@posthog/agent 1.20.0 → 1.22.0

package/dist/src/workflow/steps/research.js

@@ -58,9 +58,9 @@ const researchStep = async ({ step, context }) => {
     let jsonContent = '';
     for await (const message of response) {
         emitEvent(adapter.createRawSDKEvent(message));
-        const transformed = adapter.transform(message);
-        if (transformed) {
-            emitEvent(transformed);
+        const transformedEvents = adapter.transform(message);
+        for (const event of transformedEvents) {
+            emitEvent(event);
         }
         if (message.type === 'assistant' && message.message?.content) {
             for (const c of message.message.content) {

package/dist/src/workflow/steps/research.js.map

@@ -1 +1 @@
-{"version":3,"file":"research.js","sources":["../../../../src/workflow/steps/research.ts"],"sourcesContent":["import { query } from '@anthropic-ai/claude-agent-sdk';\nimport { RESEARCH_SYSTEM_PROMPT } from '../../agents/research.js';\nimport type { WorkflowStepRunner } from '../types.js';\nimport type { ResearchEvaluation } from '../../types.js';\nimport { finalizeStepGitActions } from '../utils.js';\n\nexport const researchStep: WorkflowStepRunner = async ({ step, context }) => {\n    const {\n        task,\n        cwd,\n        isCloudMode,\n        options,\n        logger,\n        fileManager,\n        gitManager,\n        promptBuilder,\n        adapter,\n        mcpServers,\n        emitEvent,\n    } = context;\n\n    const stepLogger = logger.child('ResearchStep');\n\n    const existingResearch = await fileManager.readResearch(task.id);\n    if (existingResearch) {\n        stepLogger.info('Research already exists', { taskId: task.id, hasQuestions: !!existingResearch.questions, answered: existingResearch.answered });\n        \n        // If there are unanswered questions, re-emit them so UI can prompt user\n        if (existingResearch.questions && !existingResearch.answered) {\n            stepLogger.info('Re-emitting unanswered research questions', { \n                taskId: task.id,\n                questionCount: existingResearch.questions.length \n            });\n            \n            emitEvent({\n                type: 'artifact',\n                ts: Date.now(),\n                kind: 'research_questions',\n                content: existingResearch.questions,\n            });\n            \n            // In local mode, halt to allow user to answer\n            if (!isCloudMode) {\n                emitEvent(adapter.createStatusEvent('phase_complete', { phase: 'research' }));\n                return { status: 'skipped', halt: true };\n            }\n        }\n        \n        return { status: 'skipped' };\n    }\n\n    stepLogger.info('Starting research phase', { taskId: task.id });\n    emitEvent(adapter.createStatusEvent('phase_start', { phase: 'research' }));\n\n    const researchPrompt = await promptBuilder.buildResearchPrompt(task, cwd);\n    const fullPrompt = `${RESEARCH_SYSTEM_PROMPT}\\n\\n${researchPrompt}`;\n\n    const baseOptions: Record<string, any> = {\n        model: step.model,\n        cwd,\n        permissionMode: 'plan',\n        settingSources: ['local'],\n        mcpServers,\n        // Allow research tools: read-only operations, web search, and MCP resources\n        allowedTools: [\n            'Read',\n            'Glob',\n            'Grep',\n            'WebFetch',\n            'WebSearch',\n            'ListMcpResources',\n            'ReadMcpResource',\n            'TodoWrite',\n            'BashOutput',\n        ],\n    };\n\n    const response = query({\n        prompt: fullPrompt,\n        options: { ...baseOptions, ...(options.queryOverrides || {}) },\n    });\n\n    let jsonContent = '';\n    for await (const message of response) {\n        emitEvent(adapter.createRawSDKEvent(message));\n        const transformed = adapter.transform(message);\n        if (transformed) {\n            emitEvent(transformed);\n        }\n        if (message.type === 'assistant' && message.message?.content) {\n            for (const c of message.message.content) {\n                if (c.type === 'text' && c.text) {\n                    jsonContent += c.text;\n                }\n            }\n        }\n    }\n\n    if (!jsonContent.trim()) {\n        stepLogger.error('No JSON output from research agent', { taskId: task.id });\n        emitEvent({\n            type: 'error',\n            ts: Date.now(),\n            message: 'Research agent returned no output',\n        });\n        return { status: 'completed', halt: true };\n    }\n\n    // Parse JSON response\n    let evaluation: ResearchEvaluation;\n    try {\n        // Extract JSON from potential markdown code blocks or other wrapping\n        const jsonMatch = jsonContent.match(/\\{[\\s\\S]*\\}/);\n        if (!jsonMatch) {\n            throw new Error('No JSON object found in response');\n        }\n        evaluation = JSON.parse(jsonMatch[0]);\n        stepLogger.info('Parsed research evaluation', {\n            taskId: task.id,\n            score: evaluation.actionabilityScore,\n            hasQuestions: !!evaluation.questions,\n        });\n    } catch (error) {\n        stepLogger.error('Failed to parse research JSON', {\n            taskId: task.id,\n            error: error instanceof Error ? error.message : String(error),\n            content: jsonContent.substring(0, 500),\n        });\n        emitEvent({\n            type: 'error',\n            ts: Date.now(),\n            message: `Failed to parse research JSON: ${\n                error instanceof Error ? error.message : String(error)\n            }`,\n        });\n        return { status: 'completed', halt: true };\n    }\n\n    // Add answered/answers fields to evaluation\n    if (evaluation.questions && evaluation.questions.length > 0) {\n        evaluation.answered = false;\n        evaluation.answers = undefined;\n    }\n\n    // Always write research.json\n    await fileManager.writeResearch(task.id, evaluation);\n    stepLogger.info('Research evaluation written', {\n        taskId: task.id,\n        score: evaluation.actionabilityScore,\n        hasQuestions: !!evaluation.questions,\n    });\n\n    emitEvent({\n        type: 'artifact',\n        ts: Date.now(),\n        kind: 'research_evaluation',\n        content: evaluation,\n    });\n\n    await gitManager.addAllPostHogFiles();\n    await finalizeStepGitActions(context, step, {\n        commitMessage: `Research phase for ${task.title}`,\n    });\n\n    // Log whether questions need answering\n    if (evaluation.actionabilityScore < 0.7 && evaluation.questions && evaluation.questions.length > 0) {\n        stepLogger.info('Actionability score below threshold, questions needed', {\n            taskId: task.id,\n            score: evaluation.actionabilityScore,\n            questionCount: evaluation.questions.length,\n        });\n        \n        emitEvent({\n            type: 'artifact',\n            ts: Date.now(),\n            kind: 'research_questions',\n            content: evaluation.questions,\n        });\n    } else {\n        stepLogger.info('Actionability score acceptable, proceeding to planning', {\n            taskId: task.id,\n            score: evaluation.actionabilityScore,\n        });\n    }\n\n    // In local mode, always halt after research for user review\n    if (!isCloudMode) {\n        emitEvent(adapter.createStatusEvent('phase_complete', { phase: 'research' }));\n        return { status: 'completed', halt: true };\n    }\n\n    // In cloud mode, check if questions need answering\n    const researchData = await fileManager.readResearch(task.id);\n    if (researchData?.questions && !researchData.answered) {\n        // Questions need answering - halt for user input in cloud mode too\n        emitEvent(adapter.createStatusEvent('phase_complete', { phase: 'research' }));\n        return { status: 'completed', halt: true };\n    }\n\n    // No questions or questions already answered - proceed to planning\n    emitEvent(adapter.createStatusEvent('phase_complete', { phase: 'research' }));\n    return { status: 'completed' };\n};\n"],"names":[],"mappings":";;;;AAMO,MAAM,YAAY,GAAuB,OAAO,EAAE,IAAI,EAAE,OAAO,EAAE,KAAI;IACxE,MAAM,EACF,IAAI,EACJ,GAAG,EACH,WAAW,EACX,OAAO,EACP,MAAM,EACN,WAAW,EACX,UAAU,EACV,aAAa,EACb,OAAO,EACP,UAAU,EACV,SAAS,GACZ,GAAG,OAAO;IAEX,MAAM,UAAU,GAAG,MAAM,CAAC,KAAK,CAAC,cAAc,CAAC;IAE/C,MAAM,gBAAgB,GAAG,MAAM,WAAW,CAAC,YAAY,CAAC,IAAI,CAAC,EAAE,CAAC;IAChE,IAAI,gBAAgB,EAAE;QAClB,UAAU,CAAC,IAAI,CAAC,yBAAyB,EAAE,EAAE,MAAM,EAAE,IAAI,CAAC,EAAE,EAAE,YAAY,EAAE,CAAC,CAAC,gBAAgB,CAAC,SAAS,EAAE,QAAQ,EAAE,gBAAgB,CAAC,QAAQ,EAAE,CAAC;;QAGhJ,IAAI,gBAAgB,CAAC,SAAS,IAAI,CAAC,gBAAgB,CAAC,QAAQ,EAAE;AAC1D,YAAA,UAAU,CAAC,IAAI,CAAC,2CAA2C,EAAE;gBACzD,MAAM,EAAE,IAAI,CAAC,EAAE;AACf,gBAAA,aAAa,EAAE,gBAAgB,CAAC,SAAS,CAAC;AAC7C,aAAA,CAAC;AAEF,YAAA,SAAS,CAAC;AACN,gBAAA,IAAI,EAAE,UAAU;AAChB,gBAAA,EAAE,EAAE,IAAI,CAAC,GAAG,EAAE;AACd,gBAAA,IAAI,EAAE,oBAAoB;gBAC1B,OAAO,EAAE,gBAAgB,CAAC,SAAS;AACtC,aAAA,CAAC;;YAGF,IAAI,CAAC,WAAW,EAAE;AACd,gBAAA,SAAS,CAAC,OAAO,CAAC,iBAAiB,CAAC,gBAAgB,EAAE,EAAE,KAAK,EAAE,UAAU,EAAE,CAAC,CAAC;gBAC7E,OAAO,EAAE,MAAM,EAAE,SAAS,EAAE,IAAI,EAAE,IAAI,EAAE;YAC5C;QACJ;AAEA,QAAA,OAAO,EAAE,MAAM,EAAE,SAAS,EAAE;IAChC;AAEA,IAAA,UAAU,CAAC,IAAI,CAAC,yBAAyB,EAAE,EAAE,MAAM,EAAE,IAAI,CAAC,EAAE,EAAE,CAAC;AAC/D,IAAA,SAAS,CAAC,OAAO,CAAC,iBAAiB,CAAC,aAAa,EAAE,EAAE,KAAK,EAAE,UAAU,EAAE,CAAC,CAAC;IAE1E,MAAM,cAAc,GAAG,MAAM,aAAa,CAAC,mBAAmB,CAAC,IAAI,EAAE,GAAG,CAAC;AACzE,IAAA,MAAM,UAAU,GAAG,CAAA,EAAG,sBAAsB,CAAA,IAAA,EAAO,cAAc,EAAE;AAEnE,IAAA,MAAM,WAAW,GAAwB;QACrC,KAAK,EAAE,IAAI,CAAC,KAAK;QACjB,GAAG;AACH,QAAA,cAAc,EAAE,MAAM;QACtB,cAAc,EAAE,CAAC,OAAO,CAAC;QACzB,UAAU;;AAEV,QAAA,YAAY,EAAE;YACV,MAAM;YACN,MAAM;YACN,MAAM;YACN,UAAU;YACV,WAAW;YACX,kBAAkB;YAClB,iBAAiB;YACjB,WAAW;YACX,YAAY;AACf,SAAA;KACJ;IAED,MAAM,QAAQ,GAAG,KAAK,CAAC;AACnB,QAAA,MAAM,EAAE,UAAU;AAClB,QAAA,OAAO,EAAE,EAAE,GAAG,WAAW,EAAE,IAAI,OAAO,CAAC,cAAc,IAAI,EAAE,CAAC,EAAE;AACjE,KAAA,CAAC;IAEF,IAAI,WAAW,GAAG,EAAE;AACpB,IAAA,WAAW,MAAM,OAAO,IAAI,QAAQ,EAAE;QAClC,SAAS,CAAC,OAAO,CAAC,iBAAiB,CAAC,OAAO,CAAC,CAAC;QAC7C,MAAM,WAAW,GAAG,OAAO,CAAC,SAAS,CAAC,OAAO,CAAC;QAC9C,IAAI,WAAW,EAAE;YACb,SAAS,CAAC,WAAW,CAAC;QAC1B;AACA,QAAA,IAAI,OAAO,CAAC,IAAI,KAAK,WAAW,IAAI,OAAO,CAAC,OAAO,EAAE,OAAO,EAAE;YAC1D,KAAK,MAAM,CAAC,IAAI,OAAO,CAAC,OAAO,CAAC,OAAO,EAAE;gBACrC,IAAI,CAAC,CAAC,IAAI,KAAK,MAAM,IAAI,CAAC,CAAC,IAAI,EAAE;AAC7B,oBAAA,WAAW,IAAI,CAAC,CAAC,IAAI;gBACzB;YACJ;QACJ;IACJ;AAEA,IAAA,IAAI,CAAC,WAAW,CAAC,IAAI,EAAE,EAAE;AACrB,QAAA,UAAU,CAAC,KAAK,CAAC,oCAAoC,EAAE,EAAE,MAAM,EAAE,IAAI,CAAC,EAAE,EAAE,CAAC;AAC3E,QAAA,SAAS,CAAC;AACN,YAAA,IAAI,EAAE,OAAO;AACb,YAAA,EAAE,EAAE,IAAI,CAAC,GAAG,EAAE;AACd,YAAA,OAAO,EAAE,mCAAmC;AAC/C,SAAA,CAAC;QACF,OAAO,EAAE,MAAM,EAAE,WAAW,EAAE,IAAI,EAAE,IAAI,EAAE;IAC9C;;AAGA,IAAA,IAAI,UAA8B;AAClC,IAAA,IAAI;;QAEA,MAAM,SAAS,GAAG,WAAW,CAAC,KAAK,CAAC,aAAa,CAAC;QAClD,IAAI,CAAC,SAAS,EAAE;AACZ,YAAA,MAAM,IAAI,KAAK,CAAC,kCAAkC,CAAC;QACvD;QACA,UAAU,GAAG,IAAI,CAAC,KAAK,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC;AACrC,QAAA,UAAU,CAAC,IAAI,CAAC,4BAA4B,EAAE;YAC1C,MAAM,EAAE,IAAI,CAAC,EAAE;YACf,KAAK,EAAE,UAAU,CAAC,kBAAkB;AACpC,YAAA,YAAY,EAAE,CAAC,CAAC,UAAU,CAAC,SAAS;AACvC,SAAA,CAAC;IACN;IAAE,OAAO,KAAK,EAAE;AACZ,QAAA,UAAU,CAAC,KAAK,CAAC,+BAA+B,EAAE;YAC9C,MAAM,EAAE,IAAI,CAAC,EAAE;AACf,YAAA,KAAK,EAAE,KAAK,YAAY,KAAK,GAAG,KAAK,CAAC,OAAO,GAAG,MAAM,CAAC,KAAK,CAAC;YAC7D,OAAO,EAAE,WAAW,CAAC,SAAS,CAAC,CAAC,EAAE,GAAG,CAAC;AACzC,SAAA,CAAC;AACF,QAAA,SAAS,CAAC;AACN,YAAA,IAAI,EAAE,OAAO;AACb,YAAA,EAAE,EAAE,IAAI,CAAC,GAAG,EAAE;AACd,YAAA,OAAO,EAAE,CAAA,+BAAA,EACL,KAAK,YAAY,KAAK,GAAG,KAAK,CAAC,OAAO,GAAG,MAAM,CAAC,KAAK,CACzD,CAAA,CAAE;AACL,SAAA,CAAC;QACF,OAAO,EAAE,MAAM,EAAE,WAAW,EAAE,IAAI,EAAE,IAAI,EAAE;IAC9C;;AAGA,IAAA,IAAI,UAAU,CAAC,SAAS,IAAI,UAAU,CAAC,SAAS,CAAC,MAAM,GAAG,CAAC,EAAE;AACzD,QAAA,UAAU,CAAC,QAAQ,GAAG,KAAK;AAC3B,QAAA,UAAU,CAAC,OAAO,GAAG,SAAS;IAClC;;IAGA,MAAM,WAAW,CAAC,aAAa,CAAC,IAAI,CAAC,EAAE,EAAE,UAAU,CAAC;AACpD,IAAA,UAAU,CAAC,IAAI,CAAC,6BAA6B,EAAE;QAC3C,MAAM,EAAE,IAAI,CAAC,EAAE;QACf,KAAK,EAAE,UAAU,CAAC,kBAAkB;AACpC,QAAA,YAAY,EAAE,CAAC,CAAC,UAAU,CAAC,SAAS;AACvC,KAAA,CAAC;AAEF,IAAA,SAAS,CAAC;AACN,QAAA,IAAI,EAAE,UAAU;AAChB,QAAA,EAAE,EAAE,IAAI,CAAC,GAAG,EAAE;AACd,QAAA,IAAI,EAAE,qBAAqB;AAC3B,QAAA,OAAO,EAAE,UAAU;AACtB,KAAA,CAAC;AAEF,IAAA,MAAM,UAAU,CAAC,kBAAkB,EAAE;AACrC,IAAA,MAAM,sBAAsB,CAAC,OAAO,EAAE,IAAI,EAAE;AACxC,QAAA,aAAa,EAAE,CAAA,mBAAA,EAAsB,IAAI,CAAC,KAAK,CAAA,CAAE;AACpD,KAAA,CAAC;;AAGF,IAAA,IAAI,UAAU,CAAC,kBAAkB,GAAG,GAAG,IAAI,UAAU,CAAC,SAAS,IAAI,UAAU,CAAC,SAAS,CAAC,MAAM,GAAG,CAAC,EAAE;AAChG,QAAA,UAAU,CAAC,IAAI,CAAC,uDAAuD,EAAE;YACrE,MAAM,EAAE,IAAI,CAAC,EAAE;YACf,KAAK,EAAE,UAAU,CAAC,kBAAkB;AACpC,YAAA,aAAa,EAAE,UAAU,CAAC,SAAS,CAAC,MAAM;AAC7C,SAAA,CAAC;AAEF,QAAA,SAAS,CAAC;AACN,YAAA,IAAI,EAAE,UAAU;AAChB,YAAA,EAAE,EAAE,IAAI,CAAC,GAAG,EAAE;AACd,YAAA,IAAI,EAAE,oBAAoB;YAC1B,OAAO,EAAE,UAAU,CAAC,SAAS;AAChC,SAAA,CAAC;IACN;SAAO;AACH,QAAA,UAAU,CAAC,IAAI,CAAC,wDAAwD,EAAE;YACtE,MAAM,EAAE,IAAI,CAAC,EAAE;YACf,KAAK,EAAE,UAAU,CAAC,kBAAkB;AACvC,SAAA,CAAC;IACN;;IAGA,IAAI,CAAC,WAAW,EAAE;AACd,QAAA,SAAS,CAAC,OAAO,CAAC,iBAAiB,CAAC,gBAAgB,EAAE,EAAE,KAAK,EAAE,UAAU,EAAE,CAAC,CAAC;QAC7E,OAAO,EAAE,MAAM,EAAE,WAAW,EAAE,IAAI,EAAE,IAAI,EAAE;IAC9C;;IAGA,MAAM,YAAY,GAAG,MAAM,WAAW,CAAC,YAAY,CAAC,IAAI,CAAC,EAAE,CAAC;IAC5D,IAAI,YAAY,EAAE,SAAS,IAAI,CAAC,YAAY,CAAC,QAAQ,EAAE;;AAEnD,QAAA,SAAS,CAAC,OAAO,CAAC,iBAAiB,CAAC,gBAAgB,EAAE,EAAE,KAAK,EAAE,UAAU,EAAE,CAAC,CAAC;QAC7E,OAAO,EAAE,MAAM,EAAE,WAAW,EAAE,IAAI,EAAE,IAAI,EAAE;IAC9C;;AAGA,IAAA,SAAS,CAAC,OAAO,CAAC,iBAAiB,CAAC,gBAAgB,EAAE,EAAE,KAAK,EAAE,UAAU,EAAE,CAAC,CAAC;AAC7E,IAAA,OAAO,EAAE,MAAM,EAAE,WAAW,EAAE;AAClC;;;;"}
+{"version":3,"file":"research.js","sources":["../../../../src/workflow/steps/research.ts"],"sourcesContent":["import { query } from '@anthropic-ai/claude-agent-sdk';\nimport { RESEARCH_SYSTEM_PROMPT } from '../../agents/research.js';\nimport type { WorkflowStepRunner } from '../types.js';\nimport type { ResearchEvaluation } from '../../types.js';\nimport { finalizeStepGitActions } from '../utils.js';\n\nexport const researchStep: WorkflowStepRunner = async ({ step, context }) => {\n    const {\n        task,\n        cwd,\n        isCloudMode,\n        options,\n        logger,\n        fileManager,\n        gitManager,\n        promptBuilder,\n        adapter,\n        mcpServers,\n        emitEvent,\n    } = context;\n\n    const stepLogger = logger.child('ResearchStep');\n\n    const existingResearch = await fileManager.readResearch(task.id);\n    if (existingResearch) {\n        stepLogger.info('Research already exists', { taskId: task.id, hasQuestions: !!existingResearch.questions, answered: existingResearch.answered });\n        \n        // If there are unanswered questions, re-emit them so UI can prompt user\n        if (existingResearch.questions && !existingResearch.answered) {\n            stepLogger.info('Re-emitting unanswered research questions', { \n                taskId: task.id,\n                questionCount: existingResearch.questions.length \n            });\n            \n            emitEvent({\n                type: 'artifact',\n                ts: Date.now(),\n                kind: 'research_questions',\n                content: existingResearch.questions,\n            });\n            \n            // In local mode, halt to allow user to answer\n            if (!isCloudMode) {\n                emitEvent(adapter.createStatusEvent('phase_complete', { phase: 'research' }));\n                return { status: 'skipped', halt: true };\n            }\n        }\n        \n        return { status: 'skipped' };\n    }\n\n    stepLogger.info('Starting research phase', { taskId: task.id });\n    emitEvent(adapter.createStatusEvent('phase_start', { phase: 'research' }));\n\n    const researchPrompt = await promptBuilder.buildResearchPrompt(task, cwd);\n    const fullPrompt = `${RESEARCH_SYSTEM_PROMPT}\\n\\n${researchPrompt}`;\n\n    const baseOptions: Record<string, any> = {\n        model: step.model,\n        cwd,\n        permissionMode: 'plan',\n        settingSources: ['local'],\n        mcpServers,\n        // Allow research tools: read-only operations, web search, and MCP resources\n        allowedTools: [\n            'Read',\n            'Glob',\n            'Grep',\n            'WebFetch',\n            'WebSearch',\n            'ListMcpResources',\n            'ReadMcpResource',\n            'TodoWrite',\n            'BashOutput',\n        ],\n    };\n\n    const response = query({\n        prompt: fullPrompt,\n        options: { ...baseOptions, ...(options.queryOverrides || {}) },\n    });\n\n    let jsonContent = '';\n    for await (const message of response) {\n        emitEvent(adapter.createRawSDKEvent(message));\n        const transformedEvents = adapter.transform(message);\n        for (const event of transformedEvents) {\n            emitEvent(event);\n        }\n        if (message.type === 'assistant' && message.message?.content) {\n            for (const c of message.message.content) {\n                if (c.type === 'text' && c.text) {\n                    jsonContent += c.text;\n                }\n            }\n        }\n    }\n\n    if (!jsonContent.trim()) {\n        stepLogger.error('No JSON output from research agent', { taskId: task.id });\n        emitEvent({\n            type: 'error',\n            ts: Date.now(),\n            message: 'Research agent returned no output',\n        });\n        return { status: 'completed', halt: true };\n    }\n\n    // Parse JSON response\n    let evaluation: ResearchEvaluation;\n    try {\n        // Extract JSON from potential markdown code blocks or other wrapping\n        const jsonMatch = jsonContent.match(/\\{[\\s\\S]*\\}/);\n        if (!jsonMatch) {\n            throw new Error('No JSON object found in response');\n        }\n        evaluation = JSON.parse(jsonMatch[0]);\n        stepLogger.info('Parsed research evaluation', {\n            taskId: task.id,\n            score: evaluation.actionabilityScore,\n            hasQuestions: !!evaluation.questions,\n        });\n    } catch (error) {\n        stepLogger.error('Failed to parse research JSON', {\n            taskId: task.id,\n            error: error instanceof Error ? error.message : String(error),\n            content: jsonContent.substring(0, 500),\n        });\n        emitEvent({\n            type: 'error',\n            ts: Date.now(),\n            message: `Failed to parse research JSON: ${\n                error instanceof Error ? error.message : String(error)\n            }`,\n        });\n        return { status: 'completed', halt: true };\n    }\n\n    // Add answered/answers fields to evaluation\n    if (evaluation.questions && evaluation.questions.length > 0) {\n        evaluation.answered = false;\n        evaluation.answers = undefined;\n    }\n\n    // Always write research.json\n    await fileManager.writeResearch(task.id, evaluation);\n    stepLogger.info('Research evaluation written', {\n        taskId: task.id,\n        score: evaluation.actionabilityScore,\n        hasQuestions: !!evaluation.questions,\n    });\n\n    emitEvent({\n        type: 'artifact',\n        ts: Date.now(),\n        kind: 'research_evaluation',\n        content: evaluation,\n    });\n\n    await gitManager.addAllPostHogFiles();\n    await finalizeStepGitActions(context, step, {\n        commitMessage: `Research phase for ${task.title}`,\n    });\n\n    // Log whether questions need answering\n    if (evaluation.actionabilityScore < 0.7 && evaluation.questions && evaluation.questions.length > 0) {\n        stepLogger.info('Actionability score below threshold, questions needed', {\n            taskId: task.id,\n            score: evaluation.actionabilityScore,\n            questionCount: evaluation.questions.length,\n        });\n        \n        emitEvent({\n            type: 'artifact',\n            ts: Date.now(),\n            kind: 'research_questions',\n            content: evaluation.questions,\n        });\n    } else {\n        stepLogger.info('Actionability score acceptable, proceeding to planning', {\n            taskId: task.id,\n            score: evaluation.actionabilityScore,\n        });\n    }\n\n    // In local mode, always halt after research for user review\n    if (!isCloudMode) {\n        emitEvent(adapter.createStatusEvent('phase_complete', { phase: 'research' }));\n        return { status: 'completed', halt: true };\n    }\n\n    // In cloud mode, check if questions need answering\n    const researchData = await fileManager.readResearch(task.id);\n    if (researchData?.questions && !researchData.answered) {\n        // Questions need answering - halt for user input in cloud mode too\n        emitEvent(adapter.createStatusEvent('phase_complete', { phase: 'research' }));\n        return { status: 'completed', halt: true };\n    }\n\n    // No questions or questions already answered - proceed to planning\n    emitEvent(adapter.createStatusEvent('phase_complete', { phase: 'research' }));\n    return { status: 'completed' };\n};\n"],"names":[],"mappings":";;;;AAMO,MAAM,YAAY,GAAuB,OAAO,EAAE,IAAI,EAAE,OAAO,EAAE,KAAI;IACxE,MAAM,EACF,IAAI,EACJ,GAAG,EACH,WAAW,EACX,OAAO,EACP,MAAM,EACN,WAAW,EACX,UAAU,EACV,aAAa,EACb,OAAO,EACP,UAAU,EACV,SAAS,GACZ,GAAG,OAAO;IAEX,MAAM,UAAU,GAAG,MAAM,CAAC,KAAK,CAAC,cAAc,CAAC;IAE/C,MAAM,gBAAgB,GAAG,MAAM,WAAW,CAAC,YAAY,CAAC,IAAI,CAAC,EAAE,CAAC;IAChE,IAAI,gBAAgB,EAAE;QAClB,UAAU,CAAC,IAAI,CAAC,yBAAyB,EAAE,EAAE,MAAM,EAAE,IAAI,CAAC,EAAE,EAAE,YAAY,EAAE,CAAC,CAAC,gBAAgB,CAAC,SAAS,EAAE,QAAQ,EAAE,gBAAgB,CAAC,QAAQ,EAAE,CAAC;;QAGhJ,IAAI,gBAAgB,CAAC,SAAS,IAAI,CAAC,gBAAgB,CAAC,QAAQ,EAAE;AAC1D,YAAA,UAAU,CAAC,IAAI,CAAC,2CAA2C,EAAE;gBACzD,MAAM,EAAE,IAAI,CAAC,EAAE;AACf,gBAAA,aAAa,EAAE,gBAAgB,CAAC,SAAS,CAAC;AAC7C,aAAA,CAAC;AAEF,YAAA,SAAS,CAAC;AACN,gBAAA,IAAI,EAAE,UAAU;AAChB,gBAAA,EAAE,EAAE,IAAI,CAAC,GAAG,EAAE;AACd,gBAAA,IAAI,EAAE,oBAAoB;gBAC1B,OAAO,EAAE,gBAAgB,CAAC,SAAS;AACtC,aAAA,CAAC;;YAGF,IAAI,CAAC,WAAW,EAAE;AACd,gBAAA,SAAS,CAAC,OAAO,CAAC,iBAAiB,CAAC,gBAAgB,EAAE,EAAE,KAAK,EAAE,UAAU,EAAE,CAAC,CAAC;gBAC7E,OAAO,EAAE,MAAM,EAAE,SAAS,EAAE,IAAI,EAAE,IAAI,EAAE;YAC5C;QACJ;AAEA,QAAA,OAAO,EAAE,MAAM,EAAE,SAAS,EAAE;IAChC;AAEA,IAAA,UAAU,CAAC,IAAI,CAAC,yBAAyB,EAAE,EAAE,MAAM,EAAE,IAAI,CAAC,EAAE,EAAE,CAAC;AAC/D,IAAA,SAAS,CAAC,OAAO,CAAC,iBAAiB,CAAC,aAAa,EAAE,EAAE,KAAK,EAAE,UAAU,EAAE,CAAC,CAAC;IAE1E,MAAM,cAAc,GAAG,MAAM,aAAa,CAAC,mBAAmB,CAAC,IAAI,EAAE,GAAG,CAAC;AACzE,IAAA,MAAM,UAAU,GAAG,CAAA,EAAG,sBAAsB,CAAA,IAAA,EAAO,cAAc,EAAE;AAEnE,IAAA,MAAM,WAAW,GAAwB;QACrC,KAAK,EAAE,IAAI,CAAC,KAAK;QACjB,GAAG;AACH,QAAA,cAAc,EAAE,MAAM;QACtB,cAAc,EAAE,CAAC,OAAO,CAAC;QACzB,UAAU;;AAEV,QAAA,YAAY,EAAE;YACV,MAAM;YACN,MAAM;YACN,MAAM;YACN,UAAU;YACV,WAAW;YACX,kBAAkB;YAClB,iBAAiB;YACjB,WAAW;YACX,YAAY;AACf,SAAA;KACJ;IAED,MAAM,QAAQ,GAAG,KAAK,CAAC;AACnB,QAAA,MAAM,EAAE,UAAU;AAClB,QAAA,OAAO,EAAE,EAAE,GAAG,WAAW,EAAE,IAAI,OAAO,CAAC,cAAc,IAAI,EAAE,CAAC,EAAE;AACjE,KAAA,CAAC;IAEF,IAAI,WAAW,GAAG,EAAE;AACpB,IAAA,WAAW,MAAM,OAAO,IAAI,QAAQ,EAAE;QAClC,SAAS,CAAC,OAAO,CAAC,iBAAiB,CAAC,OAAO,CAAC,CAAC;QAC7C,MAAM,iBAAiB,GAAG,OAAO,CAAC,SAAS,CAAC,OAAO,CAAC;AACpD,QAAA,KAAK,MAAM,KAAK,IAAI,iBAAiB,EAAE;YACnC,SAAS,CAAC,KAAK,CAAC;QACpB;AACA,QAAA,IAAI,OAAO,CAAC,IAAI,KAAK,WAAW,IAAI,OAAO,CAAC,OAAO,EAAE,OAAO,EAAE;YAC1D,KAAK,MAAM,CAAC,IAAI,OAAO,CAAC,OAAO,CAAC,OAAO,EAAE;gBACrC,IAAI,CAAC,CAAC,IAAI,KAAK,MAAM,IAAI,CAAC,CAAC,IAAI,EAAE;AAC7B,oBAAA,WAAW,IAAI,CAAC,CAAC,IAAI;gBACzB;YACJ;QACJ;IACJ;AAEA,IAAA,IAAI,CAAC,WAAW,CAAC,IAAI,EAAE,EAAE;AACrB,QAAA,UAAU,CAAC,KAAK,CAAC,oCAAoC,EAAE,EAAE,MAAM,EAAE,IAAI,CAAC,EAAE,EAAE,CAAC;AAC3E,QAAA,SAAS,CAAC;AACN,YAAA,IAAI,EAAE,OAAO;AACb,YAAA,EAAE,EAAE,IAAI,CAAC,GAAG,EAAE;AACd,YAAA,OAAO,EAAE,mCAAmC;AAC/C,SAAA,CAAC;QACF,OAAO,EAAE,MAAM,EAAE,WAAW,EAAE,IAAI,EAAE,IAAI,EAAE;IAC9C;;AAGA,IAAA,IAAI,UAA8B;AAClC,IAAA,IAAI;;QAEA,MAAM,SAAS,GAAG,WAAW,CAAC,KAAK,CAAC,aAAa,CAAC;QAClD,IAAI,CAAC,SAAS,EAAE;AACZ,YAAA,MAAM,IAAI,KAAK,CAAC,kCAAkC,CAAC;QACvD;QACA,UAAU,GAAG,IAAI,CAAC,KAAK,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC;AACrC,QAAA,UAAU,CAAC,IAAI,CAAC,4BAA4B,EAAE;YAC1C,MAAM,EAAE,IAAI,CAAC,EAAE;YACf,KAAK,EAAE,UAAU,CAAC,kBAAkB;AACpC,YAAA,YAAY,EAAE,CAAC,CAAC,UAAU,CAAC,SAAS;AACvC,SAAA,CAAC;IACN;IAAE,OAAO,KAAK,EAAE;AACZ,QAAA,UAAU,CAAC,KAAK,CAAC,+BAA+B,EAAE;YAC9C,MAAM,EAAE,IAAI,CAAC,EAAE;AACf,YAAA,KAAK,EAAE,KAAK,YAAY,KAAK,GAAG,KAAK,CAAC,OAAO,GAAG,MAAM,CAAC,KAAK,CAAC;YAC7D,OAAO,EAAE,WAAW,CAAC,SAAS,CAAC,CAAC,EAAE,GAAG,CAAC;AACzC,SAAA,CAAC;AACF,QAAA,SAAS,CAAC;AACN,YAAA,IAAI,EAAE,OAAO;AACb,YAAA,EAAE,EAAE,IAAI,CAAC,GAAG,EAAE;AACd,YAAA,OAAO,EAAE,CAAA,+BAAA,EACL,KAAK,YAAY,KAAK,GAAG,KAAK,CAAC,OAAO,GAAG,MAAM,CAAC,KAAK,CACzD,CAAA,CAAE;AACL,SAAA,CAAC;QACF,OAAO,EAAE,MAAM,EAAE,WAAW,EAAE,IAAI,EAAE,IAAI,EAAE;IAC9C;;AAGA,IAAA,IAAI,UAAU,CAAC,SAAS,IAAI,UAAU,CAAC,SAAS,CAAC,MAAM,GAAG,CAAC,EAAE;AACzD,QAAA,UAAU,CAAC,QAAQ,GAAG,KAAK;AAC3B,QAAA,UAAU,CAAC,OAAO,GAAG,SAAS;IAClC;;IAGA,MAAM,WAAW,CAAC,aAAa,CAAC,IAAI,CAAC,EAAE,EAAE,UAAU,CAAC;AACpD,IAAA,UAAU,CAAC,IAAI,CAAC,6BAA6B,EAAE;QAC3C,MAAM,EAAE,IAAI,CAAC,EAAE;QACf,KAAK,EAAE,UAAU,CAAC,kBAAkB;AACpC,QAAA,YAAY,EAAE,CAAC,CAAC,UAAU,CAAC,SAAS;AACvC,KAAA,CAAC;AAEF,IAAA,SAAS,CAAC;AACN,QAAA,IAAI,EAAE,UAAU;AAChB,QAAA,EAAE,EAAE,IAAI,CAAC,GAAG,EAAE;AACd,QAAA,IAAI,EAAE,qBAAqB;AAC3B,QAAA,OAAO,EAAE,UAAU;AACtB,KAAA,CAAC;AAEF,IAAA,MAAM,UAAU,CAAC,kBAAkB,EAAE;AACrC,IAAA,MAAM,sBAAsB,CAAC,OAAO,EAAE,IAAI,EAAE;AACxC,QAAA,aAAa,EAAE,CAAA,mBAAA,EAAsB,IAAI,CAAC,KAAK,CAAA,CAAE;AACpD,KAAA,CAAC;;AAGF,IAAA,IAAI,UAAU,CAAC,kBAAkB,GAAG,GAAG,IAAI,UAAU,CAAC,SAAS,IAAI,UAAU,CAAC,SAAS,CAAC,MAAM,GAAG,CAAC,EAAE;AAChG,QAAA,UAAU,CAAC,IAAI,CAAC,uDAAuD,EAAE;YACrE,MAAM,EAAE,IAAI,CAAC,EAAE;YACf,KAAK,EAAE,UAAU,CAAC,kBAAkB;AACpC,YAAA,aAAa,EAAE,UAAU,CAAC,SAAS,CAAC,MAAM;AAC7C,SAAA,CAAC;AAEF,QAAA,SAAS,CAAC;AACN,YAAA,IAAI,EAAE,UAAU;AAChB,YAAA,EAAE,EAAE,IAAI,CAAC,GAAG,EAAE;AACd,YAAA,IAAI,EAAE,oBAAoB;YAC1B,OAAO,EAAE,UAAU,CAAC,SAAS;AAChC,SAAA,CAAC;IACN;SAAO;AACH,QAAA,UAAU,CAAC,IAAI,CAAC,wDAAwD,EAAE;YACtE,MAAM,EAAE,IAAI,CAAC,EAAE;YACf,KAAK,EAAE,UAAU,CAAC,kBAAkB;AACvC,SAAA,CAAC;IACN;;IAGA,IAAI,CAAC,WAAW,EAAE;AACd,QAAA,SAAS,CAAC,OAAO,CAAC,iBAAiB,CAAC,gBAAgB,EAAE,EAAE,KAAK,EAAE,UAAU,EAAE,CAAC,CAAC;QAC7E,OAAO,EAAE,MAAM,EAAE,WAAW,EAAE,IAAI,EAAE,IAAI,EAAE;IAC9C;;IAGA,MAAM,YAAY,GAAG,MAAM,WAAW,CAAC,YAAY,CAAC,IAAI,CAAC,EAAE,CAAC;IAC5D,IAAI,YAAY,EAAE,SAAS,IAAI,CAAC,YAAY,CAAC,QAAQ,EAAE;;AAEnD,QAAA,SAAS,CAAC,OAAO,CAAC,iBAAiB,CAAC,gBAAgB,EAAE,EAAE,KAAK,EAAE,UAAU,EAAE,CAAC,CAAC;QAC7E,OAAO,EAAE,MAAM,EAAE,WAAW,EAAE,IAAI,EAAE,IAAI,EAAE;IAC9C;;AAGA,IAAA,SAAS,CAAC,OAAO,CAAC,iBAAiB,CAAC,gBAAgB,EAAE,EAAE,KAAK,EAAE,UAAU,EAAE,CAAC,CAAC;AAC7E,IAAA,OAAO,EAAE,MAAM,EAAE,WAAW,EAAE;AAClC;;;;"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@posthog/agent",
-  "version": "1.20.0",
+  "version": "1.22.0",
   "description": "TypeScript agent framework wrapping Claude Agent SDK with Git-based task execution for PostHog",
   "main": "./dist/index.js",
   "module": "./dist/index.js",
@@ -49,7 +49,7 @@
     "typescript": "^5.5.0"
   },
   "dependencies": {
-    "@anthropic-ai/claude-agent-sdk": "^0.1.30",
+    "@anthropic-ai/claude-agent-sdk": "^0.1.47",
     "dotenv": "^17.2.3",
     "yoga-wasm-web": "^0.3.3"
   },

package/src/adapters/claude/claude-adapter.ts

@@ -18,72 +18,71 @@ export class ClaudeAdapter implements ProviderAdapter {
     };
   }
 
-  transform(sdkMessage: SDKMessage): AgentEvent | null {
+  transform(sdkMessage: SDKMessage): AgentEvent[] {
     const baseEvent = { ts: Date.now() };
 
-    // Handle stream events
     if (sdkMessage.type === 'stream_event') {
       const event = sdkMessage.event;
 
       switch (event.type) {
         case 'message_start':
-          return {
+          return [{
             ...baseEvent,
             type: 'message_start',
             messageId: event.message?.id,
             model: event.message?.model
-          };
+          }];
 
         case 'content_block_start':
           const contentBlock = event.content_block;
-          if (!contentBlock) return null;
+          if (!contentBlock) return [];
 
-          return {
+          return [{
             ...baseEvent,
             type: 'content_block_start',
             index: event.index,
             contentType: contentBlock.type as 'text' | 'tool_use' | 'thinking',
             toolName: contentBlock.type === 'tool_use' ? contentBlock.name : undefined,
             toolId: contentBlock.type === 'tool_use' ? contentBlock.id : undefined
-          };
+          }];
 
         case 'content_block_delta':
           const delta = event.delta;
-          if (!delta) return null;
+          if (!delta) return [];
 
           if (delta.type === 'text_delta') {
-            return {
+            return [{
               ...baseEvent,
               type: 'token',
               content: delta.text,
               contentType: 'text'
-            };
+            }];
           } else if (delta.type === 'input_json_delta') {
-            return {
+            return [{
               ...baseEvent,
               type: 'token',
               content: delta.partial_json,
               contentType: 'tool_input'
-            };
+            }];
           } else if (delta.type === 'thinking_delta') {
-            return {
+            return [{
               ...baseEvent,
               type: 'token',
               content: delta.thinking,
               contentType: 'thinking'
-            };
+            }];
           }
-          return null;
+          return [];
 
         case 'content_block_stop':
-          return {
+          return [{
             ...baseEvent,
             type: 'content_block_stop',
             index: event.index
-          };
+          }];
 
         case 'message_delta':
-          return {
+          return [{
             ...baseEvent,
             type: 'message_delta',
             stopReason: event.delta?.stop_reason,
@@ -91,20 +90,20 @@ export class ClaudeAdapter implements ProviderAdapter {
             usage: event.usage ? {
               outputTokens: event.usage.output_tokens
             } : undefined
-          };
+          }];
 
         case 'message_stop':
-          return {
+          return [{
             ...baseEvent,
             type: 'message_stop'
-          };
+          }];
 
         case 'ping':
           // Ignore ping events
-          return null;
+          return [];
 
         case 'error':
-          return {
+          return [{
             ...baseEvent,
             type: 'error',
             message: event.error?.message || 'Unknown error',
@@ -115,18 +114,20 @@ export class ClaudeAdapter implements ProviderAdapter {
               code: event.error.code,
             } : undefined,
             sdkError: event.error
-          };
+          }];
 
         default:
-          return null;
+          return [];
       }
     }
 
     // Handle assistant messages (full message, not streaming)
     if (sdkMessage.type === 'assistant') {
       const message = sdkMessage.message;
+      const events: AgentEvent[] = [];
 
       // Extract tool calls from content blocks
+      // A single assistant message can contain multiple tool_use blocks
       if (message.content && Array.isArray(message.content)) {
         for (const block of message.content) {
           if (block.type === 'tool_use') {
@@ -139,31 +140,36 @@ export class ClaudeAdapter implements ProviderAdapter {
               args: block.input || {},
               parentToolUseId: sdkMessage.parent_tool_use_id
             };
-            // Enrich with tool metadata
-            return this.toolMapper.enrichToolCall(toolCallEvent);
+            // Enrich with tool metadata and add to events array
+            events.push(this.toolMapper.enrichToolCall(toolCallEvent));
           }
         }
       }
 
+      // If we found tool calls, return them
+      if (events.length > 0) {
+        return events;
+      }
+
       // If no tool calls, emit status event
-      return {
+      return [{
         ...baseEvent,
         type: 'status',
         phase: 'assistant_message',
         messageId: message.id,
         model: message.model
-      };
+      }];
     }
 
     // Handle user messages
     if (sdkMessage.type === 'user') {
       const message = sdkMessage.message;
+      const events: AgentEvent[] = [];
 
-      // Check for tool results in content blocks
-      if (message?.content && Array.isArray(message.content)) {
+      // Check for tool results in content blocks, A single user message can contain multiple tool_result blocks
+      if (message?.content && Array.isArray(message.content)) { 
         for (const block of message.content) {
           if (block.type === 'tool_result') {
-            // Create tool_result event and enrich with metadata
             const toolResultEvent = {
               ...baseEvent,
               type: 'tool_result' as const,
@@ -173,29 +179,33 @@ export class ClaudeAdapter implements ProviderAdapter {
               isError: block.is_error,
               parentToolUseId: sdkMessage.parent_tool_use_id
             };
-            // Enrich with tool metadata
-            return this.toolMapper.enrichToolResult(toolResultEvent);
+            events.push(this.toolMapper.enrichToolResult(toolResultEvent));
           }
         }
       }
 
+      // If we found tool results, return them
+      if (events.length > 0) {
+        return events;
+      }
+
       // Otherwise extract text content
       const textContent = this.extractUserContent(message?.content);
       if (!textContent) {
-        return null;
+        return [];
       }
-      return {
+      return [{
         ...baseEvent,
         type: 'user_message',
         content: textContent,
         isSynthetic: sdkMessage.isSynthetic
-      };
+      }];
     }
 
     // Handle result messages
     if (sdkMessage.type === 'result') {
       if (sdkMessage.subtype === 'success') {
-        return {
+        return [{
           ...baseEvent,
           type: 'done',
           result: sdkMessage.result,
@@ -206,9 +216,9 @@ export class ClaudeAdapter implements ProviderAdapter {
           usage: sdkMessage.usage,
           modelUsage: sdkMessage.modelUsage,
           permissionDenials: sdkMessage.permission_denials
-        };
+        }];
       } else {
-        return {
+        return [{
           ...baseEvent,
           type: 'error',
           message: `Execution failed: ${sdkMessage.subtype}`,
@@ -220,14 +230,14 @@ export class ClaudeAdapter implements ProviderAdapter {
             num_turns: sdkMessage.num_turns
           },
           sdkError: sdkMessage
-        };
+        }];
       }
     }
 
     // Handle system messages
     if (sdkMessage.type === 'system') {
       if (sdkMessage.subtype === 'init') {
-        return {
+        return [{
           ...baseEvent,
           type: 'init',
           model: sdkMessage.model,
@@ -239,18 +249,18 @@ export class ClaudeAdapter implements ProviderAdapter {
           slashCommands: sdkMessage.slash_commands,
           outputStyle: sdkMessage.output_style,
           mcpServers: sdkMessage.mcp_servers
-        };
+        }];
       } else if (sdkMessage.subtype === 'compact_boundary') {
-        return {
+        return [{
           ...baseEvent,
           type: 'compact_boundary',
           trigger: sdkMessage.compact_metadata.trigger,
           preTokens: sdkMessage.compact_metadata.pre_tokens
-        };
+        }];
       }
     }
 
-    return null;
+    return [];
   }
   
   createStatusEvent(phase: string, additionalData?: any): StatusEvent {

package/src/adapters/types.ts

@@ -12,10 +12,10 @@ export interface ProviderAdapter {
   name: string;
 
   /**
-   * Transform a provider-specific SDK message into an AgentEvent.
-   * Returns null if the message should be ignored.
+   * Transform a provider-specific SDK message into one or more AgentEvents.
+   * Returns an array of events (can be empty if the message should be ignored).
    */
-  transform(sdkMessage: unknown): AgentEvent | null;
+  transform(sdkMessage: unknown): AgentEvent[];
 
   /**
    * Create a standardized status event.

package/src/agent.ts

@@ -214,10 +214,9 @@ export class Agent {
             this.logger.debug('Received message in direct run', message);
             // Emit raw SDK event
             this.emitEvent(this.adapter.createRawSDKEvent(message));
-            // Emit transformed event
-            const transformedEvent = this.adapter.transform(message);
-            if (transformedEvent) {
-                this.emitEvent(transformedEvent);
+            const transformedEvents = this.adapter.transform(message);
+            for (const event of transformedEvents) {
+                this.emitEvent(event);
             }
             results.push(message);
         }
@@ -308,11 +307,16 @@ export class Agent {
         return commitHash;
     }
 
-    async createPullRequest(taskId: string, branchName: string, taskTitle: string, taskDescription: string): Promise<string> {
+    async createPullRequest(
+        taskId: string,
+        branchName: string,
+        taskTitle: string,
+        taskDescription: string,
+        customBody?: string
+    ): Promise<string> {
         this.logger.info('Creating pull request', { taskId, branchName, taskTitle });
 
-        // Build PR body
-        const prBody = `## Task Details
+        const defaultBody = `## Task Details
 **Task ID**: ${taskId}
 **Description**: ${taskDescription}
 
@@ -320,6 +324,7 @@ export class Agent {
 This PR implements the changes described in the task.
 
 Generated by PostHog Agent`;
+        const prBody = customBody || defaultBody;
 
         const prUrl = await this.gitManager.createPullRequest(
             branchName,
@@ -447,11 +452,15 @@ Generated by PostHog Agent`;
         }
 
         const branchName = await this.gitManager.getCurrentBranch();
+        const finalizeResult = stepResults['finalize'];
+        const prBody = finalizeResult?.prBody;
+
         const prUrl = await this.createPullRequest(
             task.id,
             branchName,
             task.title,
-            task.description ?? ''
+            task.description ?? '',
+            prBody
         );
 
         this.emitEvent(this.adapter.createStatusEvent('pr_created', { prUrl }));

package/src/file-manager.ts

@@ -1,5 +1,5 @@
 import { promises as fs } from 'fs';
-import { join } from 'path';
+import { join, extname } from 'path';
 import type { SupportingFile, ResearchEvaluation } from './types.js';
 import { Logger } from './utils/logger.js';
 
@@ -9,6 +9,14 @@ export interface TaskFile {
   type: 'plan' | 'context' | 'reference' | 'output' | 'artifact';
 }
 
+export interface LocalArtifact {
+  name: string;
+  content: string;
+  type: TaskFile['type'];
+  contentType: string;
+  size: number;
+}
+
 export class PostHogFileManager {
   private repositoryPath: string;
   private logger: Logger;
@@ -223,11 +231,7 @@ export class PostHogFileManager {
       const content = await this.readTaskFile(taskId, fileName);
       if (content !== null) {
         // Determine type based on file name
-        let type: SupportingFile['type'] = 'reference';
-        if (fileName === 'plan.md') type = 'plan';
-        else if (fileName === 'context.md') type = 'context';
-        else if (fileName === 'requirements.md') type = 'reference';
-        else if (fileName.startsWith('output_')) type = 'output';
+        const type = this.resolveFileType(fileName);
         
         files.push({
           name: fileName,
@@ -240,4 +244,53 @@ export class PostHogFileManager {
     
     return files;
   }
+
+  async collectTaskArtifacts(taskId: string): Promise<LocalArtifact[]> {
+    const fileNames = await this.listTaskFiles(taskId);
+    const artifacts: LocalArtifact[] = [];
+
+    for (const fileName of fileNames) {
+      const content = await this.readTaskFile(taskId, fileName);
+      if (content === null) {
+        continue;
+      }
+
+      const type = this.resolveFileType(fileName);
+      const contentType = this.inferContentType(fileName);
+      const size = Buffer.byteLength(content, 'utf8');
+
+      artifacts.push({
+        name: fileName,
+        content,
+        type,
+        contentType,
+        size,
+      });
+    }
+
+    return artifacts;
+  }
+
+  private resolveFileType(fileName: string): TaskFile['type'] {
+    if (fileName === 'plan.md') return 'plan';
+    if (fileName === 'context.md') return 'context';
+    if (fileName === 'requirements.md') return 'reference';
+    if (fileName.startsWith('output_')) return 'output';
+    if (fileName.endsWith('.md')) return 'reference';
+    return 'artifact';
+  }
+
+  private inferContentType(fileName: string): string {
+    const extension = extname(fileName).toLowerCase();
+    switch (extension) {
+      case '.md':
+        return 'text/markdown';
+      case '.json':
+        return 'application/json';
+      case '.txt':
+        return 'text/plain';
+      default:
+        return 'text/plain';
+    }
+  }
 }

package/src/git-manager.ts

@@ -30,6 +30,14 @@ export class GitManager {
     this.logger = config.logger || new Logger({ debug: false, prefix: '[GitManager]' });
   }
 
+  private escapeShellArg(str: string): string {
+    return str
+      .replace(/\\/g, '\\\\')
+      .replace(/"/g, '\\"')
+      .replace(/`/g, '\\`')
+      .replace(/\$/g, '\\$');
+  }
+
   private async runGitCommand(command: string): Promise<string> {
     try {
       const { stdout } = await execAsync(`cd "${this.repositoryPath}" && git ${command}`);
@@ -130,7 +138,7 @@ export class GitManager {
   }
 
   async addFiles(paths: string[]): Promise<void> {
-    const pathList = paths.map(p => `"${p}"`).join(' ');
+    const pathList = paths.map(p => `"${this.escapeShellArg(p)}"`).join(' ');
     await this.runGitCommand(`add ${pathList}`);
   }
 
@@ -204,7 +212,7 @@ export class GitManager {
   }
 
   private buildCommitCommand(message: string, options?: { allowEmpty?: boolean; authorName?: string; authorEmail?: string }): string {
-    let command = `commit -m "${message.replace(/"/g, '\\"')}"`;
+    let command = `commit -m "${this.escapeShellArg(message)}"`;
 
     if (options?.allowEmpty) {
       command += ' --allow-empty';
@@ -446,7 +454,7 @@ Generated by PostHog Agent`;
 
     await this.pushBranch(branchName);
 
-    let command = `gh pr create --title "${title.replace(/"/g, '\\"')}" --body "${body.replace(/"/g, '\\"')}"`;
+    let command = `gh pr create --title "${this.escapeShellArg(title)}" --body "${this.escapeShellArg(body)}"`;
 
     if (baseBranch) {
       command += ` --base ${baseBranch}`;

package/src/posthog-api.ts

@@ -1,4 +1,15 @@
-import type { Task, TaskRun, LogEntry, SupportingFile, PostHogAPIConfig, PostHogResource, ResourceType, UrlMention } from './types.js';
+import type {
+  Task,
+  TaskRun,
+  LogEntry,
+  SupportingFile,
+  PostHogAPIConfig,
+  PostHogResource,
+  ResourceType,
+  UrlMention,
+  TaskRunArtifact,
+  TaskArtifactUploadPayload,
+} from './types.js';
 
 interface PostHogApiResponse<T> {
   results?: T[];
@@ -170,6 +181,27 @@ export class PostHogAPIClient {
     });
   }
 
+  async uploadTaskArtifacts(
+    taskId: string,
+    runId: string,
+    artifacts: TaskArtifactUploadPayload[]
+  ): Promise<TaskRunArtifact[]> {
+    if (!artifacts.length) {
+      return [];
+    }
+
+    const teamId = this.getTeamId();
+    const response = await this.apiRequest<{ artifacts: TaskRunArtifact[] }>(
+      `/api/projects/${teamId}/tasks/${taskId}/runs/${runId}/artifacts/`,
+      {
+        method: 'POST',
+        body: JSON.stringify({ artifacts }),
+      }
+    );
+
+    return response.artifacts ?? [];
+  }
+
   /**
    * Fetch logs from S3 using presigned URL from TaskRun
    * @param taskRun - The task run containing the log_url