through queued, completed, and failed lifecycle states.

  • Verification:

    • node --check public/assets/post_analysis.js

    • Focused smoke checks rendered one shared catalog and response chooser on both root and reply cards, hid only the requested card’s chooser, and displayed the selected-mode completion label.

    • ./v3 test AgentReplyGenerationTest AgentResponseTaskTest AgentResponseGeneratorTest WriteApiSmokeTest — passed.

    • ./v3 test — 720 run, 712 passed; the 8 failures are all classified by the suite as long-standing and outside this slice.

  • Notes:

    • The chooser uses the browser’s native dialog focus and Escape behavior. A final human browser pass at desktop and narrow viewport remains the release handoff check.

Stage 4 - Modal chooser refinement

  • Changes:

    • Made the native dialog presentation explicit with modal semantics, backdrop, and a visible close control.

    • Removed per-choice sub-captions; the chooser presented five mode labels with a shared analysis limitation notice.

  • Verification:

    • php -l src/ForumRewrite/Agent/AgentResponseTask.php

    • node --check public/assets/post_analysis.js

    • ./v3 test AgentResponseTaskTest AgentResponseGeneratorTest WriteApiSmokeTest::testApprovedViewerSeesAgentReplyRequestButtonUntilRequestExists WriteApiSmokeTest::testPostAnalysisScriptDoesNotExposeInProgressReplyGeneration — 11 passed.

  • Notes:

    • No request, fulfillment, or safety behavior changed.

Stage 5 - Dropdown chooser correction

  • Changes:

    • Replaced the modal with an anchored, non-modal dropdown menu.

    • Retained label-only choices and added Escape and outside-click dismissal with trigger state restoration.

  • Verification:

    • node --check public/assets/post_analysis.js

    • ./v3 test AgentResponseTaskTest AgentResponseGeneratorTest WriteApiSmokeTest::testApprovedViewerSeesAgentReplyRequestButtonUntilRequestExists WriteApiSmokeTest::testPostAnalysisScriptDoesNotExposeInProgressReplyGeneration — 11 passed.

  • Notes:

    • No request, fulfillment, or safety behavior changed.

Stage 6 - Server-wide request feature flag

  • Changes:

    • Added the private, server-wide AGENT_RESPONSE_REQUESTS_ENABLED flag, enabled by default and configurable through the private config file or environment.

    • Applied it to both canonical card surfaces and the request endpoint, so disabling it hides the dropdown and prevents direct requests from creating work.

    • Restored the dropdown trigger label to “Request agent response.”

  • Verification:

    • ./v3 test FeatureFlagEvaluatorTest WriteApiSmokeTest::testAgentResponseRequestFlagHidesChooserAndRejectsRequests WriteApiSmokeTest::testApprovedViewerSeesAgentReplyRequestButtonUntilRequestExists WriteApiSmokeTest::testGenerateAgentReplyPersistsAndFulfillsSelectedResponseMode — 17 passed.

  • Notes:

    • Set AGENT_RESPONSE_REQUESTS_ENABLED=false in the server private config or environment to disable the feature everywhere; re-enable it with true or by removing the override.

Stage 7 - Logic analysis replacement

  • Changes:

    • Replaced the selectable facts_analysis mode with logic_analysis, labelled “Logic analysis.”

    • Replaced evidence-verification instructions with a charitable analysis of premises, conclusions, assumptions, and logical gaps.

    • Removed the dropdown’s facts-analysis disclaimer and its unused styling.

    • Kept the old task identifier readable and fulfillable solely for requests already stored before this change; new API requests accept only selectable current modes.

  • Verification:

    • Focused task, generator, request API, and feature-flag tests cover the new mode, rejection of no-longer-selectable modes, and legacy fulfillment compatibility.

Stage 8 - External response-mode prompts

  • Changes:

    • Moved the shared system prompt and every response-task instruction into individually reviewable files under prompts/.

    • The generator now loads, trims, validates, and caches those files using the same repository-root convention as other prompt-backed features.

  • Verification:

    • Generator coverage exercises every selectable task and verifies each loaded instruction.

Stage 9 - Readable plain-text task prompts

  • Changes:

    • Replaced the dense all-caps/delimiter layout with readable plain-text headings for the task, response instructions, and forum context.

    • Kept each supplied forum field explicitly enclosed in forum-content markers and stated that those sections are untrusted text, preserving the prompt-injection boundary.

Stage 10 - Legacy automatic-response flag separation

  • Changes:

    • Made DEDALUS_AGENT_REPLIES_AUTOMATIC_ENABLED the complete gate for the legacy analysis-suggested response path, including direct analysis requests as well as page-triggered work.

    • Clarified in the private feature-flag catalog that it is independent from AGENT_RESPONSE_REQUESTS_ENABLED, which controls reader-selected response tasks.

  • Operational note:

    • Set DEDALUS_AGENT_REPLIES_AUTOMATIC_ENABLED=false to disable legacy automatic replies while retaining the “Request agent response” feature.